Research On Vehicle Target Detection And Tracking Algorithm Based On Deep Learning

The use of automated vehicle driving technology and intelligent traffic management systems can effectively reduce the incidence of road traffic accidents and ease the pressure on urban traffic as the number of motor vehicles in China continues to increase and urban road traffic congestion and accidents become more and more serious.The detection of and tracking for vehicle targets is crucial to autonomous driving and intelligent traffic systems,but a complex road environment and severe mutual occlusion between vehicles make vehicle detection and tracking ineffective.Deep knowledge has been developed rapidly in recent years,and the application of deep learning in vehicle detection and tracking is a current research focus in the field of machine vision.So this paper will use deep learning as the basis to study the vehicle detection and tracking method,the main research work and improvement points of the paper are as follows.(1)The convolutional neural network principle and structure are firstly analyzed,followed by two different types of target detection methods,and the two methods are compared,finally the YOLOv5 s version algorithm with better detection effect and lighter weight is chosen as the basic detection algorithm to implement this topic.For the tracking algorithm,this paper chooses DeepSORT algorithm to implement vehicle target tracking,it has real-time and good tracking effect.(2)In regard to vehicle detection,the YOLOv5 s algorithm is improved for complex road scenarios.The specific improvement objectives and measures are: in order to improve the real-time vehicle target detection,the feature extraction network in the original YOLOv5 s algorithm is improved into a BiFPN network to achieve a balance between speed and accuracy;in order to improve the accuracy of small target vehicle detection in complex scenarios,the SA attention mechanism;finally,the original loss function is improved to SIOU loss function,which can be effective in improving the wrong detections or missed detections of small targets.The experimental results show that the m AP of the improved YOLOv5 s algorithm is improved by 3.1% and has better accuracy for vehicle detection.(3)For the vehicle tracking algorithm,which is difficult to track because the car is obscured,this paper adopts the combination of improved YOLOv5 s and tracking algorithm to track the vehicle.By retraining the feature extraction network using the vehicle re-identification dataset,the deficiency of missing vehicle-related features in the feature extraction network of the traditional DeepSORT algorithm is improved.The StrongSORT algorithm is then used,with the appearance that the feature bank in the traditional DeepSORT algorithm is replaced with an exponential mean shift EMA.Motion-wise,to address the problem that the traditional Kalman filter ignores the scale message of the detection noise,the StrongSORT algorithm gives an adaptive calculation of the noise covariance formula with the help of the NSA Kalman filter algorithm,thus improving the effect on vehicle tracking.For different road scenarios,the UA-DETRAC dataset was transformed to obtain a tracking test set that meets the form of multi-target tracking requirements.By evaluating the performance of the improved DeepSORT algorithm and StrongSORT algorithm on this test set,the improved DeepSORT and StrongSORT algorithms have improved the MOTA and MOTP metrics by 3.8% and 2.9%respectively,while the StrongSORT algorithm has improved the MOTA and MOTP metrics by 5.9% and 6.4% respectively,the improved vehicle tracking model The improved vehicle tracking model is able to ensure better tracking results while ensuring real-time performance.