Hybrid Environment Basedartificial And Natural Features For Monocular SLAM

Monocular Simultaneous Localization and Mapping (Monocular SLAM) has served as anexcellent research domain for mobile robots. In order to improve the overall performance of thesystem effectively in case of both artificial and natural features in an environment, this paperproposed a monocular SLAM algorithm based on hybrid artificial and natural features, whichcontained three modules: a visual feature modelling approach, a loop closure detection methodbased on visual dictionary, a map construction and optimization module. The main work andinnovations were as follows:First, a feature modelling approach based on visual dictionary was proposed. Every imageacquired contained a series of unordered natural features or artificial features. Visual features ofeach image were extracted by SURF, then a fuzzy K-means algorithm was applied to cluster thesevisual features into visual words based on which a visual dictionary was constructed. In order torepresent the similarities between each visual word and corresponding local visual features precisely,the Gaussian mixture model was employed to learn the probability model of every visual word.Secondly, aiming at the problem of loop closure detection in monocular SLAM, this paperpresented a visual dictionary based method. Based on the established hybrid probability model ofartificial and natural visual words, every image can be denoted as a probabilistic vector of visualdictionary so that the similarities between any two images can be computed by the vector innerproduct. Meanwhile, a Bayesian filter technology was applied to track the historical detection whichensured a success rate and continuity. A twofold memory management mechanism (the shallowmemory and the depth memory) improved the process speed of the proposed algorithm.Finally, our method was performed in hybrid environments with artificial and natural features. Itrealized through two steps: map construction and optimization. The camera relative poses wereestimated through geometric constraint within image sequences, and a graph was established basedon pose associations on different times. The constraints between nodes in the graph contained thelocal constraints of visual odometers and a global constraint of the closed-loop detections. Then, astochastic gradient descent approach was applied to minimize the errors introduced by constraintsduring map optimization in order to achieve a maximum likelihood map of the camera trajectory.