I used indirect method to complete a stereo visual odometer. The main ideas are as follows: 1. The front end uses optical flow to track fast feature points 2. The backend uses g2o to optimize the sliding window. It is similar to the traditional indirect method. 3. But using optical flow and stereo as the front end makes it more robust to fast motion and pure rotation,and get the real scale of motion.
1. Continuing our previous work, I used a tightly coupled approach to fuse wheel speeds, making it robust to weak texture regions and obvious lighting changes. 2. Added Dbow2 to loopclose detection . After loopclosed, use pose-graph to optimize the trajectory. 3. Implement map preservation and reuse.
1.calibrate stereo image 2.use sgbm for match method 3.real time 20fps on cpu