Evaluating Motion Model Hypotheses for Automotive Radar Tracking

Sammanfattning

Automotive radar is a core ADAS sensor due to weather robustness and directDoppler velocity measurements, but radar multi-object tracking is challenged byheterogeneous and maneuvering target dynamics. This thesis evaluates a white-noisejerk model (CCA) and a curvilinear motion model (CTCA), each implemented in anEKF, and assesses whether combining them in an Interacting Multiple Model (IMM)filter improves robustness across scenarios. To enable an unbiased comparison betweenCCA, CTCA, and IMM, process-noise parameters and IMM transition/interactionsare tuned automatically by formulating tracking as a black-box optimizationproblem. Performance is optimized using the probabilistic GOSPA (P-GOSPA)metric, which penalizes localization error as well as missed and false tracks undermulti-Bernoulli set representations. CMA-ES is used to search the nonconvex parameterspace without gradients. Evaluation is performed in a controlled MATLABsimulation with a four-corner radar configuration and known ground truth, fusingradar range-rate detections with object-level pseudo-measurements of positionand orientation. Results show strong scenario dependence for single-model trackingand indicate that automated tuning is necessary to avoid biased motion-modelconclusions, the IMM provides more consistent performance across diverse drivingscenarios than either single model.