3D-Voxel Based Object Detection And Segmentation In Point Clouds For Autonomous Driving

In this paper,we address 3D object detection and instance segmentation from LiDAR point clouds for autonomous driving application.Autonomous driving needs to be aware of other road participants' location and motion for friendly interaction.Real?time object detection and localization is critical job to environment perception for autonomous driving.Multi-sensor data is used in this work,including camera,LiDAR and Radar.Although many methods have been proposed to detect objects using images.These outstanding 2D detection results lacks the 3D pose,which is key to tracking and prediction of moving objects.Since LiDAR point clouds enjoy rich geometry feature and high precision of location.We use LiDAR data as single input to do detection,recognition and segmentation for road moving objects in this paper.We propose a two-stage method for object detection and instance segment in point clouds.In stage-I,a 3D-Voxel and Region Proposal Network based 3D object detection network is extended by adding a front-ground and background segment branch in parallel with the existing branch for bounding box recognition and regression.Three branches share the feature extraction network.Replaced of multi-layer perception,graph convolution network is adapted to local geometry feature extraction.The new segment branch improves the bounding box regression performance and make the instance segmentation become easy task in the following step.Stage-II dose instance segment by clustering the point cloud inside 3D bounding box generated in stage-I.For object detection in point clouds,customized augmentation and evaluation are also proposed in this paper.Extensive experiments on the 3D detection benchmark of KITTI dataset show that our proposed architecture outperforms 3D-Voxel baseline methods.Besides,segmentation results as a kind of new output from our architecture,is also evaluated by metrics defined in this paper.