Standard Kalman filters used in 3D multi-object tracking assume a static motion model, which is fast but prone to state drift when an object's motion changes—especially for distant or occluded targets. Recent multimodel approaches improve robustness by running several motion models at once, but they add significant computational overhead.

The authors propose a motion dynamics Kalman filter (MD-KF) that keeps a single-model structure. It models how an object's motion varies over time and adaptively scales the contribution of motion components to correct prediction drift, preserving efficiency while improving accuracy.

Evaluations on KITTI and the Waymo Open Dataset, plus vehicle experiments, show MD-KF achieves tracking performance comparable to multimodel KF with an average 1.72x speedup. It also handles highly distant and prolonged occluded objects better than the multimodel approach.