LoRa-MOS: Density-Decay-Robust LiDAR Moving Object Segmentation via Logarithmic Spatial Representation and Range-Balanced Learning
Abstract
Moving object segmentation (MOS) is a fundamental component for safety in LiDAR-based autonomous driving. However, the performance of existing MOS frameworks drops sharply with distance, mainly due to the inverse-square decay of LiDAR point density and the near-field bias of standard training pipelines. Conventional bird's eye view (BEV) representations amplify this by allocating spatial resolution uniformly, leaving far-field regions severely under-represented. We propose LoRa-MOS, a far-field robust MOS framework that addresses these limitations through three complementary strategies. First, we introduce a logarithmic BEV projection that mitigates radial density decay by redistributing points to balance density across varying ranges for consistent feature extraction. Second, a range-balanced learning strategy prevents optimization from being dominated by dense near-field points. Third, we adopt a dual-branch semantic guidance scheme that couples a range-view movable predictor with a BEV moving predictor to exploit their complementary strengths. Extensive distance-stratified experiments show that LoRa-MOS achieves substantial far-field gains while maintaining competitive near-field IoU and real-time inference speed.