Research On Vehicle-road Cooperative System Detection Method Based On Behavior Analysis

In recent years,breakthroughs in hardware and software technologies have promoted the rapid development of vehicle-road collaborative technology.Vehicle-road cooperation technology provides computer vision inspection function by camera and embedded system.However,in practical applications,considering the cost,computing power and the limitation of the memory of embedded devices,it is a very challenging problem to realize the high-precision identification of the blocked crowd at the intersection by the vehicle-road cooperation system in the real-time state.Drivers are prone to traffic accidents when driving in the complex environment of traffic junctions.In view of this phenomenon,China’s traffic department also issued new traffic rules of comity to pedestrians,the introduction of new traffic rules to a certain extent to reduce the occurrence of traffic accidents.However,in the dark environment at night,it is still difficult for drivers to observe the path of pedestrians in the complex traffic system.Especially for the more travel taxi master is difficult to completely solve the traffic intersection potential dangerous situation.Aiming at this problem,a vehicle-road cooperative system which can warn and monitor pedestrians in real time is proposed.A pedestrian detection model and a pedestrian pose estimation model are constructed to monitor and analyze pedestrian behavior in the vehicle-road cooperative system.The spatial pyramid module（SPP）is established in YOLOv3 network.By learning BDD100 K,Street Scenes and the data set of dangerous Scenes made,the vehicle-road cooperative system can accurately identify the problem of pedestrians in dangerous Scenes.Then the optimized YOLOv3 detection model was improved by combination pruning.Based on this improved strategy,we compared the optimized YOLOv3-Faster algorithm with the original YOLOv3 algorithm.YOLOv3-Faster has a more compact network structure and fewer floating point operations.As a feasible scheme,it can be applied to pedestrian target detection in vehicle-road cooperation.Compared with the unoptimized YOLOv3,the optimized YOLOv3-Faster achieved excellent detection performance,with its floating point operation reduced by about 78.50%,trainable parameters reduced by 99.1%,model space reduced by99.02%,and detection time shortened by 41.79%.The average accuracy（m AP）of the optimized YOLOv3-Faster is increased by 17.16% compared with that of the original YOLOv3.The above experimental results prove that the optimized YOLOv3-Faster has simplified running space and excellent detection accuracy and speed compared with the original YOLOv3.Established a behavior analysis algorithm optimized based on RMPE,The multi-person pose estimation framework based on RMPE needs to locate pedestrians in complex traffic environments before conducting behavior analysis.Pedestrians are located using the optimized target detection algorithm YOLOv3-Faster.In order to further optimize the detection efficiency of the traffic system,Based on the RMPE multi-person pose estimation framework,the correlation between cross-frame pose and pose flow,To optimize the efficiency of behavior flow tracking,and combined with the behavior flow non-maximum suppression algorithm PF-NMS to improve the overall function of the behavior analysis algorithm.Finally,the network model is trained,tested and verified on the open behavior analysis data set,and the efficiency of behavior flow tracking is improved obviously.Collaborative system adopts the intelligent Vehicle-road collaboration cameras（RGB camera,long wave infrared cameras,communication module and industrial PC）,intelligent traffic lights（traffic lights,traffic lights control board and the communication module）,variable information board,on-board unit joint target detection and behavior analysis algorithm assisted comity pedestrian monitoring and early-warning prompt path of ascension.