Study On The Key Issues Of Cognitive Mapping For Robots

World maps show the surface of the Earth, including the localities and their connections.Thatishowtheworldisknowntoourhumanbeings. Similarly,ifrobotsweretounderstandtheirworking surroundings, map of the environment is a necessary. To enable robots to understandtheir surroundings autonomously, the problem of unknown environment mapping or more for-mally Simultaneous Localization and Mapping (SLAM) is proposed. After nearly three decadesof development, SLAM has achieved significant progress; some of the technologies have evenmatured. However, robots are still unable to provide the very basic services in daily life up todate. An important reason for this is that, traditional maps are navigation-oriented and lack ofcognitive information. Robots cannot understand the environment as human beings do withoutnecessary cognitive information, and the unavailability of it also hampers the applications oftraditional maps.In order to improve the cognitive ability of robots so that they can be cognizant of theirsurroundings, the thesis focus on two key issues related to the cognition map building prob-lem, which are the object perception and the scene understanding. Several related problems areexplored, including segmentation of unknown objects from indoor scenes, learning object mod-els from scenes of daily living and scene recognition through observed images etc. The maincontributions are as follows.1) A frame registration method with local consistency constraints is proposed for3D modelreconstruction. Inconsistency is a key factor that affects the precision of registration. Afterthe analysis of the possibility of two kinds of inconsistencies, we exploit the relationshipof the correspondences among multiple frames, and introduce an inference process so asto maintain the correspondence consistency. Besides, the relationships of the transforma-tions are invoked as constraint conditions, and correspondence error is minimized to obtainthe transformations. Due to the special form of the object function, a method combiningalternating direction and nonlinear least squares is employed to solve the optimization prob-lem. The experiments demonstrate that the proposed method can improve the precision ofregistration.2) Local surface convexity is an important clue which helps to distinguish between the rims of objects and the edges between objects and backgrounds, and has been applied to seg-ment objects from daily scenes. However, the Boolean-value representation is sensitive tomeasurement noises, and incorrect judgements in convexity lead to bad results in segmen-tation. To address the issue, a method for measuring the degree of local surface convexity isproposed, and is applied to compute segmentation weights for adjacent local surfaces. Theweights are computed by combining the degree of local surface convexity and the local sur-face normal, which provide a way for better characterizing the properties of object segmen-tation. For scenes to be segmented, undirected weighted graphs are created through severalsteps, including noise filtering, normal estimation, graph construction and edges weighted.Unknown objects in the scenes are finally obtained with fast graph segmentation algorith-m. Experiments show that, compared to traditional methods based on local convexity andsurface normal, the proposed method is more robust to measurement noise and estimationerror; it also achieves comparable results with other techniques based on complex learningand inference.3) To achieve concurrent multiple object modeling, a framework named Simultaneous Recog-nition and Modeling (SRAM) is proposed. SRAM builds up object models in an on-linefashion while at the same time using the uncompleted models for recognition. To addressthe update issue during on-line object learning, a view graph model is proposed for ob-ject representation, where object views are nodes in the graph and constraints between theviews are the connections. Therefore, object models can be updated by modified the con-nections easily. To describe the relationships between measurements and object models,a probabilistic observation model is presented, where both the appearance and the spatialstructure of the object are examined. The difference of the appearances is characterized bythe distance of the corresponding descriptors, while that of the structures is measured by there-projection error. Based on the probabilistic model, recognition and modeling are formu-lated as a unified inference problem, and maximum likelihood estimation is employed toachieve joint optimal object recognition and modeling. When applying SRAM to the con-current multi-object model learning problem, the results show that the proposed method iscapable of learning object models from various kinds of scenes effectively, under both static and dynamic environments.4) Salient features extracted from images are important cues for scene recognition, howev-er, raw feature representation suffers from several drawbacks, such as high computationalcomplexity and sensitive to noise, which affect the recognition results. To solve these is-sues, feature coding techniques are employed. Different coding schemes are analyzed so asto find out the key rules for feature coding, and Laplacian Regularized Locality-constrainedCoding (LapLLC) is proposed. LapLLC utilizes a linear combination of nearby bases in thedictionary to represent the feature to be encoded, and uses a Laplacian matrix to impose con-straints among codes of similar features. This not only allows low reconstruction errors andgood maintenance of locality, but also well preserves the code consistency of similar fea-tures. Additionally, by introducing a set of template features, the coding procedure can bedecoupled, and each feature is encoded by solving a linear system analytically. Experimen-tal results show that LapLLC is not only with high efficiency, but also achieves outstandingperformance in scene recognition.