Study On Surveillance Video Object Detection Based On Deep Learning

With the increasing of security awareness,the surveillance system becomes huge day by day.But the development of surveillance system brings new challenges.Researchers have introduced computer vision to develop intelligent surveillance system,enabling the computer to have the ability of video analysis and object detection in surveillance video.The traditional video object detection algorithms can't satisfy the needs of the actual applications,such as inter frame difference,background subtraction,and optical flow field,and so on.The object detection based on deep learning can solve this problem well.So in this paper,we taken the campus surveillance video of Tianjin University of Science and Technology as the research object and introduced the characteristics of campus surveillance video.Based on that,we analyzed the problems existed in the campus surveillance video object detection with SSD network and given improvements.The main contents of this paper as follows.Firstly,we introduced the research status of object detection,as a result of overall consideration of detection accuracy and speed of these detection algorithms,we selected SSD as our object detection algorithm.Then,we trained and tested SSD using VOC dataset.Secondly,we applied SSD to campus surveillance video to have a test,deeply analyzed the result of detection and the main problems existed with SSD network structure and its Implementation principle.Thirdly,in campus surveillance video,most of the objects belongs to small objects because of the large scope of vision of Surveillance camera,this brings difficulties to the object detection.Aim at this problem,we combined image region division with SSD to achieve the goal of object detection,this method could increase the size of object in input images.We also analyzed the impact of object proportion on the detection result.The test results show that the method of image region division not only effectively improves the detection effect of daytime surveillance video,but also improves the performance of SSD in the night surveillance video,it's further confirms the effectiveness of the method.