Research On Cluster-tree RSOM Based Loop-closure Detection For Visual SLAM

With the rapid development of artificial neural network and robot technology, the solution of Simultaneous Localization and Mapping (SLAM) for mobile robot has become one of the prevalent issues in the field of artificial intelligence. For a robot in an unknown environment, due to the lack of prior knowledge of the surroundings, it has to recognize whether current place is being revisited. This process is known as loop-closure detection. The main task of vision-based loop-closure detection is to retrieve the most similar image of the image currently obtained from historical database and judge whether these images come from a same scene.First off, the thesis overviews the current situation of research about mobile robot and the core issues of loop-closure detection for SLAM. Also, several typical local invariant features and two kinds of image modeling methods are analyzed, which provide theoretical foundation for image representation, image matching and image retrieval in loop-closure detection.Since the similarity among adjacent frames in image sequence obtained in real-time is high, it easily causes wrong loop-closure detection if images are input into database directly in real-time. To solve this problem, this thesis designed a grouping approach for image sequence to select input images. Meanwhile, in most loop-closure detection algorithm, the computation time increases sharply with the enlargement of database, which leads detection system can not operate in real-time. As RSOM (Recursive Self-organizing Map) tree has excellent performance in large-scale image retrieval, this thesis proposed a RSOM-based scene modeling algorithm to construct landmarks and a RSOM-based landmark retrieval algorithm to retrieve similar landmarks in database. When operating, the system would remove redundant landmarks and images in the database.In loop-closure detection, perceptual aillasingdas is one of the main reasons for incorrect detection. For the purpose of minimizing the effects of this problem, the thesis proposed a threshold weighting method which used similarity information and geometric constraints among each landmark. At the same time, by applying incremental learning, the system would learn new information of the visiting scene to improve detection accuracy.The experimental results conducted in six different environments shows that the proposed loop-closure detection algorithm could operate in real-time in both indoor and outdoor environments, and has excellent recall performance. Also, it has adpativity to dynamic objects.