Research On Vehicle Detection And Tracking Algorithms In Traffic Monitoring Scenes

In recent years,vehicle detection and tracking algorithms based on deep learning show excellent performance and have gradually replaced traditional algorithms in traffic monitoring scenes.However,deep learning algorithms consume a lot of computing resources and storage memory.At present,deep learning algorithms are mainly deployed in the cloud,to realize data analysis and research.It is difficult to achieve real-time feedback under the explosive growth trend of data.Based on the above problems,edge computing comes into being,which processes and makes decisions on the edge side of the network,so as to realize the rapid response to traffic events.In this paper,the deep learning algorithm is deployed to the embedded platform based on the idea of edge computing.By referring to the solution of current multi-object tracking algorithm and using the multi-object tracking framework based on detection,a lightweight multi-vehicle detection and tracking algorithm is designed to achieve accurate and real-time tracking of vehicles within the monitoring range.The main research contents of this paper are as follows:(1)Inspired by lightweight object detection algorithm,weighing the detection accuracy,running speed and model size,a lightweight object detection algorithm R-YOLOv3 based on feature fusion is proposed.Firstly,based on YOLOv3-Mobilenetv2,the structure of backbone convolutional network is improved to enhance the capability of feature extraction.At the same time,FPN multi-feature fusion network and multi-scale aggregation module are designed to enhance the information fusion among multi-scale feature images.In addition,Dilated convolution is introduced to construct receptive fields of different sizes to improve the ability of feature extraction for targets of different scales.Then,combined with KITTI data set,K-means algorithm is adopted for dimension clustering to get the appropriate prior box parameters.Finally,experiments on the KITTI detection data set show that R-YOLOv3 is suitable as a lightweight detector for vehicle detection tasks.(2)Aiming at the situation of identity switch in multi-vehicle tracking,an improved multi-vehicle tracking algorithm L-Deepsort based on Deepsort is proposed.Firstly,using long short-term memory network instead of Kalman filter,a completely data-driven trajectory prediction algorithm is designed to solve the prediction error caused by Kalman filter in nonlinear system.Then,introduce a Siamese Network and use Mobilenetv2 as a sub-network to train the appearance of lightweight vehicles matching model;At the same time,improve the appearance matching strategy of Deepsort tracking algorithm to enhance the relevance between targets.Finally,the algorithm performance is verified on the KITTI-Tracking data set.The experiment shows that L-Deepsort meets the real-time requirements with improved tracking performance.(3)Aiming at the problem that deep learning based object detection algorithm is difficult to deploy in embedded platform,R-YOLOv3 model is accelerated.Firstly,the R-YOLOv3 detection model is accelerated by using Tensor RT inference optimization framework of NVIDIA embedded platform.Then,the optimized detection model is adopted as the detection module of the L-Deepsort tracking algorithm,and the detection-based multi-vehicle tracking algorithm RYL-Deepsort is constructed.Finally,the algorithm is deployed to the Jetson TX2 platform and tested in a real scene.The experiment shows that the detection based multi-vehicle tracking algorithm RYL-Deepsort can improve the speed significantly without obviously reducing the tracking performance.The paper has 42 pictures,13 tables,and 86 references.