Real-time Detection Method Of Suspicious Person In Surveillance Video

In recent years,with the social public security has received more and more attention,the use of surveillance videos has become popular,using these surveillance videos to detect suspicious persons has also become an important research topic.Suspicious person detection means using computer vision,pattern recognition or other methods to determine whether there are suspicious persons in the surveillance video,and if so,get their position.Suspicious persons are those who are abnormally different from ordinary pedestrians in appearance,for example,using hat and mask to cover the face or ducking from the camera.Suspicious person detection technology can help users effectively supervise passing pedestrians and respond to different kinds of emergencies.This thesis focuses on the positioning and discriminating problems of suspicious person detection,and carries out related researches around the pedestrian position detection algorithm and suspicious person identification algorithm.The main research contents and contributions are as follows:1.For the pedestrian position detection problem,considering that the suspicious person's features are mainly concentrated in the head-facial regions,and in the actual scene,it is often impossible to obtain a complete pedestrian area,this thesis proposes the following two class algorithms to detect the pedestrian position:(1)Face detection algorithm based on cascaded convolutional neural network.The cascaded network is composed of two convolutional networks,the first one is used to quickly filter the background area,and the second one is used to accurately identify the human face.Finally,the cascaded network achieves the detection of face regions with a multi-scale sliding window.(2)Head-shoulder detection algorithms based on Faster R-CNN(Faster Regions with Convolutional Neural Network)and YOLO(You Only Look Once).These algorithms use the end-to-end detection model Faster R-CNN or YOLO to detect head-shoulder regions directly.Through the detection of the face or head-shoulder region,the position of the pedestrian is indirectly confirmed,and at the same time,we obtained a pedestrian's feature concentration area for subsequent algorithms.Experimental results show that the detection algorithm proposed in this thesis has achieved good results in terms of accuracy and detection speed.2.For the suspicious person identification problem,taking the pedestrian's head-shoulder region as the research object,this thesis proposes a suspicious person identification algorithm based on multi-task CNN(Convolutional Neural Network)and one-class SVM(Support Vector Machine).Firstly,we design a multi-task CNN classification model for the attributes of the head-shoulder region.After optimizing the calculation speed,this network is used to extract features of the head-shoulder images.Then,considering that the suspicious person samples are scarce and all suspicious cases cannot be exhausted,this thesis uses one-class classification algorithm to indirectly identify suspicious persons,and designs a one-class SVM classifier based on composite kernel function.Finally,experiments show that the suspicious person identification algorithm proposed in this thesis has good accuracy and real-time performance.3.Using the above research results,combined with some pre-processing steps,such as foreground extraction based on motion vectors,this thesis has constructed a verification system for real-time detection of suspicious persons,and verifies the effectiveness of the proposed algorithm under the actual scenario.