| With the development of smart city construction,Urban street view images,As a new way of map service,are applied in city construction.The main way of image service is street view images map browsing service.Object detection in street view images map browsing service is one of the popular directions of computer vision research.Object detection in urban street view has become an important part of comprehensive improvement of urban governance ability.However,there are still some problems to achieve the object detection effect in the street view images: the background complexity of the street view,the image retrieval time is long,the pedestrian and vehicle detection in the complex scene are faced with the scale changes,appearance and attitude change,target repetition and occlusion.Aiming at the problems of object detection in street view images,this paper studies an improved Yolo V3 deep learning algorithm to detect pedestrian and vehicle information in street view images and improve the speed and accuracy of object detection.The main research contents of this paper are as follows:1.The development course of target detection is systematically summarized.This paper expounds the basic flow and frame structure of traditional object detection from the theoretical part and introduces the convolutional neural network,the main part of deep learning.And two main representative algorithms in deep learning are analyzed as the comparison part of the experimental conclusion of this paper.2.According to the characteristics of urban street view images object detection,an improved algorithm is proposed,which mainly improves the Yolo V3 algorithm from the aspect of precision and speed.This paper introduces the idea of advanced semantic embedding(SEB)in the aspect of precision,and fuses the high-level features and low-level features of the image.In terms of speed,the idea of lightweight detection end is introduced to improve the speed of image output.In the experimental part,different algorithms are compared to obtain the real effectiveness of the improved algorithm.The results showed that the MAP value reached 66% and the Mean Precision value reached 85% on the street view images data set.The main research content for the object detection is in the city street view images.According to different experimental data set deep learning algorithm to meet the requirements of scenario object detection,One adopts the public KITTI data set,This data set choose the typical image of more than 7000 images data sets and format conversion to the network training process of the algorithm.The other uses the road street view images data of Xicheng District,Beijing.More than 4,000 representative images are selected from the data set.It is necessary to mark the data set of pedestrian and vehicle information in the image data set and participate in network training.Finally,the improved algorithm is obtained by using the precision,speed evaluation index and the output result of the image.It has good detection performance for pedestrians and vehicles in street view images. |
