Cross Line Counting Of Crowd Based On Single Camera And Implementation Of Embedded System

As the monitoring system becomes more and more popular in today's life,the research on the system is getting deeper and deeper.Due to the general trend of low cost,the relevant deployment conditions have become more and more stringent.The detection and tracking of the crowd can easily achieve cross-line counting and other tasks.On the practical application level,it can realize the analysis of the main trend of the crowd,passenger flow statistics and other very practical project landings,and provide more accurate traffic scheduling basis for some public places.Therefore,it is of great significance to build an efficient system that can complete multi-target detection and tracking tasks in real time.The focus of this work is to solve the problem of pedestrian detection and counting based on monocular camera.The camera is installed with a downward-looking solution,which can solve the problem of pedestrian occlusion when the crowd is dense to the greatest extent.Based on a comprehensive understanding of multiple target detection methods,we finally selected YOLO in the One-Shot category and redesigned the backbone feature extraction network to achieve fast target detection.In order to achieve the purpose of realtime detection on the embedded board,the Kalman filter is used to solve the problem of target position detection in the next frame.The Hungarian algorithm is used to solve the target matching problem between frames.Therefore,the target trajectory chain can be tracked in real time.According to the set access area,the pedestrian's cross-line behavior is counted.Based on the improved algorithm,it is also deployed on the RK3399 development board to conduct relevant real-time comparative research.It is found that the tracker can make up for the lack of prediction accuracy in single-frame target detection by combining related frames before and after.This also means that the actual significance of faster prediction is greater than when a single frame detection has a higher mAP and the prediction is time-consuming.The prediction speed of the method proposed in this paper is significantly faster than that of the designed controlled experiments.Basically meet the needs of practical applications,and provide new ideas for further research.