Deep Learning-based 3D Object Detection In Point Cloud

Object detection is used in numerous applications,ranging from autonomous driving to robot vision.Autonomous driving vehicles use multiple cameras and single/multiple LiDAR sensors for object perception.Comparing with camera,LiDAR can generate accurate 3D point cloud of environment,so LiDAR plays an important role in autonomous driving.The problem is the point cloud data is sparse data and hard to apply image-based object detection.To solve this problem,we use voxel and sparse convolution-based deep convolutional network to perform 3D object detection in point cloud.Experiments shows good result of our method.The main research contents are as follows:(1)The deep-learning based object detection methods are analyzed,include single-stage method and two-stage method.The two-stage method is especially analyzed.A typical 3D object detection method for point cloud is analyzed.(2)The convolution and sparse convolution algorithm are analyzed.To solve the slow speed of current rule generation algorithm in sparse convolution,a parallel rule generation algorithm is introduced to increase calculate speed of sparse convolution in GPU.(3)Several improvements of voxel-based 3D object detection method are introduced.Firstly,sparse convolution is applied in LiDAR-based 3D object detection,thereby greatly increasing the speeds of training and inference.Secondly,a novel angle loss regression approach is introduced to solve adversarial example problem.Thirdly,a novel data augmentation method for LiDAR-only learning problems is proposed.In this method,we firstly get all point cloud of ground truths from train set,then randomly put them into train point cloud during training.(4)To solve problem in previous introduced problem,the sparse convolution-based backbone network is directly used in point cloud data.To solve the receptive field problem,the RPN network is improved by reducing it to single stage.Super converge method is used to greatly increase converge speed.The experiment results show that:(1)The proposed sparse convolution algorithm can increase network speed.(2)The novel angle loss demonstrates better orientation regression performance than other methods do.(3)The proposed data augmentation method for LiDAR-only learning problems that greatly increases the convergence speed and performance.(4)Pure sparse convolution network can improve network performance.(5)Super converge can increase converge speed while keeping good result.The proposed network produces state-of-the-art results on the KITTI 3D object detection benchmarks when submit while maintaining a fast inference speed.