A real-time monocular SLAM system that creates large-scale semi-dense maps using a fully direct approach without feature extraction.
LSD-SLAM is a real-time monocular SLAM system that creates large-scale semi-dense maps from single camera input. It solves the simultaneous localization and mapping problem using a fully direct approach that processes image intensities directly without feature extraction. The system can run in real-time on standard laptops while building detailed 3D maps of environments.
Robotics researchers, computer vision engineers, and developers working on autonomous systems that require real-time 3D mapping and localization from monocular cameras. Particularly useful for those building drone navigation, robotic exploration, or AR/VR systems.
Developers choose LSD-SLAM for its fully direct approach that avoids feature extraction limitations, its real-time performance on consumer hardware, and its ability to create semi-dense maps rather than sparse point clouds. It provides robust monocular SLAM capabilities without requiring specialized sensors.
LSD-SLAM
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 image intensities directly without feature extraction, making it robust in low-texture environments where feature-based methods fail, as highlighted in the key features.
Runs in real-time on consumer laptops while building large-scale maps, enabling practical use without specialized hardware, as demonstrated in the quickstart.
Generates detailed semi-dense point clouds rather than sparse features, providing richer 3D scene representation for applications like robotics and AR.
Handles extensive environments through pose-graph optimization and loop closure detection, supporting long-term navigation and mapping tasks.
Tied to specific ROS versions (fuerte or indigo) and Ubuntu releases, with no catkin support and complex installation steps, complicating modern integration.
As a monocular system, it cannot estimate absolute scale and requires sufficient translational camera movement, limiting use in static or rotation-only scenarios.
Results vary between runs due to parallelism and keyframe timing, making it unreliable for applications needing consistent, reproducible outputs.