Author: Chang
Year: 2021
- VSLAM based on YOLOv4 TINY + dynamic feature elimination
- epipolar geometry + LK optical flow
- Only comment performance of the tracking thread ??
- Masl-RCNN not applicable to real time => YoloTinyv4
- DS SLAM = ORB SLAM2 with semantic information
Dynamic pixels filtering:
- dynamic object determined with semantic information
- filtering dynamic also with distance wrt the epipolar line
- and with optical flow (if the optical flow of a feature point is superior to the average optical by 2 sigma)
Experiments:
- improve orbSLAM 2 with their dynamic object removal
- many object detection failure in blurred images