基于行人组运动信息表达的人群异常检测

Crowd abnormal detection based on pedestrian group movement information expression

人群异常检测是智能化人群监控技术下的一个重要研究方向,在现有的方法中,异常检测的首要步骤就是获取运动信息,传统通过对视频帧均匀分块的方式并不能保证行人的完整性,提取的特征也不能准确反映行人的运动状态。本文提出了递进式人群分组的方法,先将人群运动场与帧差法结合分割图像获取人群前景,然后依据人群运动方向获取方向组,结合时空信息对方向组再聚类,得到更细致的行人组。对于每个行人组,利用人群能量特征去表征行人整体运动信息,并依据能量场构造了环块能量直方图特征来削弱行人四肢摆动的影响,最后与图像外观特征相结合用于人群异常检测。实验结果表明,本文方法在两个不同场景下帧级准确率达到83%和92%,像素级准确率达到64%和83%,与传统方法相比有较大提升。

The crowd anomaly detection is an important research direction under intelligent crowd surveillance technology. In the existing methods, the first step of anomaly detection is to obtain motion information. The traditional way of uniform chunking of video frames does not guarantee the integrity of pedestrians, and the extracted features do not accurately reflect the motion status of pedestrians. In this article, an incremental crowd grouping method is proposed, which firstly combines crowd motion field with frame difference method to segment the image to obtain crowd foreground. Then, the direction group is achieved, which is based on crowd motion direction. Finally, Spatio-temporal information is combined to re-cluster the direction group to get a more detailed pedestrian group. For each pedestrian group, the crowd energy feature is used to characterize the overall pedestrian motion information, and the ring block energy histogram feature is constructed based on the energy field to weaken the effect of pedestrian limb swing, and finally combined with the image appearance features for crowd anomaly detection. Experimental results show that the proposed method achieves 83% and 92% accuracy at frame level and 64% and 83% accuracy at the pixel level in two different scenes, which is a significant improvement compared to the traditional method.