Research On 3D Object Detection Algorithm Based On Deep Learning

In recent years,powered by the strong performance of deep learning,one of the most important task of computer vision 2D object detection has made a great breakthrough and has been widely used in many fields of social.However,2D object detection only detect and locate the object in the planar image,and it cannot meet the challenges in autonomous driving,mobile robot and other applications due to the lack of critical depth information.Differ from 2D object detection,3D object detection needs detect the object in real 3D space with depth information,and finally provide such as the position,direction and size of the object.Therefore,3D object detection is a vital part in autonomous driving,mobile robot and other fields.At present,there is still a big gap between the performance of 3D object detection and 2D object detection.Therefore,the research on 3D object detection algorithm has not only profound theoretical value,but also broad application prospect.Therefore,the research contents of this thesis are as follows:1.Summarizes the current 3D object detection algorithms and related theories,introduces and analyzes the research status of 3D target detection algorithms,classifies 3D object detection algorithms according to different data types,and analyzes their advantages and disadvantages.Then introduces the concept of deep learning,explains the structure and working principle of convolutional neural network in deep learning network,and analyzes the important function of convolutional neural network in object detection.2.Research on feature extraction algorithm of point cloud based on deep learning.This thesis first expounds the key steps of 3D object detection,then introduces the data structure characteristics of point cloud,and studies the current point cloud feature learning algorithm based on deep learning,analyzes mechanisms of different network structure,according to their merit and demerit,proposes an improved point cloud feature extraction method.3.By analyzing 3D object detection algorithms based on point cloud,aiming at the existing problems such as unnecessary computation,low detection efficiency and error detection,propose a redundant point filtering-based 3D object network.Firstly,2D detector is used to locate the object in RGB image,and the point cloud space where the object within is extracted through the frustum proposal,filter the redundant points,reduce the amount of calculation and increase the effective search space.The proposed algorithm is trained and tested on the KITTI data sets,and the experimental results are compared with other 3D object detection algorithms,the detection results are significantly improved.