基于马尔科夫决策过程的多目标跟踪算法

目前，多目标跟踪算法仍面临诸多挑战，例如遮挡、快速运动等所造成的影响难以完全规避。为了解决上述问题，提出一种基于马尔科夫决策过程的多目标跟踪算法。该算法将每个目标建模成一个马尔科夫决策过程，通过最大化奖励函数来驱动状态间的转移，并将强化学习训练用于数据关联相似度函数，有效地解决了目标遮挡问题。同时，为了解决物体快速运动导致跟踪算法丢失目标问题，利用超像素建立表观模型，充分考虑历史图像信息，提高跟踪算法的准确性与可靠性。实验评估表明，该跟踪器在公开的MOT15数据集上具有良好的性能。提出的跟踪器在多目标跟踪精度（multide object tracking accuracy，MOTA）指标上达到36.5，远高于其他对比算法，而在ID switch指标上仅仅为308次，低于其他对比算法，显著地减少了目标丢失率以及身份交换率。

Multi-target tracking algorithm based on Markov decision process

Multiple Object Tracking still face many challenges, such as occlusion, rapid motion, and other effects that are difficult to avoid completely. In order to solve the above problems, a multi-target tracking algorithm based on Markov decision process is proposed. The algorithm models each target as a Markov decision process, drives the transition between states by maximizing the reward function, and uses reinforcement learning to train the similarity function for data association, which can effectively solve the target occlusion problem. At the same time, in order to solve the problem that the tracking algorithm loses the target due to the rapid movement of the object, the superpixel is used to establish the apparent model, fully consider the historical image information, and improve the accuracy and reliability of the tracking algorithm. Experimental evaluation shows that the tracker has good performance on the public MOT15 data set. The tracker proposed in this paper achieves 36.5 on the MOTA index, which is much higher than other comparison algorithms, and only 308 times on the ID switch, which is lower than other comparison algorithms, which significantly reduces the target loss rate and identity exchange rate.