Topological localisation based on monocular vision and unsupervised learning

摘要

Global localisation is a very fundamental and challenging problem in robotics. This paper presents a new method for mobile robots to recognise scenes with the use of a single camera and natural landmarks. In a learning step, the robot is manually guided on a path. A video sequence is acquired with a font-looking camera. To reduce the perceptual alias of features easily confused, we propose a modified visual feature descriptor which combines colour information and local structure. A location features vocabulary model is built for each individual location by an unsupervised learning algorithm. In the course of travelling, the robot uses each detected interest point to vote for the most likely location. In the case of perceptual aliasing caused by dynamic change or visual similarity, a Bayesian filter is used to increase the robustness of location recognition. Experiments are conducted to prove that application of the proposed feature can largely reduce wrong matches and performance of proposed method is reliable.