Research On Active Loop Closing Based On Muti-sensor Fusion In Indoor Environment

Simultaneous Localization and Mapping is the critical technology for robot to achieveautonomous navigation. However, the capability of localizing and mapping in SLAM isconstrained due to the environmental conditions in indoor environment. It is urgentlyneeded to find a loop closing mechanism to detect whether the robot is revisiting a region.With this mechanism, the compact on localization due to cumulative error of robot posescan be deduced. To improve the efficiency and accuracy of loop closing in indoorenvironment, in this paper, a research on active loop closing is conducted based on laserand image data.Firstly, considering the particularity of space structure in indoor environment, a fastline feature extraction algorithm is designed based on Turning Angular Function, and afterthat the space scenes are classified through these line features. Under theline-feature-representation of scene structure, a geometrical scene representation isconstructed based on features such as scene entropy, area of laser scan and closed area oflaser scan. With the proposed similar frames determining method based on LongestCommon Subsequence and Hu invariant moment, a scan matching model is constructed.Secondly, considering the particularity of visual scene in indoor environment, thefeatures of visual appearance are extracted by combining the Maximally Stable ExtremalRegions detector and Scale-Invariant Feature Transform describer. With the PCT method,a kind of robust dimension-reduced feature is made by fusing geometric and visualfeatures. In addition, by introducing the BoW model, inverse indexing mechanism andcreating the ChowLiu tree of scene features, this paper constructs a scene similaritymeasure model with low “perceptual aliasing” rate and high matching rate.Finally, according to the proposed classification method, the BoW framescorresponding to laser-image frames are classified in real time. By calculating scenesimilarity in each single class, the matching quantity is reduced, and by matching thegeometric scenes corresponding to scenes in “candidate loop sequence” based on the scenescan matching measure model proposed above, the mismatching rate is reduced.