Computer Vision-based Pedestrian Unsafe Behavior Detection Technology And Implementation In Public Places

In this peak period of accelerating basic construction in our country,many cities are building public places which served the prople to satisfiy the appetite for convenient and comfortable life,based on the continuous improving of building structure required by the social public.However,as a public and essential place with crowd people,the probability of accidents increase rapidly in these years.The human stampedes and violence happended these years not only caused a large waste social resources,but also brought theat to people’s physical and mental health.Under this background,by the real-time video monitoring of public places with high incidence of accidents,through the methods of detection,re-identification,tracking,and Behavior Identification,we can forecast the accident before it occurs.The main topic of this paper is below:(1)Aiming at the non-uniqueness of safety accidents in public places,this paper classifies all kinds of pedestrian behaviors that are easy to cause accidents.Considering the personal feature such as clothing,route direction and actions,this paper regards different behaviors as three color judgment categories such as green,yellow and red.When behaviors are judged as green,it indicates that the behavior is normal.If the behavior is judged to be yellow,it indicates that the behavior of the pedestrian is different from that of the general behavior,for example,stepping on the yellow line or lines that people have no permission to enter.When the behavior is judged to be red,it indicates that the target is carrying out fights or other illegal acts that seriously threat social security.(2)For the real-time detection of the unsafe behaviors in public places,this study use pedestrian detect algorithm based on YOLOv3 algorithm to detect pedestrian.This study take advantage of Res Net50 network to solve the problem of pedestrian re-identification across cameras.In the further research of sort tracking algorithm,this study use Deep Sort algorithm to do real-time tracking on pedestrians inside the buildings.Finally,the real-time detection and tracking of pedestrians in public places are realized by the algorithm combination framework including YOLOv3 pedestrian detection algorithm,Res Net50 pedestrian re-identification algorithm,and Deep Sort algorithm.(3)After analyzing and comparing various mainstream behavior recognition algorithms,we finally choose Open Pose algorithm to carry out attitude recognition for pedestrians in public places.Moreover,we use ST-GCN algorithm to carry out action recognition of attitude data.Through these methods,we realized the real-time recognition of pedestrians behavior in the monitoring video.