Abstract: When running particle filter tracking algorithm, the loss of target due to occlusion will seriously reduce tracking accuracy and robustness. To solve this problem, a method to determine the state of target loss and a method of target position reconstruction based on improved sequential similarity detection are proposed. When the target loss is detected, the tracking algorithm is restarted. The improved sequential similarity detection uses the Bhattacharyya distance instead of pixel cumulative error, which can better adapt to the situation of target rotation, deformation, scale variation and so on. Using OTB-100 standard datasets, the proposed algorithm is compared with traditional particle filter tracking algorithm, SCM and other classical algorithms. The experimental results show that the tracking success rate of proposed algorithm is 36.6% and 3.2% higher than that of traditional particle filter tracking algorithm and the best anti-occlusion algorithm of OTB-100 datasets, respectively. The proposed algorithm improves the stability of the tracking process. In addition, the experimental results are compared with the latest research results of particle filter tracking algorithm.

Key words: object tracking, particle filter, sequential similarity detection, anti-occlusion

摘要： 在粒子滤波跟踪算法运行过程中，由于目标遮挡导致丢失目标，将严重地降低跟踪精度与鲁棒性。为了解决此问题，提出了目标丢失状态判定方法和基于改进序贯相似性检测的目标位置重建方法，当检测到目标丢失时，重启跟踪算法。改进序贯相似性检测使用Bhattacharyya距离代替像素累积误差，更好地适应检测目标发生旋转、形变、缩放等情况。使用OTB-100标准数据集，将该算法和传统粒子滤波跟踪算法、SCM等经典算法比较。实验结果表明，对于含遮挡特性视频序列，本文算法比传统粒子滤波跟踪算法和OTB-100抗遮挡最优算法跟踪成功率分别提高36.6%和3.2%，提升了跟踪过程的稳定性。此外，还将实验结果与最新粒子滤波跟踪研究成果作对比分析。

关键词: 目标跟踪, 粒子滤波, 序贯相似性检测, 抗遮挡