Autonomous Ground Vehicle Path Planning in Urban Environments Using GNSS and Cellular Signals Reliability Maps: Models and Algorithms SONYA RAGOTHAMAN © , Member, IEEE Aerospace Corporation, El Segundo, CA, USA MAHDI MAARI ZAHER M. KASSAS ©, Senioe Memb, IEEE Universi of California Irvine,CA, USA. In this aril, autonomous ground vehicle (AGY) path planning Is considered. The AGV Is assumed to be equipped With receivers ‘capable of producing pseudorange measurements to overhead global navigation satelite systems (GNSS) satellites and to cellalar base ‘avironment Parameters ofthe cellular psewdoranges tranemitter clock blas are estimated in an inialeation sn open-sky environment. The AGV fuses these pseudoranges {toproduce an estimate about ts own states, The AGV is also equipped «a three-dimensional building map ofthe envionment. Starting ‘roma known starting poat, the AGV desires to reach 2 known target point by taking the shortest distance, while minimizing the AGV's positon estimation error and guaranteeing thatthe AG's position ‘estimation uncertainty Isbelow adesired threshold Toward tis objec- tive, socalled signal reliability map isfirst generated, which provides fahout repions where large errors die to poot GNSS. lar signal multipath are expected. The vehicle p tocaleulat the peition mean-squared ‘error (MSE), Am analytical expression forthe AGV'sstate estimates s Manuscript received August 12,2019; revised February 8, 2020, Septem ber 28,2020, and November 1, 2020: released fr publication November 30, 2020. Date of publication January 26, 2021; date of eurent version Tune 9, OL, No. 1.1 109/TAES. m1 3053690 Referceing ofthis contribution was handled by O. Osechas ‘Ths work was supported in pat by the Ofce of Naval Research (ONR) under Grint NQQ14-19-1-2613 and in part by the National Science Foundation (NSF) under Grant 1920571 and Grant 1929968, Authors’ addresses: Sonyas Ragothaman was with the Deparment of lcetial and Computer Engineering, University of California, Riverside {CA 9285 USA. She ismos with Acrospace Compoaton, El Segond, CA ‘90245 USA, E-mail: (Sonya ragothaman gmail com) Maaref Mae and ‘Zaher (Zak) M. Kasse ae with Departmen of Mechanical and Aerospace Engineering, University of California, vine, CA 90245 USA, E-mail: Aonmsire@ ue eds rhassas ie.) (Comsponiine chor: Zaher (Zak) M. Kass) (On18 9951 © 2021 THEE 1E TRANSACTIONS ON AEROSPACE AND ELECTRONI derived, whichis used to find an upper bound on the position bias due kecping the position estimation uncertainty and position estimation blas due to multipath below desired thresholds. The path planning approach yields the optimal path together with a st of feasible paths tnd reliable GNSS satelite a F base stations to use along re pats NOMENCLATURE ep Node that proceeds a along a path. Hou Spur node. a(x) Complex amplitude of signal path x and LTE symbol f bap Multipath bias assi map. Speed of light. ag) Cost along the path from sto g dros Length of the LOS path. dist(p) Distance represented by location p. JB.) Path planning weight assigned to the edge by tween nodes a and f. fil.) Path planning weight assigned to the edge be- tween nodes a and B. FX) Path planning weight assigned edges along the zh path AS 8 ‘Target node hay Channel impulse response at time ¢ and LTE, symbol i, LTE symbol ‘Threshold for position uncertainty. Fras Upper bound on position bias, Br Node that proceeds f along a path. Xn © yi ull) + X2m(i). multipath interfere Bicanm math cellular clock bias. or, GNSS receiver clock bias. Biren Cellular receiver clock bias. Btonon —_‘Honospheric delay. Bison Tropospheric delay. Bly Experiment cellular receiver clock bias or, periment GNSS receiver clock bias. ar, Receiver clock bias. en Constant bias between cellular clock biases, Ts Pseudorange bias threshold. mw Weighted pseudorange bias threshold, vig) A path from nodes s to g Yost A FOOL path FFOM.§ €0 dbp You A spur path from cpr 10 & vig) ‘A path from nodes $10 g. « Threshold for amplitude of received path, Aowx(P. 1) Maximum eigenvalue used in path planning. g (4,0), directed graph with «nodes and « di ‘Sequence of all cellular signal reliability maps. ‘Signal reliability maps for the ath satellite ‘Sequence of all satellite signal reliability maps. Set of all paths from start to target FEMS VOL. 57,NO.3- JUNE2021 UE, EePeston Pens 0) M it MSE(p.1) RaGOTA MAN oy ‘Sequence of time intervals for the nth satellite and the pth location, Set of ¢ shortest paths. Set of candidate & shortest paths. Partition of H corresponding tothe clock states. = BB = BrBNG Partition of H corresponding to the position states. 1G. Partition of G corresponding (0 the cellular measurements, Partition of G corresponding t the satellite ‘measurements, Measurement Jaco 1G, BI Measurement noise covariance. Cholesky factor of R. Path from start to target location, Celular pseudorange measurement. Satelite pseudorange measurement Path delay between the sth impulse tothe LOS path, Symbol tin =R," Unbiased position error Ar, cA&"]", unbiased error state. State estimation error. Residual vector. Unbiased measurement etror vector Vector of deterministic measurement biases Position of the mth cellular transmiter. Position of the mh satelite transmit. Position bias Vehicles 3-D positon. Vector of measurement noises 7. e871, state bias AGV Sate vector. Moiied cellular pscudorange measurement Modified satellite pseudorange measurement. Measurement vector Correlatr pacing in the LTE receiver racking loop. [Number of paths in the search space. Signal power Number of measurements used incellularmea- surement initialization, Number of subcarrier symbols in the pilot se- quence M, cellular transmitter index Number of cellular base stations in the envi- ronment. Number of reliable cellular measurements Position MSE used in path plannit 1V, satelite transmitter index. Number of GNSS satellites in the environment Number of reliable GNSS measurements. Location index vows P Number of locations in the reliability map. P(B.@) Location indices from nodes to a. Pe ‘Target position index. De Start position index. s Eyaluated edges. s Start node, t Time, Tr Number of time intervals. feodr End time for time interval fea art time for time interval 1, Sampling interval. Vv Unvisited nodes eon ~N(0,02). taov AGV constant speed eet NO, 23h) Ya Metts — Perms ~ NO. O29 + OE) Upnsn NO. Oa x Number ef impulses. Shiu Modified cellular measurement for initializa- tion. 1. INTRODUCTION Autonomous ground vehicles (AGVs) are predicted to, improve the quality of life by automating the monotonous task of driving, while reducing crash fatalities due to hu- ‘man error. Considering the breadth of applications AGV could revolutionize (e.g.,cargodelivery, taxi services,emer- ney response, intelligent farming, etc.), a myriad num- ber of corporations have been investing in AGV enabling technologies [1] In light of recent tragedies [2], itis evident that AGVs. need extremely reliable sensing and navigation systems, Virtually all current vehicular navigation systems rely on global navigation satellite systems (GNSS). While GNSS. provide an accurate position estimate with respect to a ¢lobal frame, its signals are unreliable forthe safety critical application of autonomous driving. On one hand, GNSS. signals are susceptible to unintentional interference, inten- tional jamming, and malicious spoofing (3]. On the other hand, GNSS signals are severely attenuated in deep urban canyons. Urban high-rise structures block, shadow, and reflect signals from GNSS satellites. This makes locales at which reliable and accurate GNSS position estin achievable to be rather spotty in urban environments [4] To overcome the limitations of GNSS, current vebicular navigation systems fuse GNSS receivers with a suite of sensor-based technologies (e.g, inertial measurement uni, lidar, and eamera). These dead reckoning type sensors nex an external aiding source to account for their accumulated error, only provide local position estimates, may not prop- erly function in all environments (e.g. fog, snow, rain, dust, ete.) and are still susceptible to malicious attacks [5 ial-based technologies alleviate some of the short- comings of sensor-based technologies and provide a global position estimate, but some require installing dedicated infrastructure (e.g, pseudolites [6)), while others produce a SIGN UA ESVEONMEN IS 503‘coarse position estimate (e.g. digital television signals [7). Among signal-based technologies, cellular signals are very attractive in urban environments due to their inherent at- tributes [8], which include the following. 1) Reception at high carrier-to-noise ratio (C/N): A CC/Ny between 55 and 80 dB-Hz has been observed in urban environments. 2) Abundance: Numerous stations corresponding to different cellular providers exist in urban environ- ‘ments. 3) Favorable geometric configurations: Cells in cellu- lar networks are designed to minimize interference between stations. Recent work have demonstrated metertevel accurate navigation with cellular signals on ground vehicles [}-{14] and centimeter-level accurate navigation on serial vehi- cles [15], [16]. In addition to utilizing cellular measure- ments, AGV can employ path planning to improve po- sitioning. The objective of path planning, also known as motion planning or trajectory generation, is 10 optimize a path over a defined objective function (e.g., path length, path dutation, position uncertainty, et.) ‘This article and its Sequel consider the following prob- Jem, originally defined in [17]. An AGV is equipped with receivers capable of producing pseudoranges to overhead GNSS satellites and to cellular long-term evolution (LTE) base stations in its environment. From a known starting point, given a 3-D building map, the AGV desires to reach 1 known target point by taking the shortest distance while minimizing the AGY’s position estimation error and guar- anteeing that the AGV’s position estimation uncertainty is below a desired threshold. This article focuses on modeli analytical derivations, and algorithm development, while the sequel paper [18] presents comprehensive simulation and experimental resus for different realistic sc ‘evaluate the accuracy and efficacy of the proposed approach ‘ona ground vehicle navigating in adcep urban environment. ‘The contributions of this article are fourfold, Firs, this article introduces GNSS and cellular signal reliability maps, which store information about areas where GNSS satellites have unobstructed line-of-sight (LOS) and areas where cel ular pseudorange measurements produce acceptable errors ‘due to multipath (i.e., below a certain threshold), Second, this article proposes a method for calculating the position ‘mean-squared error (MSE), which measures the expected ‘quality of the position estimate at different locations and time along the road network, Third, this article derives an analytical expression for the position bias due to multipath, which is shown to be parameterized by a bound on the pseudorange bias and the largest eigenvalue of the position estimation error covariance. Fourth, this article proposes @ path planning method that is based on Dijkstra’s algorithm, which considers path length, position MSE, and the largest eigenvalue of the position estimation error covariance. ‘This article is organized as follows. Section I overviews related work in path planning. Section II describes the AGV. path planning problem, Section IV deseribes the GNSS and arios to IEEE TRANSACTIONS ON AEROSPACE AND ELECT cellular measurement models and estimation framework used in this article, Section V presents methods to generate GNSS and cellular signal reliability maps, Section VI shows how to calculate the position MSE from signal reliabil ity maps, and finds an upper bound on the position bias. Section VII describes optimal path planning generation, which considers the path length, position MSE, and upper bound on the position bias. Finally, Section VIII concludes this article Il. RELATED WORK Path planning has been extensively studied in different contexts, In robotics simultaneous localization and mapping. (SLAM), path planning has been considered to steer the robot in the most informative direction [19] or minimize the probability of becoming lost [20]. Ininteligent vehicles, predictive and multirate reactive planning was considered by using the vehicle dynamics, position uncertainty, and obstacle volume to derive Lagrange-Euler equations and potential fields [21]. Also in intelligent vehicles, a partially observable Markov decision process (POMDP) was used for belief state planning to account for uncertainty from noise in sensor data and the intentions of human drivers [22] In traffic management, the probability of arriving o% in a bus network under travel time uncertainty was solved using path planning [23]. In target tracking, path planning ‘was considered to minimize the time between observations of atarget made by several mobile unmanned aerial vehicles (UAVs) [24]. UAV path planning was considered in [25] to account for map, location, and sensing uncertainty, while it dered in [26] to generate collision avoidance paths using a Markov decision process. In marine vehicles, path planning was considered to localize underwater acoustic transponders (27]. In navigation via signals of opportunity, path planning was considered to maximize the information gathered from ambient signals to yield accurate positioning and timing [28]. Path planning has been also proposed to reduce position errors from GNSS signals. These position errors can be predicted based on GNSS signal power and availability prediction, proposed in [29]-(35]. In [36], a UAV was steered toward higher received global positioning system (GPS) signal-to-noise ratio (SNR) to improve the signal quality for positioning. In [37] and [38], path planning ‘was considered to account for uncertainty in GNSS-derived position estimates by using dilution of precision (DOP) of GNSS satellites. Stochastic reachability analysis was used {in [39] 10 optimize over DOP-based position uncertainty for A UAV while also avoiding collisions. In [40], predictions ‘of DOP-based position uncertainty and multipath errors for GNSS signals were considered to define and calculate quality of service (QoS) parameters (e.g,, availability, curacy, reliability, and continuity) along a given route, A. localization error map, with GPS positioning errors due to ‘multipath-induced biases, was used for UAV path planning in [41]. In [42], a UAV ignored measurements with high ‘multipath-induced biases and avoided areas with poor DOP, time (ONICSYSTEMS VOL. 57,NO.3 JUNE 2021predicted by so-called obstructed LOS volumes and multi path volumes. ‘The aforementioned metrics can be used for path plan- ning to reduce position errors for an AGV that utlzes GNSS and cellular signals. However, addtional challenges need to be accounted for when fusing cellular signals alongside GNSS signals for AGV navigation, Due to the terrestrial nature and modulation scheme of cellular signal, their received signals in urban environments suffer from LOS blockage and severe multipath compared to GNSS signals (43. Ths induces erors in cellular navigation ob- servables (pseudoranges), which contaminate the postion ate, jeopardizing the sae operation of AGN, Though ‘multipath mitigation techniques for navigation with cella lar signals in the presence of the LOS signal have been proposed [44], 145], an AGV could furher improve the positioning by optimizing its path for small positioning In this article, path planning is used to guarantee a desired level of accuracy, by choosing a path that yields acceptable DOP-based uncertainty and multipath-induced biases through the postion MSE metric, In contrast to the cost fonction in [31], which accounts for multipath-indvced bias, the position MSE cos function proposed inthis article also accounts for other sources of eror through DOP-based uncertainty. Also, this article explicitly considers orthogo- nal frequency-ivision multiplexing (OFDM) signal models in multipath calculations. Furthermore, in contrast to the path planning algorithm in [42], which uses the predicted Positioning error in the path planning constraints, ths ar ticle uses positioning etror in the cost function and the constraints This article demonstrates the feasibility of using path planning to improve positioning using GNSS and cellular measurements. For practicality, this work can be combined ‘with more sophisticated techniques for GNSS positioning ‘with 3-D maps [46 integrity monitoring notions (47), map matching [48], and sensor fasion and filtering techniques. Il, PROBLEM DESCRIPTION ‘This article considers the following problem. An AGV drives in an urban environment. The AGV is equipped with receivers capable of producing pseudorange measurements ‘on GNSS satellites and nearby cellular base stations. The AGY uses these pseuddorange measurements to ¢ vehicles’ 3-D position, AG) 7 and the clock bins of the cellular base stations. The AGV desires fo reach a target location by taking the shortest possible path while guaranteeing that the uncertainty about its own position estimate is below a specified threshold (e.g, for satety concerns), A trajectory that satisfies this “objective is generated either locally (ie. within the AGV’s processor) or at a cloud-hosted path planning generator. ‘The path planning generator uses a 3-D building map of the environment to generate a so-called signal reliability map. The signal reliability map is a spatiotemporal map of the environment that measures the expected accuracy from RAGOTHAMAN ET ALL: AUTONOMOUS GROUND VEHI PATH PLANNING. Fig Depiction of an AGY navigating with GNSS and cellar Signal. Here, the ed cites represent lesions where GNSS and abl (i the position estimate proce with cela signs are such signal will vo Tinted LOS to GNSS satlits andr age eel The bue circles represent locations where GNSS and el reliable, Te proposed frame ork gencrats the optimal lve trajectory. ‘which stses the contin while minimizing the distance tive! Tetween the start and target postions together with minimizing the Position MSE. This hgure wis obtained with ArGIS (29), using GNSS and cellular signals to produce an estimate of the AGV’s state. For GNSS signals, the signal reliability ‘map is spatiotemporal, and specifies the GNSS satellites to which the AGV would have a blocked LOS for different locations at different times in the environment, For cellular signals, the signal reliability map is a spatial map specifying the expected pscudorange hias due to multipath, The signal reliability maps are used to calculate the position MSE at each location, which in turn is used to generate an optimal path for the AGV to follow. This path is generated by ‘minimizing the total distance traveled and MSE, while guaranteeing that the bias in the position estimate due to ‘multipath is below a desired threshold as well as ensuring that the maximum position uncertainty is below a desired limit. The thresholds can be chosen by the user (e.g., based oon the lane-level accuracy). In addition, the path generator produces a table of reliable GNSS satellites and cellular base stations along the optimal path for the AGV to use as it traverses the optimal path. The path planning generator also ‘outputs other feasible paths the AGV could take. These sub- ‘optimal, yet feasible paths could be useful, should the AGV choose to not follow the optimal path, e.g,, o avoid trafic jams and road blockages duc to construction or emergency. Fig. I depicts the objective ofthe optimal path planning gen- trator, Here, the red circles on the street represent locations that violate the user-specified constraints (position bias or position uncertainty exceeding their respective thresholds). ig. 2 illustrates a flowchart of the optimal path gener- ator framework developed in this article, IQ MODEL DESCRIPTION ALGORITHM, A. AGV-Mounted Receiver States AND ESTIMATION ‘The AGV receives signals from M spatially stationary cellular base stations. It is assumed that the coordinates LURIAN ENVIRONMENTS 165Environmental prot browse ‘GNSS ephemerides| [Cala bae | Endronmant maps Snduinnae alone S'S Cay a L_ ‘Signa eli User input map generation Departure time. + Peston bins __ PY Poon MSE and ‘iveshod” ——P] uncertainty constraint Postion ‘telson ncaa cea jeOptinal path Star postion Path planing [abe of reabe Gal Seite gas “Toret postion generation Fig.2. Flowchart ofthe propose path planing generator The path planing generator i ssumed to have knowledge of GNSS orbital data. «elu hse station postions and 3D envionment map. The user inputs far: dspam time, tr poston gst position, poston bias threshold ‘nd position uncertainty threshokd. The thresholds are wed as constraints to fnd the optimal path, The outputs of the path panning {he optimal path anu table of elise GNSS stellites and cella base station tows along the optimal pth of the cellular base stations ate known a prior’ (e.g via radio mapping or satelite images [50], [51]) and are stored locally on the AGV or on a cloud-hosted database, The 3D positon of the mth cellular base station is denoted Teotin © Lit Yen Ze” in the Earth-centered Earth- fixed (ECEF) coordinate frame. The AGV also receive signals from N GNSS satellites with known positions ‘The 3D position of the uth GNSS satellite is denoted 8S Fysisn = [Xgaains Yasue Zgmesa]” in the ECEF coordinate frame. ‘The unknown states include the vehicle's 3-D position Yoo 2r]" in the ECEF coordinate frame, the AGV- mounted receiver’s lock bias 8%, and the clock bias of the M cellular base stations {89}. The cellular LTE technical specification requires transmitters in neighboring cells to be synchronized in phase up to 10 sis [52]. The cellular base station clock bias is a time-varying state, but synchronization between the clock biases canbe exploited. Many cellular providers synchronize nearby base stations ina much tighter fashion as was demonstrated over a 24-h period in recent experimental studies [15], [S3}. This syn- hronization will be exploited in the proposed framework to minimize the number of states that will be estimate. Specifically only the clock bias of one ofthe base stations willbe estimated (refered tos the ftstbase station, without Joss of generality). The lock bias of the other celular base stations will be expressed as deviations from the clock bias of the first base station, The model of such deviation and the estimation algorithm will be discussed in the following subsections. B._AGV Measurements ‘The AGV: surements to the N GNSS satellites jounted receiver makes pseudorange mea- ‘The nth GNSS 166 pseudorange measurement is modeled as Pawn) = IrAK) —rars.nfy +e [8 K) ~ Btgns I] CBs) CBfp BD + Vanes ne) where ¢ is the speed of light: Bann. aNd Btuggan aC known ionospheric and tropospheric delays, respectively and Stpsan iS the known satelite clock bias. The model coefficients for the terms Ps.ns Stone ANd Stern AC transmitted in the satellite's navigation message. The term Brogan i8 found using the Hopfield model [54]. The term Ups 8 the measurement noise, which is modeled as a zero-mean white Gaussian random sequence with variance The measurements noise across different satellites [gnel%y are assumed to be independent. The nth GNSS pseudorange measurement is modified by subtracting the Know Stoqo.93 Bs,» AN CBtasay (0 yield [54] CBtionan ~ CBlaspan + CBts.e Ponsa — IF — Femanl|s + CBte + Ypmisn a Multipath interference and non-line-of-sight (NLOS) errors are not included in the measurement model because NLOS. ‘measurements are not used, and the GNSS receiver mit gates multipath when the LOS signal is present. ‘The AGV-mounted receiver also makes pseudorange measurements to the M cellular base stations. The mth cellular pseudorange measurement is modeled as [55] Pesta) = fet) — rest (816K) — Bt (AD) + sta) te Where tare is the measurement noise, which is modeled as a zero-mean white Gaussian sequence with variance qe The measurement noise across different cellular base stations (sto1yn}fly are assumed to be independent. The model assumes that there is no multipath-induced bias. 1a the case that there is multipath, a bia is introduced, caus- ing a mismatch with the model. The subsequent sections address bounding the pseudorange error, thereby, bounding the induced bias in the position estimate By exploiting the synchronization between nearby cel- lular base stations, the transmitter clock bias of the mth cellular measurement can be expressed as Btn nt 2 Bice.) + GF Yen (h) for m=2,....M, where éq isa deterministic constant bias, and v,,» is approximated as a zero-mean white noise sequence with variance 02,,. This model is valid over rel- atively short periods of time [53], but more sophisticated ‘models capturing the long-term behavior of synchroniza- tion, of lack of, ean be used [11], [15}, [56}, [57]. For all lular measurements other than the first cellular measure- went, the mth cellular pseudorange can be rewritten in terms of ede, namely tea.) 8 Wrek) — rel] + 81-48) w Peck ~ ent IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL.57,NO.3. JUNE 2021for m 6M, where ta ‘mean white noise sequence with variance 2, Finally, using (2), the cellular pseudorange measure: ‘ment (3) (othe M cellular base stations is modified accord ing to east) = [rB)~ nena +e [BAK ~ Stan) + 9,100) etm) & Peeiim(K) + Ee Ph) — Fly +e [6 ~ Bt 1] + Unm(kD for m mation proc: M. The next subsection describes an esti lure for éq and 9 C. Estimation of Cellular Measurement Clock Bias Perturbations ‘The perturbation parameters of the mth cellular clock bias from the first cellular clock bins [ef., 2)], namely, the ‘constant bias ¢, and the variance a2, can be estimated by the AGV.mounted receiver locally’ or assumed to be available from a cloud-hosted database. To estimate the ‘constant bias éy, and variance o?,,, the measurements are differenced according to pects E—Peatin(h = |retK) rel = Yrs) — rota + Frat = thal). AD ‘as (k) canbe swapped with another cellular measurement ‘without loss of generality. Its assumed thatthe dfferenc operation in (4) is performed in an open area where ris accurately estimated (eg., while the AGY is intialy sta tionary with clear LOS to GNSS satellites). Subsequently define the measurement ict) © pen (8) ~ Peal — for) + [retk) — rectly m+ Vean 1(K) = Wann) + Ve lB Assuming the measurement noise to be ergodic (ie., the ‘ensemble average equals the time average), é» and 2, can be estimated using a sample mean and a sample variance over K measurements, namely nth K “e) [iets Elem The value of K can be a fixed value chosen prior 10 the initialization, or can be determined during initialization by until the sample mean and variance converge. In practice, the value of K depends on the batch size for which (2) holds. Experimentally, it was observed that the sample mean and variance converged in around 0.5 s with RAGOTHAMAN ET AL: AUTONOMOUS GROUND VEHICL PATH PLANNING. ‘measurements at a s Kx 50 samples) umpling time T = 0.1 s (ic D._ Estimation of AGV States, The AGV's a8 (Fl, c81;,eBiausI" iS estimated from the measurement ss tal sar]? through Squares (WNLS) estimator. The measurement Jacobian used in the WNLS estimator is © [GB], where state vector defined as, and aftr Ora s where 1 and 0 are matrices of ones and zeros, respect The G matrix isthe partition of the measurement Jacobian corresponding to the AG position states, and ean be found innavigation extbooks ($4). The B matrix isthe partition of the measurement Jacobian corresponding tothe clock bias In some cases, the GNSS are not synchronized, which implies that each receiver had a different elock bias, denoted a dt, and Sty... f= spectively, To account for this, the state is adjusted to include the GNSS receiver's clock bias, namely Bt, a8 well as the difference between the LTTE receiver's clock bias and the first LTE. base stations clock bias, namely (hn ~ Ste). Therefore te state vector that is estimated becomes, = [FF CBligesC(Stray — Ata)" The matrix B is then adjusted in accordance with this new state to become a flyer Ova meliin] nd cellular receiver clocks The weighting matrix in the WNLS is chosen as inverse of the measurement noise covariance R diaglo2 2asees OS t \V._ SIGNAL RELIABILITY MAP GENERATION, The purpose of signal reliability maps is to find the reliable measurements that can be used by the AGY, and inform the path planning generator (discussed in Section VI) of these measurements. Therefore, signal reliability maps are generated only using information that is known a priori, i.e, before the path planning generator prescribes a path to the AG’ xis known a prior! includes 3-D building maps and other static objects in the environment. Information that is not known a priori includes pedestrians or other vehicles. Information t URBAN ENVIRONMENTS, 1567‘A signal reliability map is a spatiotemporal map spec- ifying for each location in the road network the GNSS. satellites to which there is expected to be blocked LOS and the pscudorange multipath error produced for each cellular base station. The following subsections define the signal reliability maps for GNSS and cellular signals and discuss their generation and storage. ‘A. GNSS Signal Reliability Map Generation GNSS signals suffer from multipath interference and NLOS conditions in urban environments. Multipath in- terference in urban environments is a dominant error source to which many mitigation techniques have been pro- posed [58}-60]. Receiver-based multipath mitigation tech- niques typically require the LOS signal to be received {61}, while more advanced techniques in NLOS conditions re- ‘quire specialized antennas and additional hardware (62 ‘The GNSS receiver used in the experiment in the sequel paper [18] used a posteriori multipath estimation (APM) mitigation technology to subtract multipath interference from GNSS measurements [63]. The proposed approach in this article will only use GNSS satellites that are classified ‘as having clear LOS. To this end, the signal reliability map for GNSS signals stores information about whether the LOS path between the receiver and satelite is obstructed. GNSS visibility prediction using 3-D building maps has ‘been proposed in prior literature [30], [31].GNSS reliability maps differ from prior work in the way that the visiitity informationis stored. Since GNSS satellite positions change with time, the GNSS signal reliability maps store the time intervals when a satellite is visible at a given location. The intervals are stored for each satellite and each location. Formally, the GNSS signal reliability map for a given satellite isa Sequence with P elements, where each element represents a Jocation in the road network. The environment consists of W transmitters, The signal ih satelite is Here, p represents a unique index corresponding to a part ular location inthe road network. Each Tens,» isasequence ‘of ordered pairs representing the start and end times for which the nth satellite has unobstructed LOS at location p. Tasyp = (Claas tnt For one day, there are a total of 7; time intervals with start and end times fy aNd fensp.7e Tespectively. The GNSS signal reliability map could be generated a priori and updated periodically (e.g., every day. whenever new cephemerides are available, or whenever the 3-D building map undergoes change). The signal reliability map can be stored locally at the vehicle or at a cloud-hosted database. Ata location p, the nth satellite has unobstructed LOS at time if aan pr £1 Stender any Lt. An Fig. 3. Visulization of GNSS signal ibility maps (a) Regio in Hit ~) where i] denotes the real part, 7 is the sampling interval, 0-<§ $055 isthe time shit in the LTE receiver's racking loop (€ = 0.5 is chosen in this article), 1. i the number of subcarrier symbols in the pilot (200 when the bandwith is 20 MHz and the cell-specific reference signal is used asthe pilot), and A is the signal power due to antenna gain and implementation loss {14], 68]. The theory for (8) and (9) is presented in {68}. The normalized symbol timing error 2 is set to zero to assume perfect tracking. Using (8) and ), the multipath interference y for all M cellular transmit is determined. Te multipath-induced bias is comprised of the multi- path interference 7, and the NLOS bias (ie. path delay between the first received path and the LOS path). That is, the multipath-induced bias is given by », (LE arene” Xn + €0(0) — dros, ao where dyes is the length of the LOS path, therefore, t,(0) — dos is the NLOS bias. If Ja,(x)] < «, for all x=0,....X 1, where « is a threshold, the LTE signal is rendered too weak to be tracked and the signal reliability ‘map assumes no cellular measurement at that location, For each cellular transmitter, the bias is stored for each location in the cellular signal reliability map. Formally, the cellular signal reliability map for the mth transmitter is a sequence with P elements Meat, opp where By. = when the mth cellular measurement is not received a the pth location, where @ denotes null. The sig nal reliability maps for M LTE transmitters are collectively referred 10 bY Mca) Fig. 4showsa visualization of acellular signal reliability map for a single cellular base station corresponding to the US. cellular provider AT&T in downtown Riverside, CA, USA. A raster feature is illustrated, where the black regions indicate that the pseudorange bias due to multipath at the pih location exceeds a threshold tpyay = 0.5 m, i bp = 05 ANNI 1369 URBAN ENVIRONMENTS,Fig. 4 Visulzation of the cellular signal ibility map. The Black segions indicate thatthe peudorange bias due to maltipat atthe xh locaton exceeds a threshold ug = 0.8 my Le yy > 0.5 oF there 510 ‘celular measurement ie Brey = © a the eh postion, This figures ‘obtained With ARIS 49), cor there is no cellular measurement, ie,, Bn,y = @ at the pth position. The threshold tiaax is the pscudorange bias threshold used in the path planning optimization problem ‘explained in Sections VIB and VIL C._ Thresholding Cellular Signal Reliability Map Analysis As the LTE wavelength is around 15 em, a change of 1 few centimeters in the receiver position can change the relative phase in the complex channel impulse response. The relative phase affects the multipath interference y». In prior ‘work, it was shown that areas where the multipath-induced bias exceeds a threshold can be clustered together [42] Therefore, a thresholded value of the multipath-induced bias is used instead of the exact simulated value in the path planning uncertainty-based constraint, discussed further in Section VI-B. This subsection studies the validity of simu- lating and using the thresholded cellular reliability maps. ‘The validity is studied over perturbations in the relative phase in the complex channel impulse response. ‘The following assumption is studied: if the simulated multipath-induced bias exceeds (or does not exceed) a threshold 7a them the bias with perturbed relative phases also exceeds (or does not exceed) rss: This claim is tested by perturbing the multipath interference, ic. the part of tte multipath-induced bias that is affected by phase, The relative phases in several simulated channel impulse re- sponses are perturbed, and the perturbed multipath inter- ference terms are compared to the unperturbed terms, The ‘complex channel impulse responses are calculated with Wireless Insite software, using 3-D building maps from downtown Riverside, CA, USA, shown in Fig. 3. Around 45 600 locations were simulated in the downtown area, For all locations that received more than one path, the relative phases for each path were perturbed according to a uniform, distribution from —180° to 180°, and 1000 perturbed mul- tipath interference terms are calculated per location. Five {ase f0F tng are considered in the results shown in Table IEEE TRANSACTIONS ON AEROSPACE AND ELECT ‘TABLET Results Feom Multipath Interference Simulation 241885 161601 14054912879, 210102 14127 OT ‘The metries in Table I are defined as follows [69] 1) Tive positive (TP) isthe number of perturbed terms that donot exceed fay, while the unperturbed terms also does no exceed Yon 2) Truenepative (TN) is the number of perturbed terms that exeeed fy, While the unpertirbed term also exceeds an 3) False positive (FP) isthe number of perturbed terms that do no exceed fa, while the unperturbed term exceeds a False negative (EN) is the number of perturbed terms thatexeeed rau, hile te waperured erm doesnt exceed ne ‘Table [ also provides the true positive rate (TPR) and “INSP ‘TPR corresponds to the percentage of the perturbed terms that are correctly identified as not exceeding tm. TNR corresponds the percentage of perturbed terms that are correctly identified as exceeding tjqas. It can be seen in Table I that for a threshold as low as I m, the percentage of terms that are correctly identified as being below the threshold is 91.07%, TPN VI. POSITION MSE AND UNCERTAINTY CONSTRAINT CALCULATION ‘This section describes the formulation of the optimiza- tion function and constraints used to generate the optimal path for the AGV to follow. The optimization function involves the position MSE, which is discussed in the first subsection, while the constraint involves the largest eige value of the position estimation error covariance, which is discussed in the second subsection, In what follows, the biased and unbiased error states are formally defined based on the measurement model and the estimator, Since the measurement model is nonlinear with respect 10 the state vector x,, the model is linearized according to Has, +0 where the Az, is the measurement error vector, which is the difference between the measurement veetor zy and its estimate 2; Ax, © x; — 8), i.¢., ax, isthe estimation error, which is the difference between x, and the WNLS estimate (ONICSYSTEMS _VOL.57,NO.3_JUNE2021Sp5and vA [gn tee oy Yen Ns Ulta e +s Weed]. To ae alyze the effect of multipath-induced bias, a deterministic bias b i introduced in the measurement Ac, Ag +b Acyl" is the unbiased measure- iment error vector. ‘The effect of the pseudorange bias on the position estimate can be found through the normal equation (see, for example, [70, (7.67))) GER HTR Az, GRAY 'HTR (Az, +5) (ATR Hy 'HTR Ac ARH) HRD. Ax, Therefore, the bias in the pseudorange introduces an addi- tive bias in the estimation error according 10 Baty = AK, + Xn eqs €82,1° results from the multipath- the measurement, and Ax) & [Ar c&&]? is the unbiased state estimation error. ‘The vector && represents the vector of clock bias states ‘Therefore, the unbiased stae estimation errr and the state bias can be, espectively, expressed as Ax, = (HRW) HR Ac! Sree = (RH), where x induced bias in an ‘The following subsection explains the steps for using the GNSS and cellular signal reliability maps whose generation was described in Section V to calculate the position MSE for the path planning cost function, and calculate the largest ccigenvalue of the position estimation error covariance for the path planning constraint A. Position MSE ‘The position MSE is a scalar measure, which accounts for the precision and bias of an estimator [71], and is ‘commonly used duet its mathematical tractability Itreers to the mean of the squared estimation error inthe posi a a specified location and time, i. 3:D position M! anlar] [E[Ar, Art} ] Ste [BUAL, + Fic MAF, +r)" =e arid") | + In fy (12) where Ef-] denotes the expected value and tr-] denotes the trace, and (12) follows from A being. zero-mean, The positon bias is Ix obtained from the firs three clements of ta. The covariance ofthe unbiased position ‘error srelated tothe weighted-positon dilution of precision (WPDOP) according to [72] wevor 4 yfr[eotar.l] RAGOTHAMAN ET AL: AUTONOMOUS GROUND VEHICL PATH PLANNING. fi tld +h where i; is the jth diagonal of (HTR-'H)-! The calculation of the position bias [fra,ll2 due to multipath uses the simulated LOS and pseudorange bias due to multipath, which were found in the signal reliability ‘maps. "The steps to calculate the postion MSE are described next. 1) Step 1. Calculate the Vector b: For location p and time f, there ate W < NN reliable GNSS measurements as determined by the signal reliability maps such that for each n= 1,...., where ¢ satislies a time interval de- fined in Tyus,.p- Also, there are Mf = M reliable cellular measurements forall m = 1,...., such that By,» is not null and [®jp.p| Stine Where my is the mth element of Ry “eggs Where R, is the Cholesky factor of R, ie, R= RIR,. The method for calculating the threshold ‘ima is shown in Section VI-B. The pseudorange bias vector is = 10,6. 1 p.---- Duy pl" 2) Step 2. Calculate the Jacobian H: The rows of H are calculated from the transmitter positions of the corre- sponding elements in band the coordinates of location p. Its assumed thatthe biased postion and true postion are close enough so that the measurement Jacobian forthe true position is lose to that of the biased position. 3) Step 3. Calculate the MSE: The position MSE is calculated from weDor* 3-D position MS TThe position MSE at a particular position p and time f, de- noted MSE(p, 1), will be used in the path planning algorithm. described in Section VIL B. Uncertainty Constraint Calculation This subsection describes the calculation of the path planning constraint on the largest eigenvalue of the position estimation error covariance. The purpose of this constraint is to restrict the AGV’s path to be within the maximum. position uncertainty. To this end, the largest eigenvalue of the position-estimation ertor covariance will be used, whi specifies the length of the largest axis of the uncertainty ellipsoid [73]. The largest eigenvalue ata particular position and at time f, denoted yu«(p, 1), is found from the upper 3 x 3 matrix block of (H™R™'H)-!, where His calculated according to the method discussed in Section VI-A. This constraint is also related to a conservative upper bound on the position bias, which can be derived from the expression «aay where LURAN ENVIRONMENTS 1s71é =", G=a- Bre" r=@"B) R,"G B-R,"B ‘The derivation of (13) is given in Appendix A. The bias corresponding to the mth cellular measurement is con- strained such that [by,o| jm where jy, is the mth el: cement of RET fx i7aws and Tax is the pseudorange bias threshold, The constraint can also be written as |b] < Aegan imax» Where | ‘of each element in the vector ‘Subsequently, the upper bound on the position bias can be found accord | corresponds to the absolute value = maximize |(@ as [|], that maximizes the objective function lies on the boundary of the constraint). Equation (16) is found through change of variables B’ 2 Equation (17) follows from the definition of the largest singular value. Equation (18) follows from the relationship between the largesteigenvalue and the largest singular value ‘ofa matrix. The term Apaxl(G"G)-'] denotes the maximum. eigenvalue of the position estimation error covariance. The upper bound in (18) shows that the maximum eigenvalue also relates tothe position error. a} C._ Simulated Receiver Spacing ‘This subsection studies the impact of spatial sampling ‘on the position MSE, defined in Section VI-A. The position MSE was caleutated over a 209-m path based on simulated LTE and GPS reliability maps. The GPS reliability map was generated on September 20, 2020 at 11:20 UTC. Otherwise, the simulation environment and settings tothose defined in the companion paper [18]. Fig, 5 shows the sulated position MSE over various simulated receiver position spacings. The figure also shows the position root- ‘mean-squared error (RMSE) for each spacing, calculated as, the root of the average position MSE along the path. 8 eenean random io Fig. 5. Postion MSE over simulated receiver positon spacings f(a 1m. ¢b) 8m, (€) I mand (d) 05 m. A spike in postion MSE occurs between distances ISIS and 155m and is eaptred in (b)-( It can be seen that there is a loss of resolution in the high spacing cases, where the position MSE appears to bbe smeared compared to the lower resolution cases. In Fig. 5(d), around 150 m, there is a 3.5-m section of high position MSE that is not captured in Fig. 5(a). Despite these Aifferences, the position RMSE averages to a similar value across all spacings to within 0.3 m, The following conclusions can be drawn as they pertain (o path planning metrics. When using position MSE as a path planning cost function, a high spacing (10 m) suffic as can be shown by the closeness of the position RMSE over the four spacing cases. As a path planning constraint, smaller spacing (1 m) would be preferable to capture the spikes and locations with too few measurements. Methods for improving the sampling method (e.g., based on the amount of surrounding building features) ean be explored in future work VIl._ PATH PLANNING GENERATION ‘The path planning generation step utilizes the signal reliability map to prescribe an optimal path for the AGV to follow. This section describes the steps to determine the ‘optimal path between a start position ata desired departure time and a target position, The optimal path is on accounts for the shortest path length and the position subject to & maximum tolerable uncertainty (as measured by the largest eigenvalue of the position estimation error covariance). To account for both position error and path length, the ‘optimization cost function is chosen 10 be the sum of the position MSE along the path, multiplied by the distance ‘between two adjacent points. The distance is explicitly eon- sidered in the cost function because only includ MSE could result in lengthy paths, e.g., paths that requi the AGV to leave and reenter the urban environment. The ‘optimization function constraints account for the position IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONICSYSTEMS _VOL.57,NO.3_ JUNE 2021mg Rei a Fig. 6, Flow chart of signal reliably map generation, position MSE. nd uneerainy consti caleulation, and path planning generation, bias due toceTlular multipath as well as uncertainty about the AGV'sposition estimate. The user-specified constraints are threshold for position bias Fx and threshold for position unceriaimy Zag. The threshold Zypay is used asa constraint for all points p and time r along the AGV's path, ie Jmon(#) Fae: The theeshold Zax is also used along with Fux (0 calculate a threshold on the pseudorange bias ‘ina: The calculation of gas can be achieved from (18) by substitotin specified Fy and Using Fy in place (of Ppaal(G'G) "| Since 4 < M, Mf is replaced with M to calculate an upper bound that is independent ofa particular location and is valid for the entire e ‘The path planning generation block solves a con- strained optimization problem, discussed next, and returns the AGV's prescribed path along with a list of reliable GNSS satellites and cellular base stations to use along the path, AS the AGV traverses this optimal path, it only uses signals from these reliable GNSS satellites and cellular base stations. Note that to make the WNLS estimation problem observable, there needs to be atleast reliable GNSS satellite signals to estimate x or +41 > Sreliable GNSS satellite and cellularsignals to estimate x, © [FF €81,.c8f,11". with Af > 1. Fig, 6 sum. marizes the flowchart of signal reliability map generation, position MSE and eigenvalue calculation, and path plan hing generation with the corresponding inputs and outputs in Fig. 7 shows the calculation of dist) when the point is not adjacent 10 an intersection, Ifthe street has multiple lanesin one direction, 8 location index is assigned 10 each lane, and dist(p) is shown divided by the number of lanes ned the dist ‘The path planning optimization problem is formulated next, Formally, a path from the start to the target location is denoted x € P, where P is the set of all paths. The path RAGOTHAMAN ET AL: AUTONOMOUS GROUND VEHICL PATH PLANNING. ae] Ee t | ! Tease] [ecioome] [itp || inet nesim| [eamin | [pants | UO Fig. 7. Steps wo caeulate stp» for 3 points whose indies ae and py Fr p, Which is adjacent oan intersection, te 3D midpoint between py aap i clcultd, then dsr i the distance etc the midpoint and the ttrscton center. Fo p, which isnot adiacent wo an fnterscton, the midpoint between ps an p> feat, then the ‘midpoint between p; and p is calculated. The, stp: sthe distance etween he two calculated midpoint 1 is composed of a sequence of position indi the start position index p, and the target p, (Br Pie Paes Pa ‘The optimization problem is expressed as subjectt0. gel 1) S Sia Urns = Fas ay) ‘The second constraint in (19) can be relaxed wsi ld the optimization problem cas) re Yds) MSEXp.0) subject to Ayyas(Pot) <2 Bis 0 Winx Note that the cost function in (20) accounts for both the position MSE and the path length, Other cost functions could be used to favor either part: position MSE versus path length, e.g. using an exponential for either term. The optimization problem (20) resembles the problem of finding the shortest path in a weighted graph, where the roads are the edges of the graph and the path planning metric determines the weight of each edge. Several algorithmshave been proposed to find the shortest path that use dynamic programming [74]-[76]. These algorithms cannot be used because the problem lacks optimal substructure due to the time-varying nature of the costs and constraints in (20), The following subsections show two possible approaches {or prescribing a path (o the AGN. ‘A. Path Planning Generator: Approach A, ‘To simplify the optimization problem and use dynamic programming for the path planning algorithm, it will be assumed that the GNSS reliability map from the depar- ture time through the time the AGV arrives at the target position is time invariant, making MSE(p, 1) = MSE(p) And Angx(P,) = Amox(P)- This assumption is reasonable for short AGV paths during which the geometry of the LURAN ENVIRONMENTS srsGNSS satellite constellation does not vary significantly. For ‘example, a GPS satellite moves less than 0.152 radians in 15 min (see Appendix B). An additional check is performed to ensure that the GNSS reliability map does not have shatp changes (ic, oss of a satellite) as the AGV the path, in which case Approach B is to be implemented instead. This check requires the assumption that the vehicle travels at constant speed vagy Based on the aforementioned assumption, dynamic pro- ‘gramming can be used in this approach, Among exist- ing path planning algorithms, Dijkstra’s algorithm is most widely used and is recognized as a classic algorithm to find the shortest path between (wo arbitrary nodes of a weighted graph [77], Dijkstra’s algorithm is readily implementable with acceptable complexity; thus, itis adopted in this ap- proach to solve the AGV path planning problem. ‘Assume that the AGV is driving in a region consisting of intersections and «roads (e.g, side streets and highways). an be modeled by a graph G = (1, «»), which ‘consists of 1 nodes and a edges. Each intersection corre- sponds to a node, and each street segment that connects ‘bv intersections corresponds to either one or two edges. In the experiments in the companion paper [18], the graph is ‘constructed using the following rules. ccutes 1) One-way street segments are assigned one edge with cone direction, 2) Two-way street sogments are assigned two edges with one direction for each way. ‘The path planning metric f(f, ) assigns a nonnegative real number corresponding to the weight of the edge from nodes f to a in G. Based on the objective function in (20), the weight is given by the position MSE at all points from nodes f and a, denoted P(B, c), multiplied by dist(p), ie, f(p.0) = SD disp) MSEXp) Ie is assumed thatthe deviation between MS MSE calculated atthe tue time i small Based on the consrins in (20) ay p)enc804s hy frp € PCB.) then the edges removed fom te graph Dilksra's algorithm is implemented as follows. Lets indicate ue start node a which the AGV start, and let ¢ indiate some target node. Let dg) dente the cox along the path fom tog let S denote the et of edges that have already been evaluted bythe algoriim, and let V denote the st of uviited nodes. Within a path y(@) € G, denote t%, a5 he prdecessor ofa and, asthe predecessor of TD sure 74g) does not have los ofa stele the algorithm Tracks be time of uve from he source tote creat node, denoted 15,4. Approach A is inlized as follows (p) and the 1) ds) =05 2) r(s) = 05 3) foreach node adjacent tos, set d(ar) 4) foreach node a such that er # s andar in not adjacent tos, set d(a) = 2: (s,@) and ‘Algorithm 1: Path Planning Algorithm, Output: dg) and ye) 1: Finda € V that minimizes da) 2: For each adjacent to a 3: Wdia)+ f(B,a) Tan for p © PCA) and for all == 1,1, ¢. The consrtint ona(°) isthe maxima eigenvalue metic along the entire path A° based on the constrains in 20) Alter the aforementioned initialization, the path plan- ning algorithm outinedin Algorithm 2, where A"(Ais used to denote the kth node in path ‘The computational complexity of Approach B is (C0), which is higher than the computational complexity of Ap- proach A, O(7) [78]. Another disadvantage for Approach Bis that the prescribe path is not guaranteed tobe optimal like in Approach A. However, fr trajectories with long du ration, Approach B will be more coreet because it accounts for satellite motion, Choosing Approach A over Approach B for paths with long duration can lead to worse navigation performance than expected, or ignoring measurements that ean be used. This point is diseussed further in the sequel paper {18} RAGOTHAMAN ET AL: AUTONOMOUS GROUND VEHI 1E PATH PLANNING IN URBAN ENVIRONMENTS, ‘Algorithm 2: Path Planning Algorithm, put Tog Sn Output: die) and ys) H 3 20g, 4: Create local copy of G: B = (} 3: Fork =1tolen(A)—1 6: ") Yropan, = AO: KY & 0 Fojetic—1 9: Yeap, = AMOK] 10: Tf Yeon, = Yroospan 1 In G copy, fil@spar, All + 1]) = 00 2 End it 13: End for 14: Yar = Dijstas(G, ter 8) 15: BB+ (Yaga + Yow 16: End for 7: Sort 18: HAgay(B) < Saas and (8!) > a) 19: Bh dig) = fb) 20: 2: Ae A+B! 22:_End for Vill, CONCLUSION This article considered the problem where an AGV equipped with GNSS and cellular receivers desires (o reach a target location while taking the shortest path with min- imum position MSE, while guaranteeing that the bias in the position estimate and the position uncertainty are below desired thresholds. Parameters of the cellular pseudoranges related to the transmitter clock bias are estimated in an ini- tialization step in an open-sky environment. A path planning ‘gencrator prescribes a trajectory that satisfies this objective a 3-D building map to create signal reliability maps for GNSS and cellular LTE signals. The signal reliability maps are used to calculate the position MSE and uncertainty based constraint at each location, which in turn is used to generate an optimal path for the AGV to follow. In art Il of this study, extensive simulation and experimental results are presented demonstrating the improvement in the position RMSE by employing the proposed framework and the consistency between the simulated results with those obtained experimentally on a ground vehicle driving in downtown Riverside, CA, USA. This article demonstrated the feasibility of using path planning to improve GNSS and cellular positioning, and there are opportunities for future work. Different methods of spatial sampling can be explored, where location indices are distributed based on properties of the road network, Other of additional information can be stored in reliability ‘maps, such as measurement noise variances as a function of simulated carrier-to-noise ratio. Other error sources can be introduced in the path planning cost and constraints,such as GNSS multipath-induced bias, or inaccuracies in the building footprints that have been modeled in prior work [80]. Methods for GNSS and cellular visibility predictions ‘can be improved through comparing signal availability and received signal strength to 3-D map predictions using power ‘and shadow matching [29], [34], orusing a fish-eye camera todetect GNSS NLOS [33]. The same costs and constraints ‘can be evaluated with different navigation frameworks that use filtering techniques and sensor fusion, Integrity moni toring techniques [47] and map matching [48] can also be used (0 improve transmitter selection or augment the path planning cost function, Time-varying path planning in ap- proach B ean be improved by using traffic information [81]. APPENDIX A RELATIONSHIP BETWEEN PSEUDORANGE AND POSH TION BIAS ‘This appendix establishes the relationship between posi- tion bias and pseudorange bias [ef.,(13)] For this analysis, itis assumed that the biased position and true position are ‘lose enough so that the measurement Jacobians evaluated ‘each are approximately equal. From (11), the relationship between pseudorange bias and the state bias is given by Aron 2 TR) TR Consider the Choleky factorization of R = RTR,, and de- fine HE 2 R, "Hand b © R,", which results in aya. Bice en Recall that H& (G, B] and jo © (rho. 681 fore, (21) can be partitioned as [ac}-[ee Se] [E> An expression for 7; can be found through block matrix 37s there- where T= (BB)! After rearranging (22), the relation ship is found 10 be OG CBB G GT EBB (Gd BBG GTA — BRB where (I~ BIBT) is a projection matrix. The aforemen- tioned equation can be rewritten as Fig.8, Description ofthe variables used to approximate he change in APPENDIX 8 CHANGE IN GNSS SATELLITE GEOMETRY This appendix provides an approximation of the change insatellite geometry asa function of time. For aconservative approximation, the GNSS satellite is assumed to start at zenith, where the angular velocity from the surface of Earth is fastest. GNSS satellite angular velocities, denoted «,, can be used, For example, for a GPS satellite , © 1.522 radi, For low to medium orbit altitudes, the magnitude of the tangential velocity of the satellite in the BCEF shows small variation and can be approximated by a constant [82]. Ifthe satellite starts at zenith, the amount of time Tops it takes for the GPS satellite to travel @ radians can be approximated by Toes * cos 0, — oe cos rors where ray ©6371 km is the radius of Earth, ap ~ 7.2921 x 10° radis isthe Earth's angular speed, / = 55° is the inclination angle of the GPS satellite, and raps © 26571 km is the altitude of the GPS satellite, Equations (23) can bbe used to calculate the time it takes forthe satellite to travel AO radians (see Fig. 8) REFEREN U1} 44 Corporations Working on Autonomous Vehicles, CBlsights, Now York, NY, USA, May 2017 PIT Caen Self Driving car acidems will keep happening, We need 4 lear from them Mar 2018. [Online]. Available: npsfmoney can comy20180421Aechnologyelf-rving ca ‘opiniones nl se—fhmoney032118el- diving care ‘opinion 63 AMStory [3]. foannides,T Png, ana G, Gibbons Known vulnerabilities of global navigation satellite systoms statu and pent mitigationtechniguce Proc. IEEE, vol 104, no 6, pp. 1174-1194, Feb. 2016 EH] N.Zhu,J. Mass, D, Betille and M, Berbinea rey leety ‘GNSS postion integrity in urban environments: A review of tron = (OGY 'CD tients . a IEEE Trans tell. Transp. St, vo. 19.0.9. where G 2 = BIBNG, Sep. 2018, 1576 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL.57,NO.3. JUNE 2021i to Is el 0} un 12) ust bay ust 6} ust 9) RAGOTHAMAN ET AL 4. Pett, BStotelua, M Fei and F. Kars] Remote achs on automate vehicle sensors experimen on camera and LiDAR In Proc Black Hat Europe, 2018, vol. Upp. 1-13 . Boro and C. Outsell Design of general pscudolite pulsing scheme IEEE Trans. Aerosp. Electron. St, Wo. 50, 0.1, pp. 2-16 Jan. 2014, Yang and. Nauyen ‘Tracking and eve positioning with mixed signal of opr J Inst aviga. vl 62, no 4, pp. 291-811, Bee. 2015. 2, Kass, J. Khali, K. Shama nd J, Morales [ar therefore I know where Tame Compensating for GNSS Timitations wath ela signals IEEE Signal Process. Mag. v0 Sep. 2017 Yang. T Nguyen, and E- Blase Mofile positioning va fasion of mixed signals of oppor IEEE Aerosp. Elecnon. Syst. Mog. ol. 29, 0.4, pp. 34-46, Aap: 2014, M. Drivsso, C. Marshall, M. Sabathy, F. Knut, H. Mathis, and F Babich Vehicular positon ticking sing LTE signals IEEE Trans. Veh. Teclnol,, Wl. 65, 10. 4, pp. 3876-3391 Ap. 2017 J. Khalifeand 2. Kassas ‘Navigation with cellular CDMA signals—Par I Peformance nasi and experimental results IBEE Trans, Signal Process 0. 66,00, 8, Pp. Ape 2018 AK. Shama, J Khlife, and, Kass Exploiting LTE signals for navigation: Theory implements IEEE Trans. Wireless Commun vol. 17.90. 4 pp. 2173-2189, ‘Ape: 2018 J. dal Poral Rosado, 1 Sevo-Grandos Postion accuracy of join time-delay and chanel estimators in IREE Acces, vo. 6, pp. 25185: XK. Shamae and 2. Kassss LETE receiver design and multipath analysis for navigation in urban envitoaments Ins Navigt, 0. 68, n0 4, pp. 5-675, De. 2018 4. Kale and 7. Kass Precise UAV navigation with clllar cartier phase measure In Proc. IEEENON Poston Ap. 2018, pp. 978-989 . Shame al 7. Kassas Subsmeter accurate UAV navigation and cycle sip detetion with LTE carr phase In Proc. ON GNSS Conf, Sep. 2019, pp. 2469-2479. S, Ragorhaman Path planning or autonomous ground vehicles using GNSS and celular LTE signal elibility maps and GIS 3-D maps Master's thesis, Uni. California, Riverside, CA, USA, 2018. S, Ragothaman, M, Muet, and Z. Kasse ‘Autonomous ground vehicle path planning in urban environ sen using GNSS and cellar signals reliability mops: Sim Tation and experimental results IEEE Trans. erp, econ. Sys 0 be published. A. Davison Realtime simultaneous oalsation and mapping with single in Pro IEEE In. Conf Comput. Vis 2008, vol. 2, pp 108. ui, 34, no. 5. pp. 194, 204-2218, Lipor-Saleodo, F. Zanior, and G. $199, Apr. 2018 Location, Navigate. Symp. AUTONOMOUS GROUND VE 0 1s ea ps) 26) er 29) 0) Bn (3) oa) 151 PATH PLANNING IN URBAN ENVIRONMENTS RR. Valencia, M. Mona. Andrade-Coto, and J. Porta Planing reliable paths with pose SLAM IEEE Trans. Robe. ol. 29.0.4. pp. 1050-1059, Aug. 2013 M. Mora and J Toemera Predictive and malate sensor-baed planning ander uncer IEEE. Trans So, 2018, (C-Hubmann, J. Sch, M. Recker, D. Althoff and C. Stier Automated deving in uncertain envionment: Plansing with interaction snd uncertain manewver prediction IEEE Trans. Imel. Ve, vo 3,0, pp. 5-17, Ma. 2018 Zheng, Zhang. and UUs Reliable path planing forbus networks considering teavel tone IEEE ivell. Trump. Syst. Mag san. 2016, Tang and U. Ongunee “Motion planning formaltitarget sreillance with mobile sensor IEEE Trans, Robo, vol. 21, n0.5, pp. 898-908, Oct, 2008, A. Som, T. Mazauch and 8. Sarkan ‘Addeesing neering in UAV navigation deision-making IEEE Trans. Aerosp. Electron. St. vl. 84,00. Lp. 295-313, Jan, 2008 X. Zh, X. Yo and X. Peng, 'UAV collision avoidance based on varying els strategy IEEE Trane. Aerosp Electron, St, Ol. 5S. m0, 4 pp. 1783-1958, Aug 2019, 1. Gici, J Farell Z. Kas, an M. Ouimet ‘Autonomous surface vehicle mulsep lok-ahead| measure ment location planning fo optimal localization of underwater cous transponders IEEE Tron. Aerosp Blecixm, pp. 2836-2849, Dec. 2018 2, Kass nT Hampireys Receding horizon tijectory optimization in opportunistic nay igaton emironmens IEEE Trans Aerosp Electron. Syst. W0.51, 90.2, pp. 866-877 Ape 2015, '. Saab and. Kassas Power matching approach for GPS coverage extension EEE Trans. Intell. Transp. Sy, vo. 7.80.2, pp. 156-166, Sun, 2006 Yee, ¥. Suh and R. Shibasaki Ajax GIS application for GNSS availability simulation J-Chil Eng, yl. 11, mo 6, pp. 303-310, Nov. 2007 Bradbury, M. Ziebart, P. Coss, P.Boulion, and A, Read ‘Cade mapa modeling nthe urban envionment sng large nual elt city models: Determining the local environment J. Navigot, wo. 60, no 1 pp. 95-105, 2007 4Lisaacs, A ish, F-Quitn, U. Madhow, and. espana Bayesian localization and mapping vsing GNSS SNR measure: in Pre. IEBEAON Position, May 2014, pp. 445-4 "T, Suaoki and N. Kubo 'N-LOS GNSS signal detection using fist-eyecumera for vehicle navigation in urn enitonments In Proc: ION GNSS Conf. Sep. 2014, pp. 1897-1006. 1. Wang, Groves, and M. Zchact Smariphone shadow matching for beter cross stcet GNSS Psion in whan ernment I. Navigat vo. 69, no, 3. pp. 41-433, 2015, 1. Hsu ¥-Gu_ and. Kamijo 3 building model-based pedestin positioning method using [GPSIGLONASSIQZSS and its selisbity calculation GPS Solon. vol. 3, pp. 413-428, 20 2016 Intell, Wo, vol. 16, m0. 3, pp. 1493-1508, vol 8, m0. pp. 35-50, Spit, vol, 58. m0. 6, Location, Neavigat. Somp. sibo) bo) 0 co wa ss} 19 Iso) iss J. saa ea ‘GPS-opimal micro sr vehicle navigation in degrade environ in Proc Amer Control Con, an. 2014, pp. 1864-187 S. Ackermann, A. Ariana, 8 Pizzo, S, Glione,C STi a, and Digital surface moves for GNSS mission planning in critical 2 Surveying Eng. v0. 140, no. 2, pp. 1, 2013, AL Nowak Dynamic GNSS mission planing wing DTM for precise av sation of autonomous vehiles J Naviga, sl. 70.20, 3, pp 483-804, May 2017 A. Shetty and G. Gao Predicting tate certsint for GNSS- based UAV pothpaning using stochastic reachability in Pm. ION GNSS Cont, 2019, pp. 151-139 1. Karimi and D, Asavasutirak ‘Anovel optimal ating for avigntionsytemsservices bined ‘on global navigation satellite system quality f service J tse. Transp, Syat vl 8, wo, 3, pp. 286-298, May 2014, Zhang and L. Hou ‘Anes path planning algorithm wing a ‘ap for UAVs na ram re J Insel Robie Set, vo. 94, pp. 219-288, 2019, S.Ragothaman, M. Maaef, and Z.Kassas Multipth-opimal UAV uajeetory planning for whan UAV ‘navigation with cellular signals in Proc IEEE Yeh Technol. Conf, Sep. 2019 9. 1-6 . Shami J. Mors, and. Kansas {framework for navigation with LTE time-corclated pseudo: "ange eros in makipathenironments in Pro. IEEE Ye, Technol. Conf, Apt. 2019. pp. 1-6. IM. Uimschnider and C. Gentner Multipath assisted positioning for pedestrians using LTE signals in Proc. IEBBION Postion, Location, Navigat Simp Ap. 2016, pp 386-392, A. Abi and Z. Kass Evalution of fedhack and feedforward coupling of synthetic perre navigation with ETE signals in Prue. IEEE Yoh Technol. Conf, Sep. 2019, pp. 16. F Pere, D, Betile,P Carlin, RTledo-Moreo, A. Gomez Skanmeta, and M, Oni {GNSS autonomous leaization: NLOS salve detoston based fon 3D mane EEE Rob, Autom, Mag. vol. 21, 01.9. $7-63, Mar: 2014 M. Mauret and. Kassas Measurement characterization and autonomous oui detec tion and exclusion for ground vehicle navigation wih ella signals BEE Troms. uel Wee, vo 5,0. 4, 9p. 670-683, Dec. 2020. 2. Kasay, M. Maaef, J. Meals, J. half snd K. Shari Robust veel localization and map matching in wba emi ronments dvough IMU, GNSS, and collar signals BEE. toe. Transp. Sst. Mag. Yo. 12, 90. 3, pp. 36-82 Sun, 2020 SRE [Online Available: hups:wwwaregs.com. 2. Kasse, Ve Ghaok nd T. Homphreys Adaptive estimation of signa of eppority in Proc ION GNSS Conf, Sep. 2018, pp. 1679-1689, J. Morales and 7. Kass Optimal collaborative mapping of teresal wants: Re ceiver placement and peformance characterization IEEE Trans. Aerosp. Elec Syst. vol SA, 00. 2, 1p. 992-1007, Ape 2018, $$ localization eror IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYS! 153) (sa) isa) (551 [561 (7) (ss) (59) (oo) (ot) (oa (on) (6s) (66) on 30PP Evolved universal terest radio oes (E-UTRA oie rent for support of radio resource management Sed Generation Partnership Proje! (3GPP), TS 36.133, Ape mio, 4. Khalife and 2. Kassas Evaluotion of elative clock stability in cellular networks in Pre JON GNSS Conf, Sep. 2017, pp. 2584-2559, Misra and P.Enge Global Positioning System: Signals, Measuremens, and Per formance, 2nd ed, Lincoln, MA: Ganga-Jamuna Press, 2010 7. Rassas and. Hunphteys ‘Opserailiy analysis of collaborative opportunistic navigation with peeuderange measurements IEEE Trans. Inet. Tranyp. Sot, wo, 15, no. 1, pp 260-273, Fob. 2014 4. Rte, K, Shama nd Z. Kasse ‘Navigation with celllar CDMA signsls—Par : Signal tod cling and sftware-detined reeiver design IEEE Tans. Signal Process. sol. 66, 0. 8, pp. 2191 Ape 2018 4. Rhulfe and. Kaas ‘Opportunistic UAV navigation withcrser hase measurements ftom asynchronous cellular signals IEEE Trans. Aero Elec pp. 3285-8301, Aug. 2020, N.Ziedan Urban positioning acurscy enhancement uilzing3-D build ings mode and acelerited ay tracing algorithm in Pre. 1ON GNSS Conf, Sep, 2017, pp. 3253-3268, IN. Kbayer and M.Sahmoudi Performances analysis of GNSS NLOS bias conection in urban ‘environment using tre- dimensional city model and GNSS Simulator IEEE Trans. Aerosp. Electron pp. 1799-1814, Fe 2018, 4. Lesouple, Rober M,SahouJ. Toumeret and W. Vigneau “Multipath mitigation for GNSS postaning in an urban enieon iment using sparse estimation IEEE Troms. sel. Transp. St, v0. 20,04, 9p. 1316-1828, Ape 2018 N. Jak, A. Vervisch- Pio, and N. Samana ‘Mulipath insensitive dela ek loop in GNSS receivers IBEE Trin. Aeros. Elecirn. St vol. 47. 00. 4 ‘2590-2619. Oc. 201 Heng, T, Walter, P Eng and G, Gao GNSS tmulipath and jamming mitigation using high mask ‘angle anennas and mutiphe constellations TELE Trans. Intel. Transp. Spe, v.18, 90. 2, pp. 781-750, Ape 2015. 4. Sleewacgen and F. Boon Mitigating shore delay multipath A promising new technique in Pro. JON In och, Meeting Con, Sep. 2010, pp. 204-213, _D. Roongpiboonsopit and H. Karim Integrated GNSS QoS prediction for navigation services in Pr ACM SIOGSPATIAL In, Morkshop Comput. Transp. Se Nov. 2013, pp. 78-78 1. Wang P. Groves, and M, Ziebart ‘GNSS shadow matching: Improving ban postoning acumey ‘using a 3D city mode! with optimized visibility sering scheme Icfnve Naviga, vl 6, no 3, pp 198-207, 2013 (Mc Huang and W. Xu Enhanced LTE TOMOTDOA extination with tating path detection in Proc. TEBE Wireless Commun. Netw: Conf Apr. 2013 pn. 3992-3997 Remncom Dec. 2018, [Online], Available hups/fwwosemcomcom 1208, Syst. vol $6 m0. 4 Syst, vol Smo 4 MS VOL.57,NO,3. JUNE2021(6s B, Yang, K. Lette, R, Cheng and Z. Cao Timing recovery for OFDM transmission 51 P.Han,N. Nilsson, and B. Raph! A Toma has or the heute determination of minimum cost IEEE J. Sel. Areas Commun, so. 18, 90.1, pp. 278-2291 pate Now 2000 IEEE. Trans. Sit. Sci. Cyberm, vol 4, mo. 2, pp 100-107, 162] Hastie, R. Tibshirani, and J. Friedman Jul, 1968 tems of Statistica! Learning: Dasa Mining ference, and [16] Dijkstra Prediction. New York, NY. USA: Springer, 2009 ‘A note on to problems in connexion with graphs 170). Kaplan and’C, Hegarty Namersche Mathematik, wl. no. pp 269-271, Dee, 1989 Unterstanding GPS: Principles ond Applications, Zale Nore 77] A. Zhang eta. ta MA, USA: Artech Howse, 205 Vehicle outing in ueban areas hase on the oil consumption Ir) S.Kay ‘weiht-Difhstra algoritm Fundamemals of Statsical Signal Processing: Estima TET tell. Transport Sit, vol. 10, m0, 7, pp.495-502 tion Theory. Upper Saddle River, NJ, USA: Prentice-Hall, Sep. 2016 1095, (78) 1.8 172) D. Won An algorithm for finding shortest ots from ll yource nodes 1a. Weighed DOP with consideration on clevaton-dependent toa given destination in general networks rings emer of GNSS saelites Quart App Mats. vo). 27.90, pp. 526-S30, 1970 IEEE Trans. lasrum Meas. Yo. 61, 0.12, pp. 3281-3250, [19] Open Steet Map Foundation (OSMF) [Online] Available: bps Des. 2012 wow pests one 73] Kasse A. Argpostais, and. Humphreys [80] H.Fan,A.Zipf,Q. Fu, and P.Neis Greedy motion planing for simllaneows signal landscape ‘Quality assessment for building footpins data on Open mapping and rctver localization SwectNap IBEE J. Sel Topies Signal Process, 0 9, 90.2, p. 287-288, In J. Geographical If. Sc, ol. 28, 0.4, pp. 100-719, 2014 Mar. 2015, (81) EL Nikoove 174) 0.6. Swathi and S. Hemamlint High-performance heuristics fr opiizaion in stochastic Prime-ided Distr algorithm for adaptive protection in i fic engineering problems ‘ropids in Pro. Large-Scale St. Comput, Jun, 2010, p. 382-360. IEEE Trans. Emerg. Se Topics ower Blectron.ol 4.0.4, [82] LAN. ALDhuhi, and J Hershey pp. 1279-1286, Dee, 2016 Sonya Ragothaman (Member, IEEE) received t engineering from the University of California respectively. oppler characterization for LEO stlites IEEE Trans. Coma, sl. 6,90. 3, pp. 309-813, Ma, 1998 c BS, and M.S. degrees in electrical Riverside, CA, USA, in 2016 and 2018, She isan Engineer with Aerospace Corporation, El Segundo, CA. She was a Researcher with the Autonomous Systems Perception, Intelligence, and Navigation Laboratory, from 2016 to 2018. Her research interests include software-in-the-loop model development for launch vehicle fight software, high fidelity simulation of global navigation satellite systems, and satellite pose estimation. Mahdi Maaref received the B.S. and M.S, degrees from the University of Tehran, Tehran, Tran, in 2008 and 2011, respectively, and the Ph.D. degree in electrical engineering from ‘Shahid Beheshi University, Tehran, in 2016, He isa Principal Engineer with OneNav, Palo Alto, CA, USA, and was a Postdoctoral Research Fellow with the University of Califomia, Irvine, CA, and a member of the Autonomous Systems Perception Intelligent and Navigation Laboratory, from 2016 to 2020. He was a Visiting Research Collaborator with the University of Alberta, Edmonton, Canada, in 2016, His research interests include autonomous ground vehicles, opportunistic perception, and autonomous integrity monitoring RAGOTHAMAN ET AL AUTONOMOUS GROUND VEHICLE PATH PLANNING IN URBAN ENVIRONMENTS 1579‘Zaher (Zak) M. Kassas (Senior Member, IEEE) received the BE, degree in electrical engineering from the Lebanese American University, Beirut, Lebanon, the M.S. degree in electrical and computer USA, and the MS.E. degree n aerospace engineering and the Ph.D. degree in electrical and computer engineering from the University of Texas at Austin, Austin, TX, USA, He is an Associate Professor with the University of California, Irvine, CA, USA, and the Director with the Autonomous Systems Perception, Intelligence, and Navigation (ASPIN) Laboratory. He is also the Director of the U.S. Department of Transportation Center: CARMEN (Center for Automated Vehicle Research with Multimodal AssurEd ion resiliency and security of highly automated tran portation systems. His research interests include cyber- physical systems, estimation theory, navigation systems, autonomous vehicles, and intelligent transportation systems. Dr. Kassas was the recipient of the 2018 National Science Foundation Faculty Early Career Development Program (CAREER) award and 2019 Office of Naval Research Young Investigator Program award. He was also the recipient of the 2018 IEEE Walter Fried Award, 2018 Institute of Navigation (ION) Samuel Burka Award, and 2019 ION Col. ‘Thomas Thurlow Award. He is an Associate Editor for the IEEE TRANSACTIONS ON "AEROSPACE AND ELECTRONIC SYSTEMS and the IEEE TRANSACTIONS ON INTELLIGENT "TRANSPORTATION SYSTEMS. IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS. VOL.57,NO.3. JUNE 2021