Intelligent service robots, namely the third generation of robots, are robots which serve for daily lives of the people and have abilities of interactive perception, independent judgment and autonomous navigation. Two characteristics of service robots are perception and judgment. Perception means that robots can recognize human's gestures or expressions and understand human's voice and intentions. Obviously the perceptual interface is an important way of humamrobot interaction. Judgment means robots can infer and decide their navigation actions independently and autonomously based on current environment information acquired by the perceptual interface. The thesis takes a human-centered perspective, focusing on gesture based human-robot interaction and autonomous navigation within a unstructured environment. Some common and essential subtopics, such as head pose estimation and tracking, hand gesture tracking and recognition, self-localization during navigation and path planning, are discussed in the thesis based on a robot prototype, namely NLPRwheelchair. Service robots usually work in a highly dynamic and unstructured environment whose uncertainties challenge the perception and judgment of robots. Therefore, probabilistic models are better to describe environment's uncertainties than deterministic models. The characteristic of this thesis is to treat problems of object detection, tracking and recognition as problems of state estimation within probabilistic frameworks. In this way, the intelligent wheelchair can perform vision-based human-robot interaction and navigation even in the unstructured environment like lobby.The novelty of this thesis comes from four points: ① Following Asimov's rules, four criterions of intelligent wheelchair design are proposed. Based on those criterions, an intelligent wheelchair is developed with multi-modal information channel, parallel tasks implementation software architecture and multiple control modes. ②Problems of head pose estimation, localization and tracking are unified as a Maximum a Posterior (MAP) problem of head pose state which is composed with position sub-state and pose sub-state. Within the probabilistic framework, head pose estimation is proposed as the system observation model (likelihood). A bowl-like band pass filter is designed to get discriminating appearance features of head poses under variable illuminations. The likelihood of head poses is defined by statistical modelling of pose clusters by Probabilistic Principal Component Analysis (PPCA). Particle filter is used to estimate the MAP value of pose state and multiple cues fusion is realized by partitioned particle sampling. ③ The Mean Shift embedded Particle Filter (MSEPF) is proposed for hand tracking in a cluttered environment. Compared with the conventional particle filter, MSEPF adds mean shift step before particle wei
修改评论