基于层级特征与相似性估计的跟踪器

Tracker Based on Hierarchical Features and Similarity Estimation

针对多目标跟踪中存在的小目标易漏检和遮挡问题造成的身份切换等问题,沿用经典的基于检测跟踪(tracking-by-detection)框架,提出了一种基于深度学习的数据关联方法.利用深度学习特征提取的能力,在基于行人检测基础上设计了层级特征提取相似性估计的多目标跟踪器.通过端到端的方式提取跟踪物体的层级特征,并计算不同帧物体间特征的相似性,得到相似性矩阵.再利用匈牙利算法根据相似性矩阵完成最优指派问题,实现数据关联部分.实验结果表明,所设计跟踪器缓解了目标跟踪过程中由于遮挡问题带来的跟踪物体身份切换问题,并且在MOT16数据集上取得了较好的效果.

Aiming at the problems of lossing small targets and the problem of identity switching caused by occlusion problems in multi-target tracking, a classic data-based association method based on deep learning is proposed based on the classical tracking-by-detection framework. Based on the ability of deep learning feature extraction, a multi-target tracker with hierarchical feature extraction similarity estimation is designed according to pedestrian detection. The similarity matrix is obtained by extracting the hierarchical features of the tracking object in an end-to-end manner and calculating the similarity of features between different frame objects. Tlie Hungarian algorithm is used to complete the optimal assignment problem based on the similarity matrix, and the data association part is realized. The experimental results show that the designed tracker alleviates the problem of tracking object identity switching clue to occlusion problem in the target tracking process, and has achieved good results on the MOT16 data set.