RT-SLAM
RT-SLAM stands for Real Time SLAM (Simultaneous Localization And Mapping).
Disclaimer: RT-SLAM is still in development and has no official release. It is provided as-is, there is no standalone installation.
Presentation
RT-SLAM is a fast Slam library and test framework based on EKF. Its main qualities are:
genericity: for sensor models, landmark types, landmark models, landmarks reparametrization, biases estimation;
speed: real time at 60 fps (VGA, gray level) on a decent machine (at least one Core2 core at 2.2GHz)
flexibility: different estimation/image processing sequencing strategies (active search), independent base brick for a hierarchical multimap and multirobots architecture;
robustness: near-optimal repartition of landmarks, data association errors detection (gating, ransac);
developer-friendly: visualization tools (2D and 3D), offline replay step by step, logs, simulation.
For now it provides:
Landmarks: Anchored Homogeneous Points (Inverse Depth) that can be reparametrized into Euclidean Points;
Sensors: Pinhole cameras;
Prediction: Constant velocity model, inertial sensor;
Data association: Active search, 1-point Ransac, and mixed strategies.
Documentation
Screenshots and videos
|
|
Demo 1 (v.a, 1'33, 23Mo, H264 AVI): hand held alone camera at 60 fps indoor, with several loop closures.
Demo 2 (v.b, 1'06, 20Mo, H264 AVI): hand held IMU+camera at 50 fps indoor, with very high dynamic.
Demo 3 (v.a, 1'25, 27Mo, H264 AVI): IMU+camera on a rover robot on grass, with short term memory (lost landmarks are removed, no loop closure)
Roadmap
- Stabilize and make a release
- Inertial Slam
- Odometry Slam
- Multimap Slam
- Segments Slam