Research On Vehicle Detection And Tracking And Vehicle Speed Detection Algorithms Based On Video Images

Intelligent vehicle detection,tracking,and speed detection have significant implications for improving urban traffic management and addressing traffic safety issues.However,the existing video-based vehicle detection and tracking technology faces various challenges such as poor detection robustness,high error rates,and significant tracking ID switch caused by weather,target occlusion,and deformation.To address these issues,this paper focuses on the research of vehicle detection,tracking,and speed detection algorithms in different traffic scenarios.The goal is to improve the accuracy of tracking and detection,obtain vehicle travel speed,and detect speeding events.The main contributions of this paper are as follows:1.Based on the high-precision and high-speed YOLOX-s object detection algorithm,this paper proposes an improved strategy for addressing the intra-class occlusion of vehicles.Firstly,Ghost convolution is used to replace the ordinary convolution in the YOLOX-s backbone network,which reduces the model’s parameter size and improves detection efficiency.Secondly,the CBAM attention mechanism is introduced to enhance important features and weaken unimportant features,thus optimizing the model.Finally,the aspect ratios of predicted and detected boxes,and the distance between their center points are considered to improve the loss function and non-maximum suppression,making the detection box more precise.At the same time,a data set is created,and the original YOLOX-s and improved YOLOX-s algorithms are trained and tested on the self-made data set to obtain a vehicle detector.Experimental results show that the improved YOLOX-s algorithm outperforms the original YOLOX-s in detection performance,achieving a detection speed of77f/s and a detection accuracy of 92.8%,which is 1.7% higher than the original model.2.Based on the results of object detection,this paper uses the DeepSort multi-object tracking algorithm to track the detected vehicles and proposes an improvement strategy for the frequent ID conversion problem of DeepSort.Firstly,an improved wide residual network is used instead of the original residual network,and the image input size of the wide residual network is modified to 128×128 to make the model more suitable for the traffic scenario in this paper.Secondly,the center loss function and softmax loss function are combined to improve the feature extraction ability of the convolutional neural network.Finally,the improved DeepSort is trained and tested on the basis of the improved YOLOX-s to obtain the final vehicle tracking results.The experimental results show that the tracking accuracy of the improved DeepSort algorithm has been improved to some extent,and the number of ID switch has also been reduced.3.This paper improves the YOLOX-s object detection and DeepSort multi-object tracking algorithm to detect vehicle speed and judge speeding behavior.Firstly,camera calibration technology is used to transform the pixel coordinate system into the world coordinate system.The actual displacement of the vehicle is calculated by the pixel distance of the vehicle’s movement position,and the real-time vehicle speed is obtained.Secondly,the speed obtained by this paper’s video-based speed measurement method is compared and analyzed with the speed obtained by manual calibration,demonstrating the feasibility and effectiveness of the proposed speed measurement method.Finally,a speeding detection model is established to achieve vehicle’s speeding behavior detection.The experimental results show that the proposed speed measurement and speeding behavior detection method is stable,effective,and has practical application value.