Research On Vehicle Recognition In Video Streaming Environment Based On Deep Learning

As one of the important technology branches of intelligent transportation system in our country,vehicle recognition technology of video flow in video flow environment is one of the key technology modules of intelligent transportation system in our country.However,in the process of practical research and application,there are complex video flow traffic environment and scene,such as the phase of multiple interference environment factors such as multi-illumination,multi-angle and occlusion Due to the mutual influence,the accuracy of vehicle detection and recognition is low.Therefore,a high accuracy,high real-time and high robustness algorithm for vehicle recognition in video stream is undoubtedly needed at present.The main contents of this paper are as follows:1)Aiming at the problem of low accuracy of vehicle detection caused by the interference of complex environmental factors in the process of vehicle recognition and vehicle detection,this paper improves a vibe moving vehicle detection algorithm.First,the algorithm adaptively updates the threshold parameters of the current vibe algorithm,occlusion and so on on on the possibility of low accuracy in the current vehicle recognition detection process.Second,combining with the three frame difference method,the "ghost shadow" that may exist in the original vibe vehicle recognition algorithm is eliminated The phenomenon.The experimental results show that the algorithm can effectively eliminate the phenomenon of "ghost" and improve the reliability and accuracy of vehicle recognition and detection.2)In this paper,a kind of residual network model Wide-Res Net based on deep learning is improved to solve the problems of slow speed and low accuracy of vehicle recognition caused by different camera position and resolution,and the direct influence of environment factors such as illumination in complex environment.Firstly,this paper improves the loss function of the residual network recognition model,improves the recognition rate of vehicles from different angles,and solves the problem that the recognition results of the same vehicle from different angles are different.Secondly,it improves the residual structure of the residual network Res Net,increases the width of the residual module,reduces the depth of the neural network,and effectively improves the recognition rate of vehicles from different angles The accuracy and speed of vehicle identification can meet the real-time and accuracy requirements of vehicle identification system.The experimental results show that the performance of the algorithm is superior to other vehicle recognition algorithms.3)Through the implementation of the prototype system,the improved vehicle detection algorithm and vehicle recognition network model are applied and verified.The system first reads in the video stream,then extracts the moving vehicle pictures in the video streamaccording to the improved vehicle detection algorithm,then trains the improved residual network model,inputs the detected moving vehicle pictures into the trained network model for vehicle classification,the network model will output the vehicle classification results,and the final results will be displayed in the system interface.