Current point cloud-based 3D object detection methods fail to fully use the local geometric features of the point clouds,leading to poor performance in detecting objects of sparse point clouds.To solve this problem,a two-stage 3D object detection algorithm named voxel graph attention region-CNN(VGT-RCNN)is proposed based on the voxel graph attention of raw point clouds.Initially,the grid center point features are calculated by multiscale voxel feature in-terpolation.Then,a local graph is constructed on the multiscale non-empty voxel features.Finally,a weighted average is conducted for the voxel features by graph attention mechanism,fully extracting and using the local geometric features of the object to complete detection.The algorithm mainly improves the defect of the present two-stage algorithm,which does not sufficiently consider the significance of different voxel features in feature clustering.In addition,a learnable weight matrix is introduced to dynamically learn the weight of the voxel feature and increase the expression ability of local features.The algorithm has been sufficiently tested on the popular KITTI autonomous driving dataset,obtaining competitive detection effects.The accuracy of cars with sparse point clouds has been markedly improved.A visualized analysis is also carried out to determine the detection effect.
关键词
点云/三维目标检测/图注意力/特征插值/多尺度特征/激光雷达/体素化/车辆检测
Key words
point cloud/3D object detection/graph attention/feature interpolation/multiscale features/LiDAR/voxeliz-ation/car detection
维普
万方数据