This paper introduces a flexible measurement model, namely, the hierarchical truncated Gaussian, to resemble the spatial distribution of automotive radar measurements on a vehicle and, along with adaptively updating truncation bounds, to account for partial-view measurements caused by object self-occlusion. Built on a random matrix approach, we propose a new state update step together with the adaptively update of the truncation bounds. This is achieved by introducing spatial-domain pseudo measurements and by aggregating partial-view measurements over consecutive time-domain scans. The effectiveness of the proposed algorithm is verified on a synthetic dataset and an independent dataset generated from the MathWorks Automated Driving toolbox.

Automotive radar; object tracking; Bayesian filtering; random matrix; autonomous driving