Evaluating Motion Model Hypotheses for Automotive Radar Tracking

Abstract

Automotive radar is a core ADAS sensor due to weather robustness and direct Doppler velocity measurements, but radar multi-object tracking is challenged by heterogeneous and maneuvering target dynamics. This thesis evaluates a white-noise jerk model (CCA) and a curvilinear motion model (CTCA), each implemented in an EKF, and assesses whether combining them in an Interacting Multiple Model (IMM) filter improves robustness across scenarios. To enable an unbiased comparison between CCA, CTCA, and IMM, process-noise parameters and IMM transition/interactions are tuned automatically by formulating tracking as a black-box optimization problem. Performance is optimized using the probabilistic GOSPA (P-GOSPA) metric, which penalizes localization error as well as missed and false tracks under multi-Bernoulli set representations. CMA-ES is used to search the nonconvex parameter space without gradients. Evaluation is performed in a controlled MATLAB simulation with a four-corner radar configuration and known ground truth, fusing radar range-rate detections with object-level pseudo-measurements of position and orientation. Results show strong scenario dependence for single-model tracking and indicate that automated tuning is necessary to avoid biased motion-model conclusions, the IMM provides more consistent performance across diverse driving scenarios than either single model.