Research On Topological Localization Based On Local Visual Feature For Mobile Robots In The Unstructured Environment

This thesis is focused on the research on topological localization problem based onlocal visual feature for mobile robots in the unstructured environment. One key issue fortopological localization is to build the topological map, which is based on the definitionof the topological node. The definition depends on the task performed by the robot andthe working environment. When the robot is located in complex unstructuredenvironment, or when the robot employs topological map to perform the task ofrelatively precise navigation, the robot needs a lot of nodes to build topological map, sowe can define the key-frames as topological nodes. When the robot is located in theenvironment which contains some distinct places, or when the task requires manyhuman-computer interactions, we can define semantic places as topological nodes.According to these two definition methods, this thesis introduces visual vocabulary treetechnique, which is popular and has been applied successfully in pattern recognitioncommunity, to the robot topological self-localization, and proposes two topologicallocalization algorithms which using key-frames as topological nodes and semanticplaces as topological nodes respectively. The on-line self-localization experiments arecarried out to verify the effectiveness of the algorithms.Firstly, the topological localization algorithm which regards key-frames astopological nodes is researched in this thesis. The robot can extract key-frames astopological nodes autonomously by itself. In off-line phase of robot topological mapbuilding, this thesis proposes a key-frame selecting algorithm based on the visualvocabulary tree to select distinctive and typical key-frames. Then we extract the localvisual features from key-frames to build a visual vocabulary tree, and employ the visualvocabulary tree to obtain a feature vector which can sparsely represent the local visualfeatures extracted from every key-frame as the nodes of topological map. In the on-linelocalization phase, we also employ the visual vocabulary tree to sparsely represent thelocal visual features extracted from the query image obtained by the robot, and localizesthe robot to the nearest topological node through matching between the feature vectorsfrom key-frames and the query image.Secondly, the topological localization algorithm which regards semantic places astopological nodes is researched in this thesis. The method is more consistent with theway human cognize the environment, and it is more suitable for human-computerinteraction. This thesis employs the visual vocabulary tree to sparsely represent the localvisual features extracted from the images acquired at different places as the training set.And we combine the Support Vector Machines (SVMs) classifier learning algorithmbased on statistical learning to obtain the classifier for place recognition. In the on-linelocalization phase, we employ the visual vocabulary tree to sparsely represent the local visual features extracted from the query image obtained by the robot as the input of theclassifier to realize the place recognition and robot topological localization.The COLD database is used to perform the topological localization experiments totest the two topological localization algorithms mentioned above, and determine the bestalgorithm parameters and training conditions. The experimental results validate theeffectiveness of these two algorithms.