Research On Lidar Based 3D Object Detection Algorithm And Annotation Tool Design

3D object detection algorithm based on point cloud data in unmanned scene has been a research hotspot.With the development and maturity of deep neural network technology,the method of using neural network for 3D object detection begins to show Great advantage.In this paper,the point cloud data collected by the vehicle 64-line lidar is used as the research basis,The KITTI data set is used as the evaluation sample,we studied how to quickly and accurately detect the position,size and direction of obstacles in the environment,which will provide reliable information for tracking and path planning.The 3D object detection algorithm proposed in this paper involves several key technologies such as point cloud voxelization,neural network,sparse convolution and deformable convolution.The main research results of this paper are as follows:(1)We used euclidean clustering to realized 3D object detection based on point cloud,which get a good performance on the KITTI dataset.According to the characteristics of KITTI dataset,we designed a data interface and preprocessed the labeled object.We proposed a suitable data enhancement method for the neural network based object detection algorithm.(2)Based on the principle of lidar,we proposed a method of dividing space in cylindrical coordinate system and transforming point cloud into voxel.Then we studied how the dividing method influence voxel uniformity and location accuracy through comparative experiments.For the first time,deformable convolution is introduced into the point cloud target detection network,which enhances the adaptability of the network to vehicles with different directions and shapes.A new method of anchor generation in RPN is proposed,which can effectively prevent the mismatch between anchor and ground truth,and remove the loss of angle classification in loss function.Combined with the above improvement strategy,the model get better AP and AOS than SECOND.the improved method proposed in this paper can also be applied to other voxel-based threedimensional target detection algorithms.(3)In order to solve the problem of too few samples in KITTI dataset,an intelligent point cloud annotation tool based on object detection algorithm is designed.The algorithm can be used to generate the initial label,and then we can label it manually on the basis of the initial label.In the manual labeling process,the region growing algorithm is used to achieve intelligent frame selection.The PCA algorithm is used to automatically generate the minimum bounding box,which significantly improves the efficiency of point cloud labeling.