You are on page 1of 3

Real-time adaptive data filtering with multiple

sensors for indoor monitoring


NOMS 2022-2022 IEEE/IFIP Network Operations and Management Symposium | 978-1-6654-0601-7/22/$31.00 ©2022 IEEE | DOI: 10.1109/NOMS54207.2022.9789786

Kuon Akiyama and Ryoichi Shinkuma Jun Shiomi


Faculty of Engineering, Graduate School of Information Science and Technology,
Shibaura Institute of Technology, Japan Osaka University, Japan
{al18004, shinkuma}@shibaura-it.ac.jp shiomi-jun@ist.osaka-u.ac.jp

Abstract—This demonstration proposes a scheme that sup-


presses the data size of a 3D-image sensing network by adaptively
filtering low-importance points, such as the points of floors and
ceilings when the task is to track pedestrians. The adaptive filter
can be dynamically changed to further reduce the amount of
data if it is difficult for packets to reach the edge computer.
We evaluate the proposed scheme through experiments and
demonstrate that it performs better than benchmark schemes
in terms of prompt arrival of the data.
Index Terms—smart monitoring, LIDAR, 3D image sensing,
adaptive filtering, sensor network Fig. 1. System model.

I. I NTRODUCTION sensor consists of a sensor and a sensor device that processes


the data acquired by the sensor unit; these are connected
Smart cities are urban platforms that provide citizens with to each other via a local network. The edge computer is a
social public services such as education, medical care, safety, machine that aggregates the data streams received from the
transportation, and utilities through the use of various informa- sensors. The proposed scheme uses a single edge computer
tion and communication technologies [1]. Smart monitoring is and multiple sensors. Each sensor is connected to the edge
a major component in smart cities that collects information to computer. One issue is that, if we use a larger number of
manage the infrastructures and resources efficiently [2]. sensors, the system will experience increased latency and
Light detection and ranging (LIDAR) technology has been jitter. The edge computer in the proposed scheme performs a
utilized to collect three-dimensional (3D) data useful for smart periodic aggregation to adjust the timing of data with various
monitoring [3] [4]. Since the amount of available information latency and jitter from each sensor. The sensors in the proposed
from a single sensor depends highly on its location and the scheme filter out low-importance points (e.g., the points of
presence of blind spots caused by obstacles, the installation floors and ceilings in a pedestrian-tracking task) to reduce
of multiple sensors has been discussed [3] [4]. This would the data size adaptively in order to avoid network overflow.
reduce blind spots and increase the amount of information We evaluated the proposed scheme by examining its real-time
that could be acquired in the overlapping area of the sensors. packet arrival ratio through software program experiments and
However, the prior works have focused mainly on outdoor found that is effective. While there have been several prior
monitoring with a relatively low number of sensors, mostly studies on 3D-image sensing networks [4] [6], our work differs
for road safety. For outdoor use, four or five sensors can cover in that our proposed scheme uses tens of sensors and performs
the monitoring area without blind spots, as the sensors can be adaptive filtering to deal with the difficulty of packets reaching
placed high. According to [5], four sensors can cover an entire the edge computer.
central intersection. In contrast, for indoor use, ten or more
sensors are required to eliminate blind spots, as the placement II. P ROPOSED SCHEME
of sensors is limited by ceilings. In addition, the greater the A. System model
number of sensors that are used, the more delay and jitter Fig. 1 shows block diagrams of the 3D-image sensing
there will be, since the communication bandwidth will have network for the proposed scheme. It consists of multiple
to be shared. Using too large a number of sensors may cause sensors and one edge computer. A local network is used to
overflow as well. connect each sensor unit to each sensor device and the sensor
In light of this background, we demonstrate a scheme that devices to the edge computer.
suppresses the data size of a 3D-image sensing network by Each sensor unit sequentially generates streaming data for
adaptively filtering out low-importance points. It consists of each frame and sends it to the sensor device it is connected
two main components: a sensor and an edge computer. The to in the form of a user datagram protocol (UDP) packet. The
978-1-6654-0601-7/22/$31.00 © 2022 IEEE filter of the sensor device performs filtering for the point-cloud

NOMS 2022 - 2022 IEEE/IFIP Network Operations and Management Symposium


Authorized licensed use limited to: University of Greenwich. Downloaded on August 21,2023 at 20:47:12 UTC from IEEE Xplore. Restrictions apply.
data. We note that the bandwidth of communication from the
transmitter of the device to the receiver of the edge computer is
strictly limited because it is reasonable that the edge computer
is located separately from devices and is connected to them
via a shared wireless channel.
The merger of the edge computer takes the received data
frames from the storage, and determines the frame that each (a) vs. frame number (b) Histogram
received packet belongs to on the basis of the timestamp
Fig. 2. Fluctuating SDRR of LIDAR-1 with benchmark.
information given by each sensor. To accommodate the varying
delay and jitter of packets arriving from sensors, the merger to packet loss or delay. Packets could be lost during UDP
waits for a while before publishing a frame. Specifically, transmission, or dropped at the merger if the frame latency
it waits until the number of arrived frames waiting to be is too long (i.e., longer than the period of frames waiting for
published reaches a maximum value for any sensor. Packets being published). If the SDRRs calculated for each m frames
belonging to frames delayed longer than this will be dropped. are less than a pre-specified threshold T , the notifier will alert
The evaluator periodically calculates the delay status for each the corresponding sensors by sending a UDP packet.
sensor. The notifier informs the detector of sensor devices
in response to the delay status. The detector changes the III. E VALUATION
parameter, used for the filter, on notifications to further reduce A. Reproduction of data sent from sensors
the amount of data sent to the edge computer.
We reproduced a large number of sensor units to evaluate
B. Adaptive filtering method the proposed scheme. We monitored our laboratory with five
sensors named puck, also called VLP-16, commercialized by
The filter of each sensor device filters out low-importance
Velodyne [7], and generated a file (pcap file) that captured
points to reduce the data size and avoid network overflow.
network traffic for each sensor. The detailed information is
The filter obtains the pre-specified importances of each region
available on Kaggle [8].
of 3D space divided into voxel grids, prioritizes voxel areas
based on their importances, keeps points inside the top n B. Setup for experiment
voxels and removes the points on the outside. The number We evaluated the proposed scheme by doing experiments
of allowed voxels n is initialized as the number of voxels with the sensing network virtually. We ran three types of
that have been allocated importance values and is decremented programs to run: tcpreplay (an open source utility [9]), sensor
when the detector is notified by the edge computer to reduce device program, and edge computer program. The 3D-image
the data size. sensing network consists of 30 sensors and one edge computer.
C. Control metric and measurement method We used two sets of five NVIDIA Jetson Nano Development
Kits for executing the tcpreplay and sensor device programs,
The evaluator of the edge computer periodically evaluates
and one NVIDIA Jetson Xavier NX Development Kit for
the state of each sensor by a calculated control metric called
executing the edge computer program. These machines were
the satisfied delay-requirement ratio (SDRR), which represents
commercialized by NVIDIA [10] [11], and connected by wire.
the latencies of UDP packets sent by the sensor. SDRR is
Data sent from a sensor was reproduced by replaying the
defined as:
network traffic recorded in the pcap file using tcpreplay. In this
1 ∑
m experiment, we evaluated using the pcap files of LIDAR-1 and
SDRR(t) = Rappended (t − i), (1) LIDAR-3 which recorded the highest number of points among
m i=1
the five sensors. The machine for the sensor device ran six
where t, m, and Rappended (t) are respectively the current programs in parallel. A unique integer was randomly assigned
frame number, the number of frames of a single cycle, and for each voxel as importance to these voxels. To reproduce the
the appended ratio, defined as: strict bandwidth limitation, we used USB 2.0 Ethernet adapter
rather than the one for USB 3.0.
{N
appended (t)
Nexpected (t) Nexpected (t) ̸= 0 C. Evaluation metric
Rappended (t) = , (2)
0 otherwise The average SDRR can be used to evaluate the status of the
network since the packet loss or the latencies make this value
where Nappended (t) represents the number of packets the low.
edge computer receives at frame t within the satisfied delay
requirement and Nexpected (t) represents the number of packets D. Benchmark for schemes
that the sensor sends for frame t. Nexpected (t) is set in We first conducted a benchmark experiment using 30 sen-
accordance with the information on the number of packets sors without executing the filter.
to be transmitted included in each packet transmitted by Fig. 2(a) shows the average SDRR for each timing using
the sensor. Nappended (t) may be less than Nexpected (t) due the pcap file of LIDAR-1. The average number of points in

NOMS 2022 - 2022 IEEE/IFIP Network Operations and Management Symposium


Authorized licensed use limited to: University of Greenwich. Downloaded on August 21,2023 at 20:47:12 UTC from IEEE Xplore. Restrictions apply.
scheme could turn off the modification of the filter by the de-
tector. Therefore, the proposed scheme receives many packets
with less packet loss and fewer latencies within the satisfied
delay requirement when the modification was not required.
IV. C ONCLUSION
This demonstration paper proposed a scheme that aggregates
Fig. 3. Stably high SDRR of LIDAR-3 with benchmark.
data from a large number of sensors and suppresses the amount
of data in real time by filtering points. The sensing network for
the proposed scheme consists of sensors and an edge computer
that are connected via a local network. The filter of the sensor
device works in real time, keeps points inside the pre-specified
voxel regions and removes points on the outside in order
to avoid network overflow. The edge computer periodically
aggregates the data received from multiple sensors, which
(a) vs. frame number (b) Histogram
accommodates for packet arrival delay and jitter. In addition,
Fig. 4. Stably high SDRR of LIDAR-1 with proposed scheme. the filter dynamically changes the amount of data when
it is difficult for packets to arrive within a satisfied delay
the point clouds was as high as 24,573.0. We can see that the requirement. We performed experiments by emulating a large
SDRR took various values when the filter was not executed. number of sensors and found that the proposed scheme is
The overall average SDRR was approximately 0.7678. The far effective in that more packets are received within the satisfied
ends of the graph were low due to the time lag caused by the delay requirement.
start order of tcpreplay. Fig. 2(b) shows a histogram of the
SDRRs, where we can see that the ratio of packets received V. ACKNOWLEDGEMENT
within the satisfied delay requirement was very low in the This work was supported in part by JST PRESTO no.
benchmark using pcap files of LIDAR-1. JPMJPR1854, JST SBIR-1 no. JPMJST2151, and JSPS KAK-
Fig. 3 shows the average SDRR using the pcap file of ENHI no. 21H03427, Japan. The evaluation results were partly
LIDAR-3. The average number of points in the point clouds obtained from research commissioned by NICT, Japan.
was 17,699.0, which is 30% less than that of LIDAR-1.
R EFERENCES
Although the filtering was not executed, the SDRR was high
[1] J. Bélissent et al., “Getting clever about smart cities: New opportunities
with an overall average of about 0.9627. This result shows that require new business models,” Cambridge, Massachusetts, USA, vol.
when data are from LIDAR-3, the amount is enough small to 193, pp. 244–277, 2010.
be received within the satisfied delay requirement. [2] R. Du, P. Santi, M. Xiao, A. V. Vasilakos, and C. Fischione, “The
sensable city: A survey on the deployment and management for smart
E. Results of evaluation city monitoring,” IEEE Communications Surveys & Tutorials, vol. 21,
no. 2, pp. 1533–1560, 2018.
We experimented with a network for the proposed scheme [3] R. Otsu, R. Shinkuma, T. Sato, and E. Oki, “Data-importance-aware
using 30 sensors while executing the filter, evaluated it using bandwidth-allocation scheme for point-cloud transmission in multiple
lidar sensors,” IEEE Access, vol. 9, pp. 65 150–65 161, 2021.
the two pcap files of LIDAR-1 and LIDAR-3, and compared [4] C. Li, R. Shinkuma, T. Sato, and E. Oki, “Two-level processing scheme
it with the results of the corresponding benchmarks. for 3d-image sensing network,” in 2021 IFIP Networking Conference
Fig. 4(a) shows the average SDRR for each timing using the (IFIP Networking). IEEE, 2021, pp. 1–2.
[5] M. Goldhammer, E. Strigel, D. Meissner, U. Brunsmann, K. Doll, and
pcap file of LIDAR-1. The overall average was about 0.9424, K. Dietmayer, “Cooperative multi sensor network for traffic safety ap-
which is higher than that of the benchmark. Fig. 4(b) shows a plications at intersections,” in 2012 15th International IEEE Conference
histogram of the SDRRs. We can see here that the proposed on Intelligent Transportation Systems. IEEE, 2012, pp. 1178–1183.
[6] C. Li, R. Shinkuma, T. Sato, and E. Oki, “Real-time data selection and
scheme receives many more packets with less packet loss and merging for 3d-image sensing network with multiple sensors,” IEEE
fewer latencies within the satisfied delay requirement. Sensors Journal, vol. 21, no. 19, pp. 22 058–22 076, 2021.
Fig. 5 shows the average SDRR of LIDAR-3. The overall [7] Velodyne, “Puck lidar sensor, high-value surround lidar
average was about 0.9627. The SDRR remained high because — velodyne lidar,” Accessed on January 19, 2021,
https://velodynelidar.com/products/puck/.
the amount of data was small, which meant the proposed [8] R. Shinkuma, “5 lidars — kaggle,” Accessed on February 9, 2022,
https://www.kaggle.com/rshinkuma/5lidars.
[9] AppNeta, “Tcpreplay - pcap editing and replaying utilities,” Accessed
on October 15, 2021, https://tcpreplay.appneta.com/.
[10] NVIDIA, “Nvidia jetson nano developer kit —
nvidia developer,” Accessed on October 15, 2021,
https://developer.nvidia.com/embedded/jetson-nano-developer-kit.
[11] ——, “Jetson xavier nx developer kit — nvidia developer,” Accessed on
October 15, 2021, https://developer.nvidia.com/embedded/jetson-xavier-
nx-devkit.

Fig. 5. Stably high SDRR of LIDAR-3 with proposed scheme.

NOMS 2022 - 2022 IEEE/IFIP Network Operations and Management Symposium


Authorized licensed use limited to: University of Greenwich. Downloaded on August 21,2023 at 20:47:12 UTC from IEEE Xplore. Restrictions apply.

You might also like