A learning-based approach for moving object segmentation in 3D LiDAR data, distinguishing moving vs. static objects in real-time.
LiDAR-MOS (LMNet) is a deep learning-based method for moving object segmentation in 3D LiDAR data. It processes sequential LiDAR scans to distinguish between moving and static objects, such as separating moving vehicles from parked ones. This capability is essential for improving the robustness of autonomous navigation systems, including odometry, SLAM, and mapping.
Researchers and engineers working on autonomous vehicles, robotics, and 3D perception who need real-time segmentation of dynamic objects from LiDAR point clouds. It is particularly relevant for those developing or enhancing LiDAR-based SLAM, mapping, and collision avoidance systems.
Developers choose LiDAR-MOS for its real-time performance, flexibility with existing segmentation networks, and proven ability to enhance downstream tasks like odometry and mapping. Its modular design allows easy integration without overhauling existing pipelines, and it comes with a dedicated benchmark for evaluation.
(LMNet) Moving Object Segmentation in 3D LiDAR Data: A Learning-based Approach Exploiting Sequential Data (RAL/IROS 2021)
Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.
Processes LiDAR data faster than the sensor frame rate, as shown in the demo video, enabling real-time applications like odometry and collision avoidance.
Works with any range-image-based LiDAR segmentation network such as SalsaNext or RangeNet++ by only changing the data loader and input, without modifying core pipelines.
Uses residual images from consecutive scans to capture motion cues, improving segmentation accuracy, as illustrated in the pipeline diagram.
Introduces a new benchmark based on SemanticKITTI for LiDAR-based moving object segmentation, providing a standard for evaluation and comparison.
Requires generating residual images and downloading large datasets like KITTI-Odometry, involving multiple steps and significant storage, as outlined in the how-to section.
Relies on third-party segmentation networks with their own setup issues and version dependencies, such as the specific PyTorch version for SalsaNext mentioned in the README.
Only supports range-image-based LiDAR processing, so it cannot be directly applied to point cloud-based methods without conversion, reducing its versatility.