Logo image
Back
Adapting 3D Point Cloud Moving Object Segmentation to Challenging Tasks
Conference proceeding   Peer reviewed

Adapting 3D Point Cloud Moving Object Segmentation to Challenging Tasks

Dylan Marcus Wright, Susanta Deka, Daniel Carvalho, Jarrod Brown, Christian Keyser, Bhuvaneswari Ramachandran and Achraf Cohen
Proceedings of IEEE Southeastcon, p.5
SoutheastCon 2026 (Huntsville, Alabama, USA, 02/20/2026–03/15/2026)
04/20/2026
DOI: https://doi.org/10.1109/SoutheastCon63549.2026.11476691

Metrics

1 Record Views

Abstract

Moving object segmentation in 3D sensing modalities remains a challenge when operating with single-view, long-range, narrow field-of-view (FOV) flash LiDAR data, which suffer from partial observability and atypical noise characteristics. In this work, we present an adaptation of the MotionFocused Model for Moving Object Segmentation framework tailored to flash-LiDAR returns: we project 3D point clouds into range images and compute inter-frame residual maps, then employ dual SalsaNext encoders to extract semantic and motion cues in parallel. Per-layer feature fusion, gated by a learned sigmoid mask and calibrated via softmax pooling, integrates static and dynamic information, while our Strip Average Pooling Layer preserves alignment across differing aspect ratios during down-sampling. A composite loss combining weighted cross-entropy and Lovász-Softmax terms supervises both semantic and motion branches. We validate our model on real-world scans collected from eight vehicle types at distances of 470 ft to 1,025 \text{ft} , achieving intersection-over-union scores above 90% on all classes and demonstrating robust segmentation under narrow FOV constraints. Runtime experiments on four NVIDIA A40 GPUs show practical training times, and qualitative analyses highlight the method's ability to delineate moving vehicles despite oversegmentation in contiguous regions. These results bridge a critical gap in moving object segmentation for single-view LiDAR systems and suggest promising directions for efficient, data-efficient 3D perception in autonomous platforms.

Details

Logo image