Key Algorithms Of Visual Simultaneous Localization And Mapping For Mobile Robot

Autonomous mobile robot has played a very important role for social services, such asindustry, agriculture, daily life and military. Visual simultaneous localization and mapping（vSLAM）is a fundamental problem in robot visual navigation, which has been attracted alot of attentions and becomes to be one of the central topics in robotic researches. To addressthe challenge of visual loop closure detection in robot vSLAM in terms of effectiveness,computational efficiency and scalability, two important and difficult issues in visual loopclosure detection, known as image similarity measure and loop-closure candidates selection,are researched systematically for the dense environment map and the large-scale environmentmap; For improving the robustness of vSLAM system to motion blur, the motion blur law inimages observed by robot is analyzed, and a highly efficient and universal anomaly detectionmethod and its framework are proposed. The above mentioned important problems areresearched for improving the performance of vSLAM system. Specifically, the maincontributions are as follows.(1) Visual loop closure detection for the dense environment map. The Redundantinformation between neighbor keyframes in dense map is utilized to improve the accuracyand efficiency of visual loop closure detection due to the specific attributes of the denseenvironment map. The mutual information (MI) is used firstly for loop closure detection indense map. The proposed method based on the similarity of neighbor keyframes acceleratesconvergence efficiency of particles without depending on the extraction of high dimensionalfeature and the incremental visual dictionary. The complexity of the proposed method isscalable to environment map, and just related with the number of particles. Hence, thecomputational efficiency of loop closure detection is manageability. The experiment at anopen dataset verifies that the method is fast and accurate enough for visual mapping in robotvSLAM.(2) Highly efficient visual loop closure detection for the large-scale environment map.The relationship between the MI of image pairs in RGB color space and the scale of image isanalyzed, and do also for the MI of image pairs in binary space and the scale of image. Acompact binary image descriptor is designed by selecting a particular scale in the scale spaceof the image, a binary version of MI is proposed to measure the similarity of image pairs byonly employing5CPU instructions POPCNT. Loop closure candidate selection for20millionkey locations is implemented in1s without depending on any index structure, and it achievesnearly100%recall with a small top k (less than8). Hence, the computational complexity of loop closure verification is a small constant. In addition, the increased computational andstorage cost for maintaining the data structure for indexing the visual features are avoided.The excellent performance of the proposed method in terms of its low complexity andaccuracy in experiments establishes it as a promising solution to loop closure detection inlarge-scale robot maps.(3) Highly accurate and scalable method for visual loop closure detection in thelarge-scale environment map. The bag of raw features is designed to describe an image byextracting low dimensional and binary features. The location sensitive hash is employed tohash the image descriptors into some tables with bit sampling randomly for managing andmatching visual features.2-NN seeking for a query feature F is implemented to select the bestfeature pairs with distance ratio and a set of keyframes similar to the current observation byrobot is collected simultaneously. Jaccard coefficient is utilized to measure the similaritybetween the query image Q and the set {F} for recalling the top K best keyframes as loopclosure candidates {CL}. Although the hash method just offers a sub-linear complexity forloop closure candidate selection, the loop closure detection still achieves high efficiency dueto the real-time binary feature matching. The experiments on two datasets verify that theproposed method has excellent performance in terms of accuracy and efficiency. Especiallynote that the method achieves100%recall with a small size of set {CL}. In addition, it savesnear93.73%storage cost by contrasting the environment map that consists of keyframes, andis scalable to robot maps.(4) For addressing the problem about the robustness of robot vSLAM system due tomotion blur, a real-time and universal method and framework are proposed. The negativeimpact of motion blur on vSLAM is analyzed quantitatively and qualitatively, the motion blurlaw in images is analyzed on a humanoid robot, and then a no-reference method is proposedfor measuring the motion blur feature of images captured by robot, an unsupervised method isemployed to cluster the blur feature of images that occurred at time sequence in an detectionframework for recalling the anomaly from observations. The purpose is improving therobustness of vSLAM system to motion blur. Simulation and visual mapping experiment onhumanoid robot verify that the proposed method is real-time (0.1s per detecting) and effective(recall:98.5%, precision:90.7%) on an open standard dataset and the dataset acquired byNAO. The detection framework of the proposed method is universal, and it is feasible andconvenient to integrate the method into a vSLAM system.