Research On Multi-camera Object Handoff And Tracking Technology Based On Deep Learning

With the development of intelligent surveillance system,especially the application of moving object detection and tracking technology,video surveillance system has ushered in a real change.Then human and material resources are liberated.At the same time,with the increasing scale of video surveillance systems,the shortcomings of small field of view and loss of information are becoming more and more obvious in single camera system.Multi-camera video surveillance system can make up for the limitation of single-camera,so this system has become a new research direction.In the indoor environment,continuous tracking of moving object in multi-camera non-overlapping fields of view has always been a research focus and difficulty in the field of computer vision.In this paper,the massive data of surveillance video is fully utilized to study the multi-camera handoff and tracking based on deep learning.1.The performance of moving object detection and tracking of single-camera has an important influence on the moving object handoff and tracking of multi-camera.In order to achieve accurate and continuous tracking of moving object between multiple cameras,it is necessary to study the moving object detection of single camera.According to the visual characteristics of the surveillance camera,the face is selected as the detection object to reduce the occlusion between people in the indoor surveillance environment.The face detection model is established to improve the accuracy of face detection based on MTCNN(Multi-task Cascaded Convolutional Networks),and the model is trained using massive data from surveillance video.The experiments results show that the accuracy of face detection is over96% in indoor video surveillance environment,which can meet the accuracy and real-time requirements of moving object detection.2.Based on the face detection model,the key points and difficulties of multi-moving object tracking are analyzed in indoor single camera.In this paper,according to thecharacteristics of MTCNN,MTCNN and Kalman filter are combined to improve the tracking accuracy of multi-moving object.Firstly,the motion state of moving object is predicted by Kalman filter.Then,MTCNN sub-network is used to detect the face at the predicted position of the object,and the position of the face frame is determined.Finally,the relationship between object is established by overlapping of the face frames.The experimental results show that the combination of MTCNN and Kalman filter can effectively accomplish the tracking task of multi-moving object in a single camera.3.Multi-camera moving object handoff has always been a research difficulty in the field of computer vision.Especially in non-overlapping field of view,the discontinuity and uncertainty of moving object make it more difficult for object handoff.According to the characteristics of object feature matching and the research on detection and tracking of moving object,the face is used as the research object in the process of moving object handoff.In this paper,The face feature extraction model is established based on deep convolution neural network,and two face similarity measure methods are adopted to select the optimal face similarity measure method for face matching.The experimental results show that the deep convolution neural network can extract the face features of moving object accurately and complete the task of the multi-camera moving object handoff and tracking effectively.