Combining photometric and depth data for lightweight and robust visual odometry

This paper presents a visual odometry system for mobile robots that works on RGB-D data from a Kinect/Xtion class sensor, and reports experimental results of evaluating this system on publicly available data. The aim of the presented research was to build a lightweight RGB-D visual odometry system, which can run in real-time on-board of such robots as walking machines that have limited computing resources. The proposed approach is based on tracking FAST keypoints over a sequence of RGB frames to establish correspondences between the photometric features in selected keyframes of the RGB-D data stream, and then on using the readily available depth data to map these features into 3D coordinates. The approach is tested on publicly available data, demonstrating satisfying performance with very low requirements as to the computing resources.