关键词: 3D点云, 点云物体识别, 深度学习, 点云柱, 类图像

Abstract: Lidar has been initially applied in autonomous driving and industrial automation, generating vast amounts of point cloud data for scenes and objects. These point cloud data are characterized by high dimensionality and irregularity, and require computationally expensive 3D convolution in existing deep learning models, leading to high spatio-temporal complexity and hindering online application. Addressing the limitations of traditional network models in processing point cloud data, this paper proposes a 3D point cloud object recognition method based on 2D convolutional neural networks. The proposed method statistically regularizes irregular point cloud data into pillars, utilizes convolutions and pooling to extract features from clusters of pillars, converts the 3D point cloud data into 2D image-like features, and employs 2D convolutional neural networks to extract multi-scale latent features from multiple receptive fields. The decoder network then identifies objects within point cloud based on locations, orientations, and object types. Experiments are conducted on Ascend Atlas 200DK edge devices, achieving a single inference time of 291?ms. Compared with traditional point cloud object detection networks, the proposed method outperforms VoxelNet, F-PointNet, and Second by 14.7, 13.2, and 3.4 times, respectively, in terms of performance gains. On the KITTI dataset, the average precision exceeds that of the second-best algorithm by more than 2.3%  compared with 14 other point cloud object detection algorithms, including ContFuse. Ablation studies focusing on 2D convolutions and attention mechanisms reveal improvements of 50.9% and 5.37%, respectively, in model size and inference accuracy. The experimental results demonstrate that the proposed method can efficiently, robustly, and accurately detect objects within point cloud data.

Key words: 3D point cloud, point cloud object recognition, deep learning, pillar, pseudo image