In context
In 2024, autonomous systems such as self-driving vehicles and mobile robots increasingly relied on LiDAR for robust perception in dynamic environments. However, moving-object segmentation (MOS) from LiDAR point clouds remained challenging due to poor real-time performance and low point-wise accuracy, especially in complex scenes. This paper, published in the Chinese journal Robot, addressed these issues by introducing instance-level information to enhance segmentation.
What was reported
Researchers from the National University of Defense Technology proposed a moving-object segmentation method based on instance enhancement on LiDAR data. The method first converts 3D LiDAR point clouds into 2D bird's-eye-view (BEV) images and computes residuals between current and historical frames to quickly extract motion features. These motion features are then combined with 3D spatial features to extract instance features, which are used to perform consistent motion segmentation for points belonging to the same instance, improving point-wise accuracy.
The network architecture comprises three modules: motion feature encoding, instance detection, and upsampling fusion. The instance detection module uses 3D sparse convolution and an object detection network to output instance information for three classes: cars, cyclists, and pedestrians. The upsampling fusion module integrates instance features with spatiotemporal features at multiple scales via an instance pyramid, enabling multi-task outputs including moving-object segmentation, semantic segmentation, and object detection.
Experiments on the SemanticKITTI dataset showed that the method achieved an IoU score of 72.2%. Ablation studies indicated that instance enhancement improved IoU by 8.7% over the baseline model. The method demonstrated real-time performance and segmentation accuracy superior or comparable to current advanced methods.
Why it mattered
This work was significant because it was among the first to leverage instance information for LiDAR-based moving-object segmentation, addressing the common issue of incomplete segmentation where parts of a moving object are misclassified as static. By enabling multi-perception tasks in a single network, the method provided richer environmental information for path planning and motion state prediction in autonomous systems, contributing to safer and more reliable operation in dynamic environments.
"Using 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."
Source: 《机器人》期刊 (robot.sia.cn) · Published 2024-09-15 · “一种实例增强的激光雷达运动目标分割方法”