An Optimized Poisson Multi-Bernoulli Multi-Target Tracking Algorithm

Poisson multi-Bernoulli filtering can track multiple targets in scenarios where point targets and extended targets coexist. However, when the target undergoes close-proximity, cross motion, or long-term occlusion, tracking accuracy will decrease, and even problems such as false detection, missed tracking, and label jumping may occur. Aiming at this problem, an optimized Poisson multi-Bernoulli multiple target tracking algorithm is proposed. Firstly, Introducing a target-type confidence function to address the challenge of correctly identifying target types when point targets and extended targets are closely adjacent or overlapping. Simultaneously determine whether there are adjacent or overlapping areas of targets, and maintain the number of targets within the area to reduce false positives, missed detections, and label jumps on targets. In addition, to address the issue of targets being easily lost due to long-term occlusion, a Long Short Term Memory Model (LSTM) is introduced to optimize the motion model prediction of the targets, and combined with three-level gating strategy to identify the reborn targets. The experiment shows that the proposed algorithm has good performance in target type recognition and continuous trajectory tracking in complex scenarios where point targets and extended targets coexist.