基于超像素时间上下文特征的异常行为检测

Abnormal Behavior Detection via Super-Pixels Time Context Feature

为了更精准地定位拥挤视频中的异常行为,提出了基于超像素时间上下文特征的异常行为检测算法。特征表示阶段,对视频帧进行超像素分割,判断超像素是否属于前景。根据超像素的灰度直方图和位置信息找出其在相邻帧中最匹配超像素,计算最匹配超像素的多尺度光流直方图特征均值作为超像素特征,以增强超像素特征在时间上的联系。检测阶段,采用稀疏组合学习算法对超像素进行异常判断。实验结果表明,该算法在UCSD和UMN库上的检测效果优于现有异常检测算法。

In order to accurately locate the abnormal behavior, an anomaly detection method based on time context features of super-pixels is proposed. For feature representation, the video frames are firstly segmented into super-pixels. The super-pixels of foreground are then selected according to their pixel ratios of foreground. Super-pixels matching adjacent frames are selected based on the gray-level histogram and the information of location to enhance the temporal context of super-pixel features. The statical value of multilayer histogram of optical flow of matched super-pixels are taken as the feature for detection. In the phase of detection, the sparse combination learning algorithm is adopted to detect abnormality. Experimental results show that the algorithm outperforms other state-of-the-art algorithms in the UCSD and UMN video libraries.