基于多特征信息融合的目标轨迹聚类方法

Object trajectory clustering method based on multiple features information fusion

提出了一种基于多特征信息融合的运动目标轨迹聚类方法.针对视频监控目标的特点,引入轨迹均值、距离方向、运动方向和平均速度4个特征空间来描述目标的运动轨迹.首先,采用Mean-Shift算法对每个特征空间进行聚类,得到基本的运动类别信息;其次,设计多特征融合算法,通过计算不同特征空间的类别间关系,进行类别信息融合;最后,得到融合了多个特征空间信息的聚类结果.由于信息融合是在聚类层面进行的,能够有效避免在特征空间层面融合时的维数统一问题.试验结果表明了本方法的有效性.

A novel video object trajectory clustering method is presented based on multiple feature information fusion. According to the characteristics of video surveillance objects, four features consisting of the target traj- ectory mean, the directional distance, the motion direction and the average target velocity are selected to de- scribe the trajectory of the objects. Firstly, the well known Mean-Shift algorithm is used to identify the modes and clusters. Then, the multiple features fusion strategy is performed by calculating and comparing the clusters among different feature spaces. Finally, the final common patterns are estimated by fusing the clustering re- sults across all feature spaces. The fusion is carried out at cluster level so that it can avoid the dimensionality problem happened at feature level. Experimental results are also provided to demonstrate the effectiveness of the proposed approach.