Extended Object Tracking Algorithms Within Interacting Multiple Model Framework for Automotive Radar Applications

Abstract

This article presents a novel computationally efficient extended target tracking (ETT) framework for automotive radar based on the interacting multiple model (IMM) algorithm. First, the detailed mathematical foundation of a new variant of a multihypothesis tracking algorithm for extended object tracking named centroid-oriented multihypothesis centroid-oriented MHT (Co-MHT) is developed, in which confirmed measurement clusters obtained after hypothesis pruning are used to construct an equivalent kinematic observation together with a spatial scatter representation. This cluster-level formulation enables joint estimation of target kinematics and physical extent, while eliminating hypothesis tree growth and significantly reducing runtime and memory requirements compared with classical IMM track-oriented multihypothesis tracking (To-MHT). The framework incorporates constant-velocity (CV) and constant-turn-rate-and-acceleration (CTRA) motion models and supports robust estimation of object orientation and dimensions using detection clustering and statistical extent modeling. The second contribution of this article is to propose a detailed mathematical foundation of a new variant of the traditional probabilistic data-association filter (PDAF) named the skew-adaptive PDAF (SA-PDAF) to address the limitations of conventional probabilistic data association (PDA) in extended target scenarios. High-resolution automotive radars often generate multiple, nonuniform reflections from different parts of an object, resulting in highly skewed association probabilities that lead to centroid drift and track fragmentation. The proposed SA-PDAF algorithm applies a novel skew-adaptive transformation to the association probabilities, thereby suppressing imbalance, stabilizing centroid estimation, and improving robustness against object fragmentation. Extensive on-road vehicle experiments and offline performance evaluation using optimal sub-pattern assignment (OSPA) metric demonstrate that both IMM-based Co-MHT and the SA-PDAF framework achieve stable tracking performance, accurate kinematic and extent estimation, and superior robustness under real-life extended target conditions.