3D Multi-Object Tracking Based on Probability Distribution and State Regression

In previous multi-object tracking paradigms, a complex data association strategy is generally needed to achieve accurate matching for detections and trajectories. In this paper, a novel end-to-end 3D multi-object tracking (MOT) framework is proposed based on probability distribution and state regression. Firstly, this framework does not rely on complex data association strategies; instead, it derives the accurate position of an object in the current frame directly by regression based on the prior information of the object’s trajectory. Secondly, a probability grid sampling strategy is then adopted to expand the regression search range of the trajectory in an adaptive manner, thereby reducing the uncertainty of the trajectory states caused by consecutive predictions. Lastly, to eliminate overlaps of trajectories, a trajectory interaction module is introduced to retain trajectories with higher confidence. Experiments are conducted on the KITTI and Waymo datasets. The results demonstrate that this framework can achieve robust tracking by utilizing only a single 3D detector, outperforming many tracking-by-detection methods.