基于深度学习与稀疏光流的人群异常行为识别

Crowd Abnormal Behavior Recognition Based on Deep Learning and Sparse Optical Flow

目前公共场所人群异常行为检测的异常种类检测准确率较低,且多数对突然奔跑等部分异常行为无法识别.为此,提出一种基于YOLO_v3与稀疏光流的人群异常行为识别算法,通过检测小团体异常为群体异常预警与采取相应的应急措施提供充足的时间.为方便定位异常发生区域,将视频分割为多个子区域,通过获取子区域的图像样本进行诱发群体异常的小团体异常检测,利用改进YOLO_v3神经网络对传统算法较难检测行人持棍、持枪、持刀与面部遮挡等异常进行检测,在未检测到上述异常诱因时,使用稀疏光流法获取人群平均动能与运动方向熵,将得到的特征数据通过PSO-ELM进行分类,区分正常行为与同向突散或无规则突散.实验结果表明,与现有同类算法相比,该算法能有效检测行人持械与面部遮挡等小团体异常,并且定位异常发生区域的准确率达到98.227%.

The current detection accuracy rate of crowd abnormal behaviors in public places is low,and most abnormal behaviors such as sudden running cannot be recognized.Therefore,this paper proposes a recognition algorithm for crowd abnormal behaviors based on YOLO_v3 and sparse optical flow,providing sufficient time for early warning and corresponding emergency measures for crowd anomalies by detecting small group anomalies.In order to locate the abnormal area conveniently,the video is divided into several sub-areas and the image samples of sub-areas are obtained to detect small group anomalies that induce crowd abnormality.The improved YOLO_v3 neural network is adopted to detect anomalies that are difficult for the traditional algorithms to detect,such as carrying stick,gun and knife and face occlusion.When these anomalies are not detected,the sparse optical flow method is used to obtain the average kinetic energy and the entropy of the movement direction of the crowd,and the obtained characteristic data is sent to PSO-ELM for classification,distinguishing normal behaviors from co-directional spur or irregular spur.Experimental results show that compared with existing similar algorithms,the proposed algorithm can effectively detect small group anomalies such as pedestrian armed anomalies and facial occlusion anomalies,and can locate the area where the abnormality occurs with an accuracy rate of 98.227%.