Research On Object Detection Algorithm Of Urban Management Road Monitoring Image Based On Deep Learning

At present,most cities in China have built large-scale road monitoring networks to achieve full coverage of cameras,which leads to the explosive growth of urban management road monitoring video.However,the urban management video monitoring center still uses the way of manually watching road monitoring videos to find violations,which has the problems of shortage of manpower and low efficiency.By using the object detection technology based on deep learning,not only can we identify the violations that affect the appearance of the city and disturb people’s life,such as illegal parking and road occupation,but also can realize24-hour continuous monitoring and early warning.Although the object detection algorithm based on deep learning has good performance on public data sets,in the real environment,due to illumination,weather,occlusion,camera jitter,and other reasons,the image quality collected by street cameras is uneven,which leads to low object detection accuracy of urban management road monitoring image.Given the above problems,this thesis improves the object detection algorithm based on deep learning to solve the problem of object detection in urban management road monitoring image in a real environment.The main results of this research are as follows:(1)Make urban road monitoring image data set.To improve the detection accuracy,a total of 32197 images are collected from the urban management video monitoring center.To make the training algorithm more robust,the urban management road monitoring image data set includes not only the images taken from different angles,different distances,different brightness,different weather,and different time intervals but also the images of occlusion,blur,and dense scenes.Combined with the actual needs of urban management,this thesis constructs five categories of illegal objects,including illegal billboards,out-of-store operation,motor vehicle illegal parking,road occupation operation,non-motor vehicle illegal parking,and subdivides them into 16 categories.Finally,Label Img software is used to label all the urban management road monitoring images into Pascal VOC data set format.(2)An improved Cascade RCNN object detection algorithm for urban management road monitoring images is proposed.Firstly,to improve the classification accuracy,the algorithm uses a split attention network as the backbone network;secondly,to further enhance the feature extraction ability,the algorithm uses a cascaded feature pyramid structure;thirdly,to improve the detection accuracy of dense objects,the algorithm uses soft NMS algorithm;fourthly,to improve the positioning accuracy,the algorithm uses CIOU loss function.The experimental results show the effectiveness of the four improved methods,and the detection accuracy of the improved Cascade RCNN algorithm is significantly improved.(3)An improved object detection algorithm of urban management road monitoring image based on YOLOv4 is proposed.Although the detection speed of YOLOv4 is fast,the accuracy of the urban management road monitoring image data set is not very high.To improve the detection accuracy of YOLOv4,the batch normalization in the path aggregation network is replaced by group normalization,and the mish activation function is applied to the non-backbone network of YOLOv4.To separate the context features of the image and increase the receptive field,two spatial pooling pyramid modules are added.In addition,the focus loss parameters suitable for the algorithm are determined through some experiments,and the performance of the improved yolov4 algorithm is significantly improved.