Application Of Improved YOLOv3-tiny In Vehicle Detection At Urban Intersections

With the rapid development of computer vision technology,intelligent application has appears widely in life.Intelligent traffic is the trend of urban development.However,the acquisition of road parameters mainly comes from the manual statistics and calculation of professionals,and its efficiency is low,which is not conducive to real-time guidance of traffic flow.In this paper,the target detection algorithm based on deep learning is used to identify and detect road vehicles at intersections.Under the experimental scene,the design of road vehicle parking detection,vehicle queue length and traffic flow statistics and detection.The main research contents are as follows:(1)The network structure of yolov3-tiny was studied and analyzed.When the trained yolov3-tiny model was directly applied to vehicle detection,obvious problems such as missing detection and incomplete detection would occur.To solve these problems,yolov3-tiny is reproduced on the experimental data set,and the backbone network of yolov3-tiny is analyzed and improved.Four different backbone network structures are designed: the first and second network structures still retain the direct connection model of yolov3-tiny,which is pooled by convolution substitution,and the third and fourth ones introduce residual structures.Training and testing were carried out under the same data set and training conditions respectively,and the experiment showed that the backbone network with the best performance was the fourth one.In addition,CBAM attention mechanism was introduced into the model with good performance,and attention weighting was carried out at the channel scale and spatial scale respectively.Through experimental analysis,the performance of the network with attention mechanism was improved in Recall,m AP and F1.(2)Through the study and analysis of this experimental scene,the edge contrast information in the scene is relatively obvious.The experiment proves that adding gaussian sharpening in the data enhancement can enhance the accuracy of the network,and adding noise in the data set can also improve the accuracy of the network.As the prior knowledge of the target frame,the Anchor value is closely related to the prediction result.Obtaining a more appropriate Anchor value in the data set is conducive to the model’s better learning and prediction.The experiment obtained the Anchor of the experimental data set through k-means.Compared with the test results of the original Anchor,the performance of the m AP was improved.(3)The improved yolov3-tiny model is applied to the event detection of vehicle parking,vehicle queue length and traffic flow at urban intersections.The detection rules and processes of three kinds of events are studied in detail,and the complete detection processes of vehicle parking detection,vehicle queue length and vehicle flow are designed.On this basis,the software of vehicle detection at urban intersections is realized by using Python3.7 development environment.