基于簇相似度的实时多尺度目标跟踪算法

Real-Time Multi-scale Object Tracking Based on Cluster Similarity

针对目标跟踪中跟踪实时性和适应目标尺度变化的问题,提出在粒子滤波框架内基于簇相似度测量的实时目标跟踪算法,算法的外观模型使用改进的均值类哈尔特征表示.首先,根据采样半径采集目标簇和背景簇.然后,定义粒子与簇之间的相似度.当新帧到来时计算每个粒子与目标簇和背景簇的相似度,并将相似度最高的粒子作为目标在该帧的位置.在每帧跟踪结束时,更新目标簇和背景簇的统计特征,并对粒子进行重采样防止退化.与当前通用的跟踪算法对比体现文中算法的优越性.

To solve the problems of real-time object tracking, a real-time object tracking algorithm tracking and scale changing of the object in object is proposed based on cluster similarity measurement （MSCSM） in particle filtering framework. The improved average haar-like features are utilized to represent the proposed appearance model. Firstly, the target cluster and the background cluster are cropped in their sample radii. Secondly, a similarity measurement between a particle and a cluster is defined. The score of each particle is calculated according to its similarity with clusters while a new frame coming. Finally, the particle with the maximum score is selected as the new target location in the current frame. At the end of tracking for each frame, statistical characteristics of clusters are updated and the particles are resampled to avoid degeneration. The proposed algorithm shows superiority in comparison with the state-of-the-art algorithms.