Mobile robot navigation in natural environments using robust object tracking

In this paper the authors introduce a method focusing on the robustness improvement of the landmark tracking system for mobile robot operation in natural environments. We extract feature points from the data obtained by a stereo vision system with CenSurE (Center Surround Extremas for Realtime Feature Detection and Matching) used as a detector, and FREAK (Fast Retina Keypoint) as a descriptor. RANSAC (RANdom SAmple Consensus) is used to remove outlier data from the feature points in order to increase precision. For self-localization, landmarks are selected from the surroundings. These landmarks are tracked by a template matching method using ZNCC (Zero-Mean Normalized Cross-Correlation) complemented with visual odometry based motion estimation. For performance purposes, this is combined with UKF (Unscented Kalman Filter) for narrowing the landmark search areas. A template update strategy suitable for long range tracking is also introduced. Finally, for increasing robustness in long range operation, we solve the issue of obscured/temporarily out of frame landmark tracking by estimating their position based on nearby visible landmarks.