Research On Multi-Vehicle Detection And Tracking Algorithm Based On Deep Learning

From a global perspective,the trend of “new four modernizations of automobile” is irreversible.As an important part of “new four modernizations of automobile”,vehicle intelligence is the focus of current major automakers and component companies.The quality of smart car's environment perception ability can measure its intelligence level to a certain extent,and the advancement of smart car's perception will promote the development of vehicle intelligent technology.Pre-vehicle detection based on monocular vision helps the vehicle to have the ability to sense the surrounding environment,analyze the possibility of vehicle collision in advance,reduce accidents,and ensure the safety of drivers,passengers and pedestrians.In this paper,the vehicle environment perception method is researched based on the deep learning algorithm.An optimizated vehicle detection method based on YOLOV3 algorithm is proposed.Combined with multi-object tracking algorithm,the detection and tracking of surrounding vehicles are realized.Convolutional neural networks have a good performance in image recognition and object detection.This paper introduces the composition of convolutional neural network,and analyzes the working principles of the current four mainstream object detection algorithms,combined with their performance on public dataset VOC.The possibility of using the YOLOV3 algorithm combined with the lightweight feature extraction network for road vehicle detection to operate in embedded hardware is analyzed.In order to realize the detection of vehicles,the YOLOV3-tiny model is used as the basic network to train the vehicle data in the KITTI dataset.The method of improving the accuracy of the model by optimizing the backbone network is proposed for the situation of poor network detection.The two models,Darknet19 and Resnet18,were used to replace the original feature extraction network,and the detection accuracy was improved by 3.30% and 8.39%,respectively.Aiming at the problem that the model does not detect small objects,the K-means algorithm is used to cluster 9 Anchor boxes,and multi-scale experiments are carried out on the original network.Finally,the detection accuracy is improved from 79.57% to 93.66%,and The hardware TX2 development board tested the speed of the network model.By analyzing the current multi-target tracking algorithm,the SORT algorithm based on "tracking-by-detection" is selected to perform multi-vehicle tracking on the road.Firstly,the Kalman filter is used to predict the Bounding box parameters of each detected object,and the optimal state parameter of Kalman filter is selected through experiments.Then,use the Hungarian algorithm to correlate the detection result and the tracking result.The IOU value of those two results is used as a measurement parameter to obtain the correlation result between the frame and frame.In order to assign different IDs to the same object with occlusion,it is proposed to use the color histogram of the object to represent the color features.The color features of each object are represented by a 192-dimensional vector,and the cosine angle of two vectors is used to represent the similarity between two objects.The product of the cosine angle and the IOU value of the two objects is used as a measure of the Hungarian algorithm matching for data association.The experimental results show that compared with the original algorithm,the proposed algorithm can greatly reduce the number of transformations of the target ID on the basis of multi-object stable tracking.