Research And Application Of Vehicle Detection And Tracking In Complex Traffic Surveillance Video

In the process of intelligent transportation development,intelligent analysis of road traffic monitoring video is a technology worthy of attention.There are still some technical difficulties in vehicle detection and tracking as key linkages.For example,when the driving environment is complex,the scale difference of vehicle objects is large,and the vehicles are blocked by each other,vehicle detection is prone to miss detection and low accuracy,and the vehicle tracking effect will also deteriorate,which directly affect the accuracy of vehicle behavior recognition.Aiming at the above problems,this thesis studies the multi-object detection and tracking algorithm of vehicles,and applies the obtained vehicle trajectory information to vehicle behavior recognition,aiming at reducing the rate of missed detection and false detection,improving the accuracy of vehicle detection and tracking,and then realizing vehicle behavior recognition more accurately.Obtaining the necessary traffic information quickly and accurately plays an especially important role in solving traffic congestion and improving traffic safety.The main research work of this thesis is as follows:(1)Aiming at the problem that complex scenes and large differences in object scales may cause false detection and missed detection,a vehicle detection network model based on YOLOv5 s combined multi-scale and attention mechanism is proposed.Firstly,we improve the regression loss function and replace the original GIOU loss with SIOU loss.Then the CA attention mechanism is integrated into the backbone network.Finally,Bi FPN is introduced into the enhanced feature extraction network.The experimental results on UA-DETRAC dataset and self-collected dataset show that compared with the original YOLOv5 s network,m AP@0.5 of the improved model increases by 2.4% and 3.6% respectively,and the problem of missing and false detection of vehicles is improved to some extent.At the same time,compared with the existing methods,the detection model proposed in this thesis has achieved relatively good results.(2)Aiming at the problems of object loss and low tracking accuracy of tracked objects due to occlusion,a vehicle tracking model based on improved Deep SORT is proposed.In the appearance branch,the large vehicle recognition Ve Ri dataset is used to retrain the appearance feature extraction network,and the exponential moving average feature update strategy is used to replace the feature library of the original Deep SORT algorithm.In the motion branch,NSA algorithm is used to replace the traditional Kalman filter algorithm.In the data association stage,GIOU measurement is used to improve IOU matching.The experimental results on the public UA-DETRAC dataset show that compared with the original Deep SORT algorithm,MOTA and IDF1 value of the improved Deep SORT algorithm are increased by 6.9% and 3.2% respectively.(3)Using the vehicle track information obtained from vehicle detection and tracking,a vehicle behavior recognition model is constructed.Firstly,a vehicle trajectory model is established based on the set of center point coordinates of the vehicle’s bounding rectangle obtained from tracking each frame.Then,based on the coordinate transformation of the vehicle trajectory model,determine the direction of vehicle movement,and determine the vehicle behavior based on the vehicle movement direction and the direction of the vehicle entering the intersection,establish a vehicle behavior recognition model to identify different vehicle behaviors.The validity of the model is verified by experiments on the self-collected dataset.