DatasetReaderKITTI is responsible for loading frames from KITTI Visual Odometry Dataset (optionally scaling them to reduce processing time) and ground truth (camera matrix, camera position and scale).

In a processing loop I convert images to greyscale, run keypoint detection using GFTT and then track these keypoints with FeatureTracker that uses OpenCV optical flow. After finding feature tracks, I remove outliers (outside of image). In the next step I estimate the essential matrix E to find relative pose R, t between consecutive frames. Calculating E matrix also allows to remove few outliers (found by RANSAC). Having that rotation and translation, I calculate absolute position and orientation of the camera and I use them to draw a trajectory.

To get better results, I rely on absolute scale provided as KITTI groundtruth when computing abs. pose.

There are plenty things to add: generating point cloud, measuring the accuracy of trajectory and so on.