Research On Vehicle Recognition And Tracking Algorithm Based On Deep Learning

With the application of artificial intelligence technology in vehicle,vehicle automated driving will be ease traffic congestion Reduce the number of traffic accidents An important means to improve the driving comfort Autopilot system including environmental perception path planning and decision-making control three parts,including the environment perception technology provide data support for automatic driving system Environmental awareness based on vision sensors because of the price is low,the advantages of simple structure and is widely used In recent years,deep learning has achieved good results in the field of visual perception thanks to the mining of deep features of images by deep learning.At present,target recognition technology and target tracking technology based on deep learning are the most potential direction in the field of visual perception.In this paper,the vehicle recognition and tracking algorithm based on deep learning is studied.The specific research contents are as follows:(1)Multi-scale vehicles in order to solve the vehicle identification process,it is difficult to identify the problem,adopted a method of multi-scale information: this method before YOLOv3 three forecast layer to join the pyramid structure of pooling,and pooling the pyramid structure to join the convolution,by reducing the number of channel control model size Pooling pyramid structure is composed of four parallel maximum pooling layer,to extract the different feelings of deep characteristics of wild,use the method to improve the ability of identifying multi-scale model.(2)In order to solve the problem of vehicle overlap in the process of vehicle recognition,an optimized non-maximum suppression method is adopted: in the process of non-maximum suppression,all the boxes will be eliminated if the average crossover ratio between the boxes with the highest score and other boxes is greater than or equal to the threshold.But there might be targets in those boxes.Therefore,the optimized non-maximum suppression method is adopted in this paper,and Gaussian function is used for attenuation when the average crossover ratio is greater than the threshold value.Using this method,the ability of the model to recognize overlapping targets is improved.(3)In order to improve the accuracy of vehicle identification algorithm,an improved default anchor frame method based on K-means++ is adopted: K-means++ algorithm is used to optimize the initial clustering center to avoid the uncertainty caused by random clustering.At the same time,the average crosswise ratio is used as the index to measure the similarity of labels.The closer the average crosswise ratio is to 1,the smaller the clustering distance is.(4)In this paper,an online multi-target tracking method based on recognition(Deep Sort)is adopted.Firstly,the motion state of the target is estimated and the motion information of the target is obtained.In this paper,we retrain our vehicle rerecognition model on the VERI data set to replace the original pedestrian model,extract the depth characteristics of the vehicle and obtain the apparent information of the vehicle.The cost matrices of the target motion similarity and the apparent similarity are obtained by cascading matching method.The Hungarian algorithm matches the real frame and the tracking frame according to the output value of the cost matrices.The experimental results show that the accuracy of the optimized YOLOv3 model on the data set is 89.76%.Compared with the original algorithm,the accuracy is improved by 6.48%.In terms of target tracking,compared with SORT algorithm,parameters of tracking box obtained by Deep Sort algorithm are closer to parameters of data set annotation box.After combining vehicle apparent information with the algorithm in this paper,the value of MOTA is increased by 4.7%,and The Times of identity exchange is reduced by 42%.Through the test in the actual traffic scene,the method has a good effect on vehicle identification and tracking.