The Design And Implement Of A System About Traffic Violations Recognition In The Vehicle-mounted Video

Drive behavior recognition is a heat topic in the field of machine vision,and principal research target in automobile aiding driving industry.Compared with goal recognition,behavior recognition has special space and time,so it's hard to get the results from normal website's bigger data collection only.Though aiding driving industry develops very fast,for the limitation on behavior recognition calculations,now it mainly depends on sensors and GPS positioning systems to collect space and time data.But these systems limited by environment conditions.It cannot work in the places such as channels,underground and other uncovered areas.To this issue,the author put forward a recognition way to violations behavior based on visual algorithm.There are three parts for the research of the essay.(1)Statical object identification of important elements in vehicle-mounted video,this paper proposes a Mobile Net-DSSD network which is dedicated to solving the main problem of statical traffic element identification in vehicle-mounted video,firstlity,multi-scale classification and location for the traffic elements;secondly,the robustness of small target recognitionnetwork;Thirdly,the requirements of lightweight network.Based on DSSD network,this paper replaces the backbone network with Mobile Net after adding convolution layer from resnet-101 network,modifies the feature fusion mode of DSSD network deconvolution module and adjusts the serial number of feature fusion layer.Aiming at the three recognition requirements of multi-scale classification and location,small target detection robustness and lightweight network architecture,this paper proposes the Mobile NetDSSD network.The experimental results show that the average accuracy of recognizing all kinds of traffic elements in vehicle video is 74.91%,the average accuracy of recognizing traffic lights,pedestrians and other small targets in vehicle video is 76.4%,and the convolution computation of the network is 7.9 million times.Compared with SSD network,the recognition accuracy of Mobile NetDSSD network is increased by 12.85%,and the convolution computation is reduced by 95.28%.(2)The scene state recognition for driving scenarios in the vehicle-mounted video,this paper proposed a scene state recognition method that can be used to identify traffic light intersections and zebra crossing intersections.scenarios recognition is always a difficulty in the field of automatic driving to be realized.driving scenarios in the vehicle-mounted video has strong spatial position characteristics and temporal displacement characteristics.While the vehicle video image is in twodimensional pixel plane,and the spatial characteristics of the scene can not be effectively interpreted by directly changing the position of pixel coordinates.In this paper,the spatial imaging model of the vehicle camera is simplified,and the spatial transfer matrix is calculated by four effective vanishing points in consecutive frames,and the spatial relative coordinates of the target in the image are further estimated.Combined with the temporal characteristics of the process of vehicle approaching the target scene,The experimental results show that the recognition accuracy rate is 0.70,the recall rate is 0.77,and the recognition accuracy rate is 0.71 and the recall rate is 0.58 for the vehicle arriving at the zebra crossing scene.(3)Recognize violation behaviors in vehicle video.Compared with the rapid development of single target tracking with fixed viewing angle,the recognition of violations in vehicle video is a multi-target tracking system with constantly changing background information.meanwhile,the one who violates the regulations will not appear directly in vehicle video.Violations need to be indirectly recognized by the space and location between various background references in the video.In this paper,the author analyze the violations by space and time sequence from status such as at the edge of violation,being violations,and after violations.And puts forward a programming idea for violation behavior migration,and through background information and the law of motion of traffic signals to recognize the driving vehicle stop and driving status.Through the adjustment of threshold,recognition effect of the system can reach optimization.And the adjustment of framerate can meet the need of liability of this system which finally successfully recognize those who violate the traffic rules such as red light running and ignoring zebra crossing by vehicle video.The experimental results show that the system can meet the requirements of real-time monitoring when the frame rate is set to 3.At this time,the recognition accuracy of the red light running recognition system is 0.63,the recall rate is 0.75,the accuracy of the zebra crossing behavior recognition system is 0.85,and the recall rate is 0.66.