Design And Implementation Of Vehicle Counting System Based On YOLOv3

Video surveillance has become a crucial barrier in the field of public safety,especially in the application of traffic scenarios.With the increasing number of cars,more and more attention has been paid to the related issues in traffic scenarios.Traffic congestion not only severely reduces travel efficiency,but also increases the possibility of causing traffic accidents.Accurately grasping the traffic flow of urban sections is conducive to alleviating urban congestion and facilitating traffic management departments to plan travel routes for citizens.Because of the high complexity of traffic scenes,the different illumination intensity in different time periods,and the high speed of algorithm processing,it brings great challenges to the task of target detection.The traditional target detection algorithm based on sliding window is lack of pertinence,and there are a large number of redundant windows,which makes the algorithm more complex and runs longer,which is not conducive to the fast processing of video stream data.In addition,because the feature relies on manual design,it does not have good robustness in the face of complex traffic scenarios.With the advent of the era of artificial intelligence,more and more related fields,such as automatic driving,face recognition,intelligent medicine and so on,have started the research process of AI.Deep learning technology has developed rapidly.Target detection algorithm based on this technology also provides more choices for solving complex traffic scene problems.In the task of vehicle detection in traffic scene,the two-stage target detection algorithm represented by RCNN and the single-stage target detection algorithm represented by YOLO provide a new solution for vehicle detection task,which is helpful for us to further count the number of vehicles after vehicle detection so as to obtain vehicle flow parameters.In the past,microwave detectors and electromagnetic induction are used to detect traffic flow.The target detection method based on deep learning can use the existing monitoring equipment as the video data input source,which has a larger detection range and is easier to install and maintain.Based on the current popular YOLOv3 target detection algorithm,this paper uses the improved K-means clustering standard to generate new anchor boxes from the data set used in this paper,in order to get a more optimized priori box to improve the accuracy of the target box.In addition,vehicle image data in traffic scenes are used to train the model.The original 80-class detection model is simplified to a single-class vehicle detectionmodel,which improves the detection speed and accuracy.After the video data passes through the target detection network,the coordinates of the two adjacent frames are saved by combining the adjacent frame center matching algorithm.According to the nearest center distance,the target matching is completed.If the boundaries of the two adjacent frames of the target are located on the opposite side of the preset detection line,the target passing and counting are indicated.The system is divided into online and offline vehicle counting modes.In order to facilitate the traffic management department's dispatch and command and the analysis of traffic flow data,the results of vehicle counting in different sections can be transferred to Aliyun ECS MySQL database,and further data visualization can be completed through DataV.