Research On Multi-scale Vehicle Detection Algorithm Based On Convolutional Neural Network

The rapid development of automobiles has brought people convenient life,but also caused problems such as traffic congestion,and the emergence of intelligent transportation systems can alleviate this problem well.Vehicle detection is the most important part of the intelligent transportation system,and its performance directly affects the application effect of the entire system.At present,vehicle detection methods based on deep learning have made great progress,but in practical applications,the problems of different vehicle scales in the image,complex and changeable scenes,and occlusion are still difficult to solve,and vehicle detection is still a challenging Mission.This article mainly studies vehicle detection algorithms based on deep learning.The research contents are as follows:For vehicle detection problems,first introduce the basic knowledge of convolutional neural network,explain its network structure and analyze the characteristics of each layer;then introduce different feature extraction networks,and analyze the advantages and disadvantages of these networks;and finally introduce the volume-based The target detection algorithms of the product neural network include Two Stage target detection algorithm and One Stage target detection algorithm,and analyze the advantages and disadvantages of these algorithms.Aiming at the multi-scale problem in vehicle detection tasks,a vehicle detection algorithm based on multi-scale is proposed.The algorithm is based on the most commonly used multi-scale target detection framework FPN(Feature Pyramid Network)as the basic model.First,deconvolution is used to replace the double upsampling in the original network from the top to the bottom of the path to enhance the deep semantic information;second Add the equalization module,and connect the prediction layers of the original FPN to this module to equalize the output of the different scale layers in the original network to obtain more balanced feature information;then add the Nonlocal module after the equalization module,The feature map after equalization is further enhanced to enrich feature information;finally,the enhanced feature map is fused with the original output and multi-scale prediction is performed.Experiments on the Pascal VOC and MS COCO datasets show that the improved FPN improves the detection accuracy of vehicles of different types or different scales.Aiming at the problem of large difference between the size of the candidate frame in the vehicle detection and the actual vehicle size,an algorithm based on K-means ++to improve the size of the candidate frame is proposed.The algorithm uses the K-means++ algorithm to cluster the experimental data set to obtain the optimal candidate frame generation size.The improved candidate frame size is closer to the actual vehicle size,which is conducive to the detection network to learn vehicle features more fully.In order to further improve the vehicle detection performance,the training strategy is optimized,and the vehicle detection network is optimized using the OHEM method and the IOU-Balanced method,respectively.Finally,experiments were conducted on the Pascal VOC and MS COCO datasets.The results show that the algorithm based on Kmeans ++ to improve the size of the candidate frame generation is used to optimize vehicle detection,which improves the detection accuracy to a certain extent.Compared with the OHEM method,IOU-The Balanced method is better for vehicle detection network optimization.By combining two optimizations,the performance of vehicle detection can be significantly improved.