3D Object Detection Based On Deep Neural Network

As one of the important tasks of three-dimensional scene perception,three-dimensional object detection is a widely studied computer vision problem,and has wide applications in many fields such as autonomous driving,robotics,and augmented reality.In recent years,with the development of smart devices such as lidar and depth cameras and the gradual maturity of neural network technology,target detection in two-dimensional scenes is still insufficient for the description of three-dimensional real-world scenes.Appling deep neural networks to three-dimensional object detection based on point clouds has become more and more popular.An improved three-dimensional target detection algorithm is proposed to obtain the three-dimensional bounding box and its category geometric information such as the position,size and pose of the target in three-dimensional space.For the feature extraction of 3D point cloud data,it is improved based on PointNet,and an Attention Aggregate Residual Network(AARNet)is given.Using AARNet as the backbone network,a two-stage structure named ARR-RCNN is designed,in which the first stage generates three-dimensional candidate anchors based on the grid,and the second stage refines the candidate anchors to obtain the final detection results.In order to obtain a more accurate three-dimensional bounding box,when the non-maximum value is suppressed,the " Intersection over Union " is introduced as the sorting basis of the candidate anchors,thereby improving the confidence of the threedimensional anchors.In the feature mapping of the target area,coordinate conversion and feature enhancement are performed on the output features.The semantic segmentation experiments on the S3 DIS datasets show that AARNet has better performance than point methods for point cloud feature extraction.A two-stage threedimensional target detection experiment was carried out on the KITTI data set of autonomous driving scenes.The results show that the improved target detection algorithm can further improve the accuracy of the three-dimensional target detection while ensuring time complexity.