IMM-based Multiple Object Tracking using a State Prediction Neural Network
| Source: arXiv AI
Tags: object tracking, autonomous vehicles, radar, Interacting Multiple Model, transformer, Doppler
A transformer-based radar tracking method (PR-IMM) integrates a neural state predictor into the classical Interacting Multiple Model framework, reducing position estimation error by 57.3% over standard IMM and cutting ID switches by 25.3% for autonomous vehicle object tracking.
Details
Object tracking for autonomous vehicles requires accurate motion state estimation even in adverse weather, where camera-based systems struggle. Radar is weather-robust and provides Doppler velocity measurements, but classical physics-based motion models (Constant Velocity, Constant Acceleration, Constant Turn) misrepresent nonlinear real-world trajectories.\n\nThis paper proposes PR-IMM: a transformer-based state predictor (PR) that uses radar Doppler measurements to predict object displacement, integrated as an additional mode alongside CV, CA, and CT models in the Interacting Multiple Model (IMM) framework. Mode probabilities dynamically weight each model's prior position estimate.\n\nOn the evaluation dataset, PR-IMM reduces position estimation error by 57.3% over baseline IMM and by 16.5% over the standalone PR model alone. ID switches drop by 25.3% and IDF1 (identity tracking quality) improves by 9.6% over IMM. The 8-page paper is submitted to IEEE Transactions on Cognitive and Developmental Systems.\n\nThe method is radar-specific and evaluated on a single dataset. It shows that combining learned predictors with physics-based multi-model frameworks improves tracking stability more than either approach alone — a useful design pattern for autonomous driving sensor fusion teams.