Bayesradar : Bayesian Metric-Kalman Filter Learning for Improved and Reliable Radar Target Classification

Abstract

Automotive radar sensors offer a promising and effective modality for perception and assessment of the surrounding environment. A key element of environment sensing in automotive radars is the reliable detection, classification and tracking of vulnerable road users such as pedestrians and cyclists. In this paper, we propose an integrated Bayesian framework dubbed BayesRadar, which improves the overall classification accuracy by tracking the embedding vector of a neural network and its prediction uncertainty via recursive Kalman filtering over time. Apart from the classification accuracy of a model, a critical measure includes the analysis of statistical confidence over the target class score. Such measures for predicting the true correctness likelihood of the classification estimates are essential in safety-critical automotive applications. Therefore, in this paper, we present and evaluate the classification, embedding cluster score and statistical confidence performance of the proposed framework in the context of classifying vulnerable road users compared to state-of-art deep learning approaches. Furthermore, we demonstrate superior performance of the BayesRadar for unseen classes compared to long short-term memory based temporal tracking of the embedding vectors.