Research And Application Of Vehicle Detection And Tracking Algorithm For Traffic Survey Scenes

Purpose of Traffic survey is to understand the characteristics of traffic flow and collect the data such as traffic flow,flow direction and vehicle type composition.In the traffic survey scenario,traffic flow,flow direction and vehicle type composition are the basic data that traffic survey needs.The automatic collection of the above data can be realized by using the vehicle detetion and tracking algorithm.This thesis intends to use a deep learning-based vehicle detection and tracking algorithm to automatically count traffic flow and other data to improve the efficiency of traffic surveys.There are three challenges in vehicle detection and tracking in traffic survey scenarios.The first is the real-time requirement.The algorithm execution speed should be fast,otherwise the traffic flow statistics will be inaccurate.The second is the vehicle sub-classification problem.In the traffic investigation scenario,the classification of vehicle types is more detailed,which increases the difficulty of vehicle detection.The third is the problem of vehicle occlusion.During the driving process of the vehicle,lane change,overtaking or yielding may occur,which will affect the tracking effect of the algorithm.In view of the above difficulties,the main work and contributions of this thesis are as follows:Aiming at real-time requirements for vehicle detection speed is not fast enough and the problem of vehicle detection accuracy,a vehicle detection algorithm based on lightweight feature fusion and attention enhancement is proposed.Based on the YOLOX algorithm,reducing the number of channels through 1×1 convolution can ensure the accuracy of the model and reduce the model inference time.Since traffic survey pays more attention to medium and large targets that are closer to the camera,the upsampling operation adds a multi-scale detection and feature fusion network output and directly fuses shallow features with deep features,which improves the model alignment between medium and large objects detection accuracy.The channel and spatial attention mechanism are used to improve the detection head and the classification accuracy of vehicle type is imporoved.In comparative experiments,The effectiveness of the improved algorithm is verified by using the constructed vehicle detection dataset.For real-time vehicle tracking and vehicle occlusion problems.A multi-task learning framework for simultaneous object detection and re-identification feature generation is proposed.Based on the improved YOLOX algorithm,a re-identification branch is proposed to generate vehicle re-identification features.In order to alleviate the problem of tracking ID exchange of different vehicle types,the data association algorithm in vehicle tracking is optimized by separating the tracking process of different vehicle types.In comparative experiments,The effectiveness of the improved algorithm is verified by using the constructed vehicle tracking dataset.Based on the vehicle detection and tracking algorithm proposed in this thesis,a traffic survey system based on Huawei’s software-defined camera is implemented,the result of system is shown,and the accuracy loss of the model is analyzed.