Software Development For Long-Distance Pedestrian Detection System Based On Background Modeling

With the development of information technology,intelligent video surveillance technology has been widely studied and applied in many fields.And pedestrian detection,as a branch of intelligent video surveillance technology,has been widely used in prison detention centers.There are two main pedestrian detection methods used in intelligent monitoring system,which are based on traditional moving object detection method and deep learning method.Background modeling method is fast,but there are many false detections and missed detections in complex environments.The method based on deep learning performs better in detection accuracy,but it is necessary to downsample the monitoring image to improve the calculation speed to achieve realtime,which leads to the small pedestrians in the distance can not be detected and affects the monitoring distance.Prison detention center,as an important place of national judicial organs,has high requirements for the accuracy,real-time performance and monitoring range of pedestrian detection technology.In order to solve the problem that the monitoring system of prison detention center can't accurately detect long-distance small-scale pedestrians in real time,this thesis developed a long-distance pedestrian detection system software based on background modeling.The system is based on NVIDIA Tegra Xavier processor platform,and the whole system is divided into video processing module,system business module and algorithm module.The video processing module realizes the receiving and pushing of streaming media as well as video decoding and encoding.The system business module realizes the communication between the system and the PC monitoring platform,manages the upgrade of the system module,and records the working state of the system.The algorithm module uses an improved background modeling algorithm to detect the foreground of the surveillance image,extracts the area where the target is located for stitching,and performs pedestrian detection based on the stitched image with YOLOv3 detection network,so as to avoid the need to compress the original image directly,which leads to the inability to detect long-distance pedestrians,and at the same time filters out the false detection targets caused by dynamic background.According to the requirements of real-time monitoring system,the detection model is compressed,and the embedded GPU and Tensor RT inference acceleration framework on Xavier platform are used to improve the processing speed of the algorithm.The system test results show that the long-distance pedestrian detection system based on foreground image stitching of this thesis improves the foreground detection algorithm,reduces the false detection caused by dynamic background and improves the integrity of the target,and accurately identify the long-distance pedestrians with the size of 10 × 20 pixels on the 1920 × 1080 image.