This document proposes using advanced radar features and machine learning to classify ground objects for autonomous vehicle systems. It discusses collecting radar data from different objects, analyzing the data to identify distinguishing features, and training an algorithm to improve classification accuracy. Specifically, it examines using radar cross-section measurements over distance to differentiate between ground objects that increase in reflection and then decrease versus non-ground objects that oscillate. Features like maximum radar cross-section value, position, and change count are analyzed to classify objects into ground or non-ground categories for autonomous driving.
This document proposes using advanced radar features and machine learning to classify ground objects for autonomous vehicle systems. It discusses collecting radar data from different objects, analyzing the data to identify distinguishing features, and training an algorithm to improve classification accuracy. Specifically, it examines using radar cross-section measurements over distance to differentiate between ground objects that increase in reflection and then decrease versus non-ground objects that oscillate. Features like maximum radar cross-section value, position, and change count are analyzed to classify objects into ground or non-ground categories for autonomous driving.
This document proposes using advanced radar features and machine learning to classify ground objects for autonomous vehicle systems. It discusses collecting radar data from different objects, analyzing the data to identify distinguishing features, and training an algorithm to improve classification accuracy. Specifically, it examines using radar cross-section measurements over distance to differentiate between ground objects that increase in reflection and then decrease versus non-ground objects that oscillate. Features like maximum radar cross-section value, position, and change count are analyzed to classify objects into ground or non-ground categories for autonomous driving.
Learning Model for Ground Object Classification in Autonomous Driving Akash Nebhwani, Arifa M Radar Sensing Department,Advanced Driver Assistance Systems Continental Automotive Components (India) Pvt. Ltd., Bengaluru akash.suresh.nebhwani@continental-corporation.com
section is reflected.There are two paths a reflected ray
Abstract will take. This paper proposes a radar based, a machine learning model for Direct path which has high RCS and indirect path with classification of Ground objects for Advanced Driver Assistance low RCS.Cone reflectors gives the most reflection with Systems (ADAS). The framework for the proposal includes same strength as incident ray(no loss). Ground object collecting data from different objects such as ground objects, won't be able to reflect much,resulting in a less obstacles etc.,labelling them with their respective classes, analysis of this data to identify various features from which the RCS,which helps us in further categorizing.i.e,less RCS appropriate features are selected that’ll help distinguish between are weak objects and high RCS are strong objects.And the classes. Algorithm is trained with these features and the KPI also a straight surface will reflect less whereas a curved is measured to check the accuracy thus improving the surface can reflect more.When an ego speed is low,more performance of ADAS frequency is got and viceversa.When a car approaches a target on the road,the radar will help record all the data Keywords - ADAS, Radar, Ground object, Machine Learning associated with the target such as the lateral and longitudinal distances,velocity from which we’d use the Distance X and RCS to do the classification. For Ground I. INTRODUCTION objects,When the distance decreases,there is a linear increase in the RCS trend and after reaching a particular maximum value,RCS gradually drops .In case of Non- 1 Motivation Ground objects,Oscillation patterns are seen as the distance decreases and as distance decreases,frequency of The classification of Objects on the roads ahead of an Oscillations increase.It is to be kept in mind,RCS is autonomous vehicle is essential to operate in safety. On higher for Obstacles compared to that of Ground objects. roads we would constantly come across different objects Ideally,Non-Ground Object can be detected at a far that are made of different materials varying in size, distance in comparison to Ground Object and Non- shapes etc., It is essential to correctly classify these Ground Object escalates to a maximum RCS value at far objects into obstacles and ground objects, with precise distance. confidence, so as to avoid false brakes. The radar for the purpose is of 83 x 69 x 22 mm with a field of view of ± 4° 200m (220m typ.) / ± 9° 120m / ± 45° 40 …70m that has the capability to carry the sufficient information required for the classification. A proper understanding of the classes are needed.There are two main divisions in classes,Ground objects(GO) and Non-ground objects(NGO)/Obstacles.Ground objects are those which are over-rideables and doesn’t hinder the smooth driving of a vehiclewhere in as Non-ground objects are those that can’t be driven over that can trigger the braking action. There’s a variety of Objects that comes under these classes and corner cases where the classifier has trouble working in normal functionality,hence challenging. Radar cross- section is the capability of a target/object to reflect the radar signals in the direction of the receiver.It defines the detectability of an object ,higher the RCS is,higher the detectability is.Radar cross-section elucidates the difference between the reflected signals’ strength from a target to the same from a smooth sphere Figure 1: RCS over the range for a ground object of cross sectional area of 1 sq m. Some targets have large (blue line) and a vehicle(black line) values of RCS owing to their size and orientation and consequently reflect a large portion of incident power.Radar cross section tells us how much cross 2 Paper Structure
This paper is organized as follows: in section II System
implementation is presented, in section III Results are discussed and section IV ends the paper with Conclusion and Future works.
II. SYSTEM IMPLEMENTATION
1 Data Extraction
The data recorded by the Radar needs to be extracted into
a more accessible form so it can be used in further task.The data is extracted with the help of a Database to make further analysis of data easier.Database is where the information is collected and organized.This makes it easier to access,update and manage.Databases are mainly of SQL based and Non-SQL based.SQL based databases are stored as instances that has a structure of Tables further divided into rows and columns.Oracle DB and MYSQL are some examples for these types of Databases.For this work,Non-SQL database Mongo DB is used.The instances in Mongo DB are stored as Collections and further divided into Documents.They are in the form of a Dictionary. A data directory needs to be created at first to store the data. Then a server is set up and required to establish a connection to upload the binary signals generated from the radar.With the help of Robo3T,which acts as an interface between the user and the server acting as a Figure 2:Process chain for enhanced feature calculation bridge, a connection is set up. A Radar labelling tool kit assists in uploading the binary signal data into server via python programming. the range at which the maximum value is observed.Typical maximum value and position observed in an ideal ground object varies from that of a non-ground Once the frames and datas are uploaded into the server object. and are available in an user friendly accessible format,the RCS Maximum change count: This is an important next step is to update the labels. The classes under ground feature as it can be seen in the pattern of RCS for ground object and non-object are defined in the scripts and it and non-ground object ,that for an ideal ground object an helps in labeling the classes properly. increasing RCS trend is observed.RCS Maximum change When executed,a new field is added named ’annotations’ count is the counter which is updated every time a new where all the objects would be listed and labeled with RCS maximum is observed proving to be a differentiating their respective valid class names or background or noise feature for ground vs non-ground object classification. if they are will be added. Maximum RCS interval: This is a second order derived feature, which depicts the distance between two 2 Feature Generation consecutive RCS maxima.”RCS Maximu change count” has a limitation for objects created at near distance due to The Least Square Fit(LSF) algorithm helps in finding curves , obscured scenarios,etc.however Max RCS features that can be used to classify the objects.LSF is interval feature can be exploited for classification in such solely based on linear regression that is to fit a line close chlenging scenarios and corner cases. to the values that has a minimized error.Less error would be of a ground object while more error is obstacle. Regarding the plot,nearly below 25m,there's not much difference between obstacle and ground object,therefore less probability from indirect path so we use LSF near for accurate classification and 25 m to 100 m shows considerable difference.Hence 2 observers are being used.LSF near keeps resetting at all new maximum values of RCS. LSFFar stands for the LSF that deals with all the detections taken from the first detection and LSFNear is that detection that starts with the first detection but resets every time a new global maximum of RCS is detected. Some of the new features derived are as follows: RCS Maximum Value and Position: With every cycle of approach Maximum RCS value is updated along with Figure 3: RCS over the range for a ground object indicating Maximum RCS values and Position along with the Maximum RCS interval features. Figure 5: Implementation of an End-to-End ML pipeline.
The pipeline takes raw labeled data as input and provides
Figure 4: RCS over the range for a vehicle indicating trained model along side all the possible performance Maximum RCS Values and Position along with Maximum RCS interval features. measures as output. It also takes care of hyperparameter tuning inherently. Pipeline returns Train Accuracy, 3 Feature Analysis Validation- Accuracy, Best Parameters, Test Accuracy, Nos. Of misclassified instances, Confusion- Matrix and The features we have derived from the earlier steps are such. In addition to all or any usual performance now undergone analysis.This is necessary because some measures, pipeline also returns few application specific features are correlated,which means that the features measures, out of which most interesting is plot for might have same information.They need to be removed misclassified signals. since they basically carry the same data.Removing them can help in maximizing the performance with a reduced 5.1 Data split computation.Less running time and less memory consumption.This is why feature analysis is an important Albeit the task is simply binary classification, the dataset step in the process and the feature selection is done is unbalanced and consists of approaches to differing adequately. types of object. Therefore, the splitting strategy is a crucial think about what and the way the model learns. to not mention, this directly changes the performance of the 4 Feauture Selection model. In scikit-learn several methods are available to With a lot of ways to generate new features, one might be separate the info . Scikit-learn official documentation allured to boost the dimensionality of the data set more explains then well [skl-split]. Not of these split strategies than the number of original available features. But are applicable / justifiable for the task. Therefore, Manual addition of more features will lead for a more complex (handpicked), StratifiedShuffleSplit, GroupKFold and model,and thus escalates the chances of overfitting.When StratifiedKFold splits are considered for the analysis. new features are added ,specifically at a high dimensional dataset ,it’s better to reduce the number of features into the most relevant ones and dispense the rest of the Grid search is that the process of performing hyper features.This results in getting a simple model which can parameter tuning so as to work out the optimal values for help the generalization better. There are mainly three a given model. This is important since the performance of ways to do feature selection: the model depend on the hyper parameter values specified. it is hard to stipulate values for hyper Univariate Selection,Model-based Selection and Iterative parameters. There are libraries that are implemented, like Selection.All of these three methods are supervised GridSearchCV of the sklearn library, so as to automate learning techniques,which means they require a target so this process and make life a touch bit easier for ML that they can fit in. And it requires for the data to be split enthusiasts. into two,I.e, training dataset and testing dataset and fit the feature selection into the training dataset. 5.2 Hyper Parameter tuning 5 Pipeline When creating a machine learning model, you will be In case of machine learning,a standard procedure needs presented with design choices on the way to define your to be followed inorder to compare the outcomes with that model architecture. Often times, we do not immediately of various models.So it is more sensible to make a know what the optimal model architecture should be for pipeline and using it.One such pipe line is given below.To a given model, and thus we'd wish to be ready to explore add scikit-learn already have an implemented a variety of possibilities. In a perfect machine learning pipeline.[pip-git].Yet inorder to follow a standard method environment, we expect the machine to perform this sklearn pipeline is used(scikit-learn principle). exploration and choose the optimal model architecture on it's own. Parameters which define the model architecture are mentioned as hyperparameters and thus this process of checking out the perfect model architecture is mentioned as hyperparameter tuning. One of the foremost important aspects when training neural networks is avoiding overfitting. Overfitting refers to the phenomenon where a neural network models the training data alright but fails when it sees new data from an equivalent problem domain. Overfitting is caused by noise within the training data that the neural network picks up during training and learns it as an underlying concept of the info .Regularization refers to a group of when combined with another feature or in simple words various techniques that lower the complexity of a neural their presence doesn’t improve the accuracy or reliability network model during training, and thus prevent the of the system which help us to come to a conclusion on overfitting. which features need to be used and which needs to be disregarded.It is clear that along different ranges,for the There are mainly two types of regularization called three different methods,the features selected and their • L1 Regularization combination are different,And they are so selected that the particular combination gives the maximum accuracy • L2 Regularization for the model.
III. RESULTS
Results are simulated for ground objects and non ground
objects of varied classes such as Car to Car scenarios, Muffler, Muffler with tube, Metal Plate, Potato chips bag, Short Ratchet strap, Small plastic box, Tin can, Wooden pallet, Adult Dummy, Child Dummy, Large tire with rim, Steel catalytic converter, Manhole cover, Long Metal profile, L profile and Mercedes E class. Figure 7: Feature selection by three methods. The data consists of a number of different classes and a variety of approaches.The objects from LC are harder to classify and that complicates functioning of classifier.So With the selected features,a model has been built and care is to be taken in splitting up of the data and balancing coefficients for the features and intercept values are properly into training data and testing data.Developing extracted. more features are yet another solution.
The Histogram shows a feature can define a boundary
between the ground objects and non ground objects,(here the usecase 0 represents ground objects and viceversa).A single feature cannot properly distinguish between the ground object and non-ground object and hence we consider other features for training a reliable model.
Figure 8: Model Generation
IV. CONCLUSION AND FUTURE WORKS
This thesis goes through various important steps, In
conclusion in this work, Logistic regression is used as ML model and for testing and training the data and to get an Figure 6: Class wise histogram for a single feature accuracy value,thus the performance.An end to end pipeline is designed which takes in the raw data as the input and produces a trained classifier with best When feature selection step is triggered,the program parameter.This along with the result has measures of the takes all the features,individually and combined with the performance constitutes the output.The design of this features and does feature selection in all the three pipeline has steps such as data analysis and methods we mentioned i.e., in univariate,model-based extraction,which extracts the radar data in a more user and iterative feature selection method at different ranges friendly,easily accessible format,feature analysis and of distance ranges.The result is generated in an excel feature selection where we select the features used for format.Table 5.1 shows the feature selection done using training,disregarding the rest with the help of three two original features methods such as univariate,model-based and iterative The accuracy when a single feature is considered is feature selection. And hyperparameter tuning with the different from when both the features were used and that help of grid-search and documentation. varies over each ranges of distances considered.The KPI There are a lot many opportunities and possibilities to is found to be improving when there is a combination of further extend this work and hence improve the model features used together rather than using a single performance.One is to introduce data augmentation with feature.And also further analysis will show some models like Ada-Boost, Gradient-Boosting, Decision combination of feature doesn’t necessarily increase the Tree etc and to analyze the results for a better accurate test accuracy which means they are only informative model. Another future scope is to implement a multi-class classifier with classes such as GO,NGO,Probably GO,Probably NGO and so on.This increases the reliability and confidence the system.Develop pipeline for a generalization so that it helps in other applications other than Ground-Object Classification.