Intelligent Transportation Vehicle Detection And Tracking Based On Deep Learning

With the development of computer technology and the improvement of science and technology,people’s demand for the quality of life is also increasing day by day.Traffic is an essential part of daily life,traffic congestion,frequent traffic accidents and other problems have attracted the attention of many countries,intelligent transportation system has become the future direction of traffic development.Among them,vehicle detection and tracking is an important link to realize intelligent transportation.Due to increasingly complex traffic conditions,traditional vehicle detection and tracking methods have been unable to meet the needs.Deep learning,as the most popular artificial intelligence implementation method in recent years,has unique advantages in image detection.Therefore,based on the deep learning method,this paper researches the detection and tracking method of intelligent transportation vehicles.The specific research contents are as follows:(1)Propose SSD-MobileNetv3 integrated network vehicle detection method for vehicle detection problems.Combined with the advantages of fast detection speed of SSD network and small number of MobileNetV3 network,the basic network VGG of SSD network is replaced by MobileNetV3,and the transfer learning method is combined with the pre-training model to accelerate the convergence of the network.At the same time,in view of the SSD network’s low accuracy in small target detection,a deconvolution network was added to improve its small target detection performance,and the self-built vehicle data set was used for experiments.The experiment shows that the number of SSD-MobileNetV3 network is significantly reduced by 83.1%compared with the original SSD network,and the accuracy is also slightly improved by 3.1%.Although the network parameters of the improved SSD-MobileNetV3 are 8.9m higher than before,the overall accuracy of the improved SSD-MobileNetV3 is 4.1%higher,and the detection of small targets is more accurate.(2)Deep Sort,a detect-based tracking method,is used for vehicle tracking,and SSD-MobileNetv3 designed above is used for detection.In view of the phenomenon that the Sort method is easy to track and lose due to the lack of vehicle appearance features,the vehicle appearance re-recognition is added,and the vehicle appearance features are obtained by using the VERI data set for training,and the matching cascade method is used to solve the tracking target loss problem to a certain extent.The results show that the Deep Sort vehicle tracking method in this paper is superior to the Sort algorithm in each index parameter.The method realizes vehicle counting and vehicle trajectory tracking through the intersection.(3)Build an embedded experimental platform to carry out experimental verification of the improved Deep Sort vehicle tracking method in this paper.Verify whether the algorithm in this paper can run in the embedded system,and compare the operation of different algorithms in the embedded system.The experiment shows that the algorithm in this paper can run smoothly in the embedded platform,and the FPS is around 18,which can basically realize real-time detection,which is helpful to the realization of intelligent transportation vehicle tracking.