- I suggest using Linux as your operating system.
- Clone this repo.
- Download the dataset from here.[1]
- Place the rgb and depth folders where the repo is cloned.
- Use the following command to install the requirements.
$ pip install -r requirements.txt
- Run the rgbd_localization.py
- By a frequency determined by
img_freq, 2 images are compared at each iteration. - SIFT [2] is used to find key points in the 2 images.
- Inliers are determined by finding Homography between matched pixel coordinates using the Least-Median robust method (LMEDS).
- Then, by using the depth images, 3D points of the matched inliers are generated.
- The rotation and translation matrices are found by the method suggested in [3] from the 3D coordinates.
- By knowing the first ground truth positions and the relative orientation between every 2 frames, the trajectory is estimated
These figures show the estimated and real trajectories in x, y, and z coordinates.
- Quantifying the error of the estimated trajectory.
- Improving the accuracy by methods such as sensor fusion or AI approaches.
[1] J. Sturm, N. Engelhard, F. Endres, W. Burgard and D. Cremers, "A benchmark for the evaluation of RGB-D SLAM systems," 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2012, pp. 573-580, doi: 10.1109/IROS.2012.6385773.
[2] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91–110, November 2004.
[3] K. S. Arun, T. S. Huang and S. D. Blostein, "Least-Squares Fitting of Two 3-D Point Sets," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. PAMI-9, no. 5, pp. 698-700, Sept. 1987, doi: 10.1109/TPAMI.1987.4767965.