Novel Stereo Visual Odometry Pipeline for Autonomous Robot Navigation on Video Data

Autonomous robot navigation relies on accurate motion estimation. Stereo Visual Odometry is used to estimate the motion of a camera by analyzing the visual information captured by a stereo pair of cameras. Visual odometry refers to the process of estimating the camera's movement by tracking the visual features in consecutive frames of a video sequence. In a stereo camera system, two cameras are positioned at known baseline separation. The images captured by the cameras are used to create a 3D representation of a scene. Stereo visual odometry leverages the disparity between corresponding points in the pair of stereo images to estimate the motion of camera. We propose a stereo visual odometry pipeline and perform a comprehensive comparison of different trajectory computation approaches on the challenging KITTI dataset. Different feature detection methods, feature matching, filter matching distance, stereo matching, different types of motion estimation, optimization techniques such as loop closure and bundle adjustment were tested in terms of error metric and computational time. SGBM matcher performed better than the BM matcher on depth and disparity estimation. SIFT was found to outperform the ORB feature detector. Brute Force feature matcher outperformed FLANN matcher. The lowest MAE error rate of 32.54 was obtained with SGBM matcher, brute force approach, PnP RANSAC, and filter matching distance of 0.3 on SIFT features.