Research On Object Detection Methods In Outdoor Street Scene

With the development of information technology and the increase of people's travel demand,how to use the most advanced technology in the field of outdoor street scenes has become a difficult problem for many scholars and enterprises.In recent years,Deep Learning technology has developed rapidly,which directly promotes the rapid progress of object detection technology and makes it possible to apply object detection technology in the field of outdoor street scenes.The 2D object detection technology can detect the class and location of common objects in the camera's image,and the 3D object detection technology can detect the class and the space location of common objects in the image and point cloud.Therefore,they can be used for the precise detection of common outdoor street objects,such as traffic signs,cars,pedestrians,etc.,so as to contribute to fields of traffic management,path planning,automatic driving,etc.This thesis attempts to improve and upgrade the most advanced object detection technology used in outdoor street scenes so as to promote its development and practicability.The main content of this thesis is as follows:1.This thesis studies an outdoor street object irregular target detection method based on multi-rectangle combination and model compression.In order to avoid a huge amount of calculation of image segmentation method,this thesis adopts the method of multirectangle combination,by combining smaller rectangular boxes generated by CNN network into a larger polygon,to fit the outline of irregular objects,and reduces the overlap between candidate boxes through the improved non-maximum suppression method at the same time.In order to meet the requirements of rapid processing of traffic scenes,this thesis adopts the highly efficient MobileNet-SSD Deep Learning model,and further speeds up the processing through model compression,so that the model can run quickly on the mobile GPU computing platform.2.This thesis studies a 3D object detection method for outdoor street scenes images based on size clustering and edge detection.In order to solve the problem of information missing in the 3D object detection task of monocular image,this thesis extracted the edge information of the object and added these features to the regression network.At the same time,through the observation of size distribution,this thesis finds that the variance of similar object's sizes is not uniform in each dimension,so this thesis proposes to use multiple candidate sizes to predict the size of the object,in order to reduce the difficulty of size regression in the detection.3.This thesis studies a 3D object detection method for outdoor street scenes images based on label optimization and point cloud instance segmentation.In this thesis,a new point cloud instance segmentation network is proposed,which directly operates on point cloud data,and it optimizes information transmission and improves detection accuracy by means of skip connections and iterative network.Meanwhile,this thesis has observed the problem of inaccurate annotation in KITTI dataset,and provides a simple and fast optimization method,by enlarging the dimension of annotations,which improved the accuracy without any increase in calculations.In order to verify the effectiveness of the proposed algorithm,this thesis constructs an object detection database for outdoor street scenes.The experimental results show that the proposed algorithm has a good performance in object detection tasks for outdoor street scenes.