Video Target Detection Based On Deep Learning And Their System Implementation

Intelligent video surveillance is now a hot topic in the field of computer vision.And with the development of network technology,digital video technology and the monitoring hardware,video surveillance is becoming more and more intelligent.Intelligent video processing that without human intervention can use computer vision technology and video analysis do the process of the video image sequence which captureed by the surveillance network,achieve the target of the positioning,identification and tracking of the targets,and give the analysis and judgment of the behavior of the targets,which can not only do the daily management work but also respond quickly in the event of abnormal circumstances.The object detection is one of key steps in the video surveillance analysis system and is also the also the fundamental preprocessing for the subsequent advanced applications(such as target tracking and recognition).In this paper,the video object detection method based on the deep learning network is studied,and the video object characteristic is modeled by the multi-layer nonlinear structure of the depth learning network.Considering that some existing video target detection algorithms are not efficient,accurate and robust in different scenes,this paper combines the merits of the Faster RCNN and ResNet.The Faster RCNN is suitable for fast and accurate detection of common video target while ResNet has priority in the detection of small objects.In this paper,we propose a new method called ERF-Net(Efficient Residual Faster rcnn)Depth network structure.The experimental results show that the proposed method achieves a better performance in the robustness of scene changes,light and shade and occlusion.We validate the proposed method on the standard datasets and our own datasets,which testify that the proposed method is efficient,accurate and robust and is very suitable for detecting small objects.