An Algorithm For Abnormal Behavior Recognition In High Definition Videos

Recognizing human behavior is a hot topic of artificial intelligence, which can be widely used in security monitoring, industrial automation, human-computer interaction and video content rating and other fields. High-definition surveillance equipments improve the resolution of the videos, which not only rich the details of the image but also improves the amount of the data which needs to be deal with. How to recognize the abnormal behaviors fast is an important problem which needs us to solve.Detecting high-speed moving targets from moving camera is always a complex and time-consuming process. To solve that problem, a fast target detection method is proposed based on motion vectors. In this paper, the data format and decoding features of surveillance videos are analyzed. Then, the ways of obtaining motion vectors directly from the video stream are concluded. Besides, the motion vectors are normalized by taking the reference frames into account and the global motion is detected by analyzing the histogram of the motion vectors. The targets are extracted by analyzing the statistical property of the motion vectors. Experiment results prove that our approach can extract the high speed moving targets from moving camera effectively.Extracting features from abnormal behaviors is always difficult. To solve that problem, a fast features extracting algorithm based on motion vectors is proposed. First, we extract the inter-frame features including the histogram of motion vectors, entropy of motion vectors, directional mean of motion vectors, and so on. Then, the intra-frame features are extracted which include histogram difference and intensity difference. Finally, we propose the descriptor of abnormal behaviors. Experimental results prove that the features extracted by our method can describe the different between each behavior clearly.Classifying the abnormal behaviors is always a complex process. To solve that problem, a three-level support vectors machine to classify the abnormal behaviors is proposed. First, we classify the fighting behaviors from other behaviors by the entropy of motion vectors. Then, for these behaviors which are not fighting, we use directional histogram difference to judge whether they are dropped behaviors. Finally, we classify the remaining behaviors by a seven dimensionality features. An extensive set of experiments are performed and this method is compared with some of the most recent approaches in the field by using publicly available datasets as well as a new annotated dataset, which results prove that our approach can detect the abnormal behaviors in the video quickly.