A Moving-object Segmentation Method Based on Instance Enhancement on LiDAR Data

Aiming at the problem of poor real-time performance and low point-wise segmentation accuracy in LiDAR-based moving-object segmentation for robots in a complex dynamic environment, a moving-object segmentation method based on instance enhancement on LiDAR data is proposed. Firstly, motion features are extracted quickly by calculating the residuals between the current frame and historical frames of the 2D BEV (bird's eye view) images, which are converted from the 3D LiDAR point clouds. Then, instance features are extracted by integrating the motion features with the spatial characteristics of the 3D point cloud. Using the instance information, consistent motion segmentation is achieved for point clouds belonging to the same instance, and thus improving the accuracy of LiDAR-based moving-object point-wise segmentation. Experiments are carried out on the open-source dataset SemanticKITTI, and the experimental results show that the proposed method achieves an IoU (intersection over union) score of 72.2%. The ablation experiments indicate that the IoU score is increased by 8.7% over the baseline model after enhancement with instance information. The real-time performance and the moving-object segmentation accuracy are superior or comparable to the current advanced methods, validating the effectiveness of utilizing instance information enhancement for multi-perception tasks.