Research And Implementation Of Real-time Detection Method For Abnormal Events In Video Surveillance

With the improvement of people's living standard and the enhancement of their safety consciousness,more and more attention has been paid to intelligent video analysis technology.The purpose of intelligent video analysis technology is to analyze and process the video sequences automatically with the knowledge of computer graphics,pattern recognition and machine learning,and abnormal behavior detection in a specific scene is an important research direction in the field of intelligent video analysis.The abnormal behavior detection technology can detect the abnormal behavior in video automatically by extracting the motion information from the video frames,then send out the warning information,which saves a lot of manpower cost.In recent years,abnormal crowd behavior occurs frequently under the specific outdoor scene.So as the facial occlusion abnormal behavior.These abnormal behaviors may cause irreparable damage to the community.But it is difficult for the traditional monitoring system to detect these abnormal behaviors on time.In this thesis,these two problems are studied deeply,and the main contents and results are as follows:1.A real-time detection method of abnormal crowd behavior based on auto-encoder is proposed.At present,most of the abnormal crowd behavior detection methods are based on optical flow,so these methods suffer a lot to achieve real-time requirements.In this thesis,we construct local spatio-temporal features based on motion vector,which takes advantage of the ability to extract motion vector in real time during decompression of the video.Then we train multiple auto-encoders relatively for each local small area by using the spatio-temporal features.In the course of the test,we use the pre-trained multiple auto-encoders to determine whether there is an abnormal event in each small area.The specific location of the anomaly is located with the comprehensive analysis of the situation of each small area and whether there is a global anomaly will be determined.The proposed detection method is tested in a public dataset,and the accuracy is more than 95%.At the same time,the processing time of each frame is about 25ms.2.A real-time detection method of facial occlusion abnormal behavior based on convolution neural network is proposed.The existing methods of detecting the abnormal behavior of facial occlusion are usually to locate the position of the head firstly,and then model the facial organs such as eyes and mouth,so as to judge whether there is an abnormal occlusion indirectly.These methods can lead to false detection and missing detection easily due to the ambiguity of facial organs.This thesis focuses on the two most common types of facial occlusion:wearing sunglasses and wearing masks.And we model the masked face directly to determine whether facial occlusion occurs.In order to meet the real-time requirement and have a good performance,this thesis makes improvement on the traditional methods in many ways,such as moving foreground extraction,head location,skin color detection,which makes the position of the face(the region of interest)be located more accurately.The experimental results show that the combination of the face location and the face occlusion anomaly model proposed in this thesis can meet the requirements of real-time and high accuracy.3.According to the above research results,the real-time monitoring system of abnormal behavior detection is designed and implemented to validate the proposed method.