Learning Representative Features for Robot Topological Localization

Abstract

This paper proposes a new method for mobile robots to recognize places with the use of a single camera and natural landmarks. In the learning stage, the robot is manually guided along a path. Video sequences are captured with a front-facing camera. To reduce the perceptual alias of visual features, which are easily confused, we propose a modified visual feature descriptor which combines the dominant hue colour information with the local texture. A Location Features Vocabulary Model (LVFM) is established for each individual location using an unsupervised learning algorithm. During the course of travelling, the robot employs each detected interest point to vote for the most likely place. The spatial relationships between the locations, modelled by the Hidden Markov Model (HMM), are exploited to increase the robustness of location recognition in cases of dynamic change or visual similarity. The proposed descriptors are compared with several state-of-the-art descriptors including SIFT, colour SIFT, GLOH and SURF. Experiments show that both the LVFM based on the dominant Hue-SIFT feature and the spatial relationships between the locations contribute considerably to the high recognition rate.