SLAM入门讲义.ppt

* Partial Observability * Indoor Laser Features Clustering of laser returns. Extraction and labelling of different feature types. Indoor features tend to be quite stable due to flat terrain and slow vehicle motion. * Outdoor Features: Tree Trunks * Tree Features What if the tree trunk is not circular? * Generalised Landmark Models * Scan-SLAM * Mapping Auxiliary Information Often want to generate a map of spatially located information in addition to just localisation. Eg., Soil salinity, temperature, terrain traversability, … DenseSLAM (Nieto et al.) is a way to embed auxiliary information in a SLAM map. Data does not participate in SLAM (except, perhaps to assist data association). Data’s spatial location moves if local SLAM features move. It is assumed to shift and warp linearly with local SLAM adjustment. * Appearance-Based SLAM The previous SLAM solutions rely exclusively on proximity measures to perform data association. Can be difficult to determine when you have returned to an old region via a large loop The loop closure problem: Given my current pose estimate, Find candidate matches to older mapped places Confirm a match hypothesis Detecting and confirming revisitation has been addressed by matching appearance rather than proximity estimates FAB-Map Mike Bosse’s keypoints * 3-D SLAM More complicated motion model. More difficult feature extraction. Well suited to IMU and vision. Not suited to 2-D scanning laser – need to reconstruct 3-D image Use an downward-tilted laser to map ground surface. Use a pan-tilt head. * More SLAM Research Active SLAM: Link control with navigation. Multiple Vehicle SLAM. SLAM in Dynamic Environments. Many variations of SLAM theme: Tree-SLAM Thin junction-tree SLAM Trajectory-based SLAM; (many variations) Graph-SLAM Frame SLAM TORO iSAM Particle filter grid SLAM Topological SLAM SLAM is an active and ongoing research problem. * Conclusion SLAM is closer than ever to being solved. Computation Non-linearity Features and envi