Car Violation Detection System For Car Recorder Video

With the development of society,the number of private cars has been increasing.However,due to the limitation of China’s traffic control system,the supervision of car violations is very limited,and many car violations cannot be punished,resulting in frequent car violations.In this thesis,we analyze and process the video data of the car recorder to determine whether there is a car violation in the video,so as to punish the violation and strengthen the supervision of car violations and create a good traffic environment.This is a new and challenging field.With the rapid development of deep learning and computer vision,related technologies have been widely used in many fields.In this thesis,deep learning algorithms are applied to detect car violations.The useful information in the video is extracted using deep learning algorithms,which involves knowledge in the fields of object detection,object tracking,text detection and recognition.In this thesis,YOLO is used to do object detection,which is mainly used to detect cars,pedestrians,traffic lights,etc.in the video.At the same time,this thesis also trains a front and rear detector based on the YOLO network.For the detection and recognition of license plates,due to the limitation of YOLO for small object detection,this thesis uses SSD to detect the license plates of cars,and then uses a CNN-based sequence recognition model to do the recognition of text.In this thesis,all these methods are combined together and tested and analyzed separately using public datasets,and finally,we also use our own video data for testing and analysis to verify the feasibility of the algorithm.Finally,the license plate number of the car,the location of the vehicle and the front and rear of the car were obtained from the video of the car recorder.Second,after getting the basic information of the vehicle also need to extract the road information,the most basic is the detection and classification of lanes.Due to the particularity of lanes,this thesis proposes a method of first detection and then classification to realize the detection and classification of lanes.For lane detection,this thesis adopts the current mainstream example segmentation method to achieve,using PINet as the lane detection network to extract lane information from the video.Then for each lane instance obtained by segmentation,a special feature point extraction algorithm is used to extract the feature points of the lanes and then classify them by CNN network.In this thesis,both detection and classification of lanes are tested and analyzed using public datasets and our own video data,and the advantages,disadvantages and feasibility of the algorithms are analyzed.Finally,this thesis uses Deepsort algorithm to track the vehicles in the video based on the obtained vehicle information and lane information,so as to integrate and match the basic information of each vehicle in the video.We design and implement the detection algorithm for two types of violations,namely vehicle retrograde and lane changing,and design and implement an asynchronous server that can be called by multiple users by ourselves,and also deploy the whole algorithm to this server,so as to realize a complete car violation detection system oriented to the driving recorder video.