Research On Loop Closure Detection For Monocular SLAM

Simultaneous Localization And Mapping (SLAM) is a core problem in mobile robotics. A really autonomous robot has to be able to localize itself in an unknown environment using its onboard sensors. SLAM techniques deal precisely with this problem. Since cameras have become a common sensor in robotics applications, more people are turning towards vision based methods to achieve it and several successful solutions have now been demonstrated by several groups around the world.Because of the presence of substantial errors in vehicle pose estimates, it is hard to correctly asserting that a vehicle has returned to a previously visited location. This problem is called loop closure detection, which is an important component in the search to make SLAM the reliable technology it should be. Loop closure detection is still challenging in large scale unstructured environments, despite significant developments in the SLAM problem. In this paper, we investigate this problem, particularly in large scale urban environment. Our goal is to design a visual system able to perform loop closure detection, within the framework of an online image retrieval task.The main work is as follows:First of all, we investigate common image features in visual system of mobile robotics, with particular focus on the extraction of local features. We give a detailed analysis of the Speeded Up Robust Features (SURF) as an example. Then we look into which of the dominating image feature algorithms is most suitable for the task. The experiments show that SURF feature performs well.Secondly, we analyze the EKF-based SLAM algorithm and the Particle Filter-based FastSLAM, followed by a comparison of the similarities and differences of these two methods. Then we describe the underlying monocular SLAM used during the experiment, which is suitable for large scale environment. Finally, we design an online loop closure detection algorithm within the framework of image retrieval. The images are captured as the robot moves along the trajectory. Then these images are used to construct the adaptive vocabulary tree. We compare the retrieval performance of the adaptive vocabulary tree with the standard vocabulary tree. Different methods are used to prune out the bad matches. We implement the algorithm in Linux operating system. Experiments show the good performance of the algorithm.