Instance Segmentation of LiDAR Point Clouds with Local Perception and Channel Similarity

LiDAR and its point cloud data are crucial visual sensors in smart driving cars for sensing the surrounding environment and achieving high accuracy localization. Compared to semantic segmentation, point cloud instance segmentation is a more complicated task. For autonomous driving systems, precise instance segmentation results can offer a more detailed understanding of the scene. The point cloud data is sparse and irregular, and the point cloud density varies with the distance from the sensor. In this paper, a LiDAR channel-aware point segmentation network (LCPSNet) is proposed to address the above problems. Given the distance-dependent sparsity and drastic scale variation of LiDAR, we adopt a top-down FPN. High-level features are progressively upsampled and summed element-wise with the corresponding shallow layers. Beyond this standard fusion, the fused features at 1/16, 1/8, and 1/4 are resampled to a common BEV/polar grid. These aligned features are then fed to the LPM to perform cross-scale, position-dependent weighting and modulation at the same spatial locations. The Local Perception Module (LPM) uses global spatial information to guide, while preserving, attention to group (scale) differences. Position-by-position weighting and re-fusion of local features of each group on the same grid. Enhances both intra-object and tele-context while suppressing cross-instance interference. The Inter-Channel Correlation Module (ICCM) uses a ball query to define local feature regions, modeling the spatio-temporal and channel information of the LiDAR point cloud jointly within the same module. And the inter-channel similarity matrix is computed to remove redundant features and highlight valid features. Experiments on SemanticKITTI and Waymo datasets verify that the two modules effectively improve the local feature and global semantic consistency modeling capabilities, respectively. The PQ of LCPSNet on the SemanticKITTI dataset is 70.9, and the mIoU is 77.1, and the instance segmentation performance exceeds the existing mainstream methods and reaches SOTA.