Research On Multi-camera Video Target Tracking Based On Deep Learning

With the deepening of the research on deep learning,many target-tracking algorithms based on the deep learning are proposed.The accuracy and robustness of the traditional tracking algorithms based on artificial selection of target feature are not high in the actual application for shade,illumination changes and target deformation.The successful application of deep learning in image recognition and classification has benefited from its strong ability to represent the target.At present,the target tracking based on deep learning benefits from the advantages of network model,computer technology and related algorithms,and it has significantly exceeded the traditional tracking methods in the tracking effect.The object-tracking methods based on deep learning proposed in recent years are studied in this thesis,and a new model is proposed which combines the MDNet(Multi-Domain Network)deep learning tracking model and the modified Faster R-CNN target detection network to achieve the goal of multi-camera tracking and applied this method to the station and building surveillance.In this thesis,the tracking network and the target detection network were re-trained using the labeled samples that collected by ourselves.The experimental results show that the proposed algorithm has better robustness and accuracy than the traditional method,and can achieve the goal of multi-camera stable tracking better.The main contributions of this thesis are as follows:Firstly,based on the existing deep network model,the network was re-trained using the labeled pictures that shot and collected from the actual scene by ourselves.The MDNet network used in this thesis is composed of shared layers,fully connection layers and multiple classification layers during the training.In the period of online tracking,the positive and negative sampling strategy,learning rate and the update strategy of network parameters are improved in this thesis.The algorithm proposed in this thesis is applied to the station and building video monitoring and achieved a better tracking effect.Secondly,modify the Faster R-CNN target detection network and retrained the network using the labeled images that collected by ourselves.In this thesis,the objects of interest in the video image are be detected by the object detection network,and do not need to complete the classification.Therefore,this thesis shields the network structure of object classification.Thirdly,combine the target-tracking network with the modified target detection network and a fusion algorithm is proposed to achieve continuous tracking of the target.When the tracking target disappears in the current video sequence,it is necessary to detect and locate the target in the other surveillance video.On the basis of the detection of many objects by using the modified target detection network,then the proposed algorithm can be achieved the target localization in this thesis,and finally call the tracking network again to achieve the goal of continuous tracking,that is multi-camera video sequence target tracking.