Long-Distance Multi-Object Vehicle Tracking And Recognition In Emergency Lane

With the gradual increase of vehicles,traffic problems have become increasingly prominent,and illegal occupation of emergency lanes has occurred from time to time,bringing huge hidden dangers to road safety.In order to improve the efficiency of intelligent supervision of illegally occupying the emergency lane,this paper is divided into three tasks to jointly obtain the information of illegal vehicles occupying the emergency lane on the road.To solve the problems of large number of vehicles,small size of long-distance vehicles and body occlusion in road images,a vehicle detection method based on inter-frame feature fusion and multi-scale feature fusion with spatial pyramid pooling is proposed in this paper and all vehicles in road images are located by using this method.Then,it is judged whether the vehicle is in the emergency lane through the emergency lane information provided in advance.When it is detected that the vehicle is in the emergency lane,the multi-target tracking technology is used to track the offending vehicle,and the trajectory secondary correlation matching strategy is introduced in the tracking process to improve the trajectory integrity of the low-confidence vehicle.Aiming at the extremely unbalanced problem that the proportion of social vehicles and special vehicles in road images is nearly 1,000,this paper extracts the image classification task from object detection as a separate module,and designs Focal Loss and adjustment sampling strategy for the class imbalance problem in the dataset to improve the classification effect of special vehicles.When the tracking vehicle drives to the front of the camera,the vehicle is cropped from the image and placed in the image classification network to filter out the special vehicles occupying the emergency lane.The main contents of this paper are as follows:1.A target detection method based on feature fusion between frames is proposed.Considering that the information difference between two adjacent frames in road video is very small,this paper processes the feature output of the previous frame and puts it into the current frame to provide prior information to improve the performance of the current frame.In order to improve the generalization performance of the model,four data augmentation methods are designed for the feature input head,which are used to simulate the target loss,false detection,and false localization that occur in actual reasoning.Finally,the method proposed in this paper has a 2.6%mAP improvement on the CenterNet detection algorithm.2.A multi-scale feature fusion network with spatial pyramid pooling is constructed.In view of the small size of long-distance vehicles in road image data,this paper uses a feature extraction network that incorporates the idea of multi-scale feature fusion——DLANet.At the same time,the spatial pyramid pooling structure is introduced to further enrich the feature fusion channel,realize the fusion of global features and local features,enrich the semantic information of shallow features,and improve the detection ability of small target vehicles.DLANet with improved structure has a 1.7%increase in mAP.3.Introduce the secondary trajectory correlation matching strategyByteTrack.The serious body occlusion in the road scene inevitably leads to low confidence in the detection results.Low-confidence targets do not participate in trajectory association in the traditional trajectory association strategy,which will lead to early termination of a large number of trajectories.Therefore,ByteTrack is introduced in this paper.After a round of trajectory association of high-confidence targets,a second round of trajectory association is performed on the remaining low-confidence targets to improve the trajectory integrity of low-confidence targets.After introducing the new association strategy,MOTA has a 2.3%improvement.4.In order to alleviate the category imbalance between special vehicles and social vehicles,this paper uses Focal Loss as the loss function of image classification tasks to increase the loss weight of special vehicles,and adjust the sampling strategy of network training to increase the proportion of training for special vehicles.In the final vehicle classification algorithm in the classification data set constructed in this paper,the accuracy and recall rate of special vehicles are 85.5%and 84.6%,respectively.