Research On Vehicle Detection Algorithm Based On Deep Learning

Vehicle detection technology is a hotspot in the field of computer vision,and it is also the research basis for license plate detection,vehicle segmentation,vehicle tracking and other applications.Due to the complicated scenes and environments of vehicle detection,especially when there are too many distant vehicles or vehicles in the scene to block each other,the difficulty of detecting the vehicle is increased.In this dissertation,the moving vehicles in the actual road scene are taken as the research object,the YOLOV3 algorithm is selected as the basic algorithm by comparing the real-time and accuracy of current mainstream target detection algorithms comprehensively.Aiming at the problems existing in traditional vehicle detection algorithm,the YOLO-v3 algorithm in deep learning is improved from three aspects,which effectively improves the detection ability of the algorithm for distant vehicles and occluded vehicles.Firstly,the pyramid-pooling algorithm is used for feature enhancement,and different maximum pooling sizes are used for feature pooling experiments to select the best pooling sizes.Secondly,the K-Means algorithm is used to cluster the sample data set,and the appropriate size and number of preselected boxes are selected by analyzing the IOU(intersection over union)between the prediction boxes and the number of prediction boxes.Finally,The non-maximum suppression method is modified to make the IOU between the prediction boxes attenuate in a certain way when the IOU is larger than the set threshold.At the same time,the algorithm proposed in the dissertation is combined with large-scale feature layer prediction and local ROI enlargement of input image,which further enhances the detection ability for smaller vehicles and occluded vehicles.By using ResNet50,ResNet152,DenseNet201 and Darknet networks as feature extraction networks of YOLO algorithm,the advantages of the proposed algorithm are further illustrated.The dataset used in the dissertation can be divided into two categories: manual data acquisition and UA-DETRAC dataset.The vehicle in five scenarios of sunny day,rainy day,night,occlusion and cloudy day are tested and compared,which shows the reliability of the improved algorithm.This algorithm is compared with Faster-RCNN,YOLO-v1,Tiny-Yolo,YOLO-v2 and YOLO-v3 algorithms in terms of recall,precision and F1 value,respectively.The advantages of the proposed algorithm are illustrated.Through testing UA-DETRAC dataset,the accuracy rate is 94%,the recall rate is 97%,and the detection speed is 32 FPS on NVIDIA 1080 graphics card,which meets the real-time requirements.