August 2020
tl;dr: Semantic feature on the ground for mapping and localization in parking lots. Similar idea to Road SLAM.
Overall impression
This paper is on localization and mapping in parking lot, but the same principles apply to urban environment, which are also narrow, crowded and GPS-denied.
The semantic features are robust to perspective or illumination changes, long-term stable (as compared to traditional features such as ORB in ORB SLAM). This helps achieve cm-level accuracy required for AD.
This paper is extremely similar to a similar one from SJTU, AVP-SLAM-late-fusion. AVP SLAM requries synchronized image feeds, and AVP-SLAM-late-fusion explicitly handles unsynchronized images through late fusion of semantic point clouds.
The paper is very well written and easy to follow.
Key ideas
- Four fisheye images from AVP cameras are stitched into one BEV image.
- The input to segmentation map is a stitched IPM image.
- Mapping:
- Semantic segmentation results are lifted into 3D and aggregated into a local map. A local map for a 30 meter window is maintained.
- Loop closure to address long time drift of odometry sensors. Once a loop is detected, a global pose optimization is performed.
- Localization:
- When vehicles are visiting the same parking lot again, semantic features are lifted into 3D space and are compared with the map with ICP.
- Good initialization is needed for ICP (as ICP matched points?): we can use parking lot entrance or GPS for initialization.
- EKF fuses visual localization results with odometry which guarantees to have a smooth output. Odometry used for prediction, and visual localization is used for updating.
- Evaluation: localization error is more important than mapping error. Even an inaccurate map can be used to guide car into parking into the right parking spot as long as the vehicle can precisely localize in this map.
- Absolute mapping error meter level. RMSE 4 m in 300 meter length.
- Localization error is cm level. Max error 5 cm.
Technical details
- AVP: automatic valley parking
- 6 semantic classes are labeled: lanes, parking lines, guide signs, speed bumps, free space, obstacles and walls.
- Evaluation with RTK-GPS in open outdoor parking lot.
- The evaluation is done on datasets collected 1hr, 3hrs, 1day, 1week and 1month apart to demonstrate the robustness of the approach.
Notes