Research On Visual Attention Model And Application For Robot Tracking

As an important research content of the robot vision, the visual tracking technology obtains more and more attentions from the researchers. Until now, this domain has acquired more plentiful research achievements. The current visual tracking algorithms are generally based on the classics imagery processing method, whose insufficiencies lie in the poor timeliness and large computation load. Therefore, developing the imagery processing technology with quicker speed and broader adaptation scope has a quite vital significance regarding the robot object tracking. The scientific research indicates that the human has the exceptionally prominent capacity of the data selection and the visual attention guarantees the high efficiency work of human eyes. The attention mechanism provides some reference for distinguishing the target rapidly from the massive image data. Therefore, solving the imagery processing question in the robot object tracking system by simulating the human visual attention mechanism has important theory value and application prospect.The current research characteristics of the visual attention model:the very strong theoretic character and quite few application oriented; only processing the implicit expression attention; gaze shifts, not the gaze controls; most models are based on space. With the aid of the superiority in project domain, the visual attention models are applied to the robot domain in this research subject and the models are instructed through actual results. The characteristic description of the objects and the consciousness organization method are discussed emphatically. The explicit visual attention model for robot integrating the bottem-up, image-driven attention and the top-down, object-driven attention is studied in order to provide one kind of novel solution for overcoming the difficult problem of key technologies in the scene processing. At the same time, the intelligent mobile robot and the robot for simulating human eyes and neck are independently developed. The understanding ability to the non-structurized vision sensation information and the processing efficiency to the magnanimous isomerism information are enhanced by the proposed visual attention models.The visual attention model for mobile robot simulates the object tracking behaviors of human in the non-structurized environment. In view of the question of fast classification to the interesting regions as well as adaptive adjustment and tracking to the focuses of attention (FOAs) from the mobile robot’s vision; the bottem-up, image-driven component is based on the scale space theory, Mean-Shift algorithm, region growing and linear fusion of salient regions to realize the classification to the walking region and the object salient regions; the top-down, object-driven component is achieved by the similarity of the projection characteristics to control the dynamic transition processing of the FOAs in the salient regions. The contrast experiment shows that this model has the obvious superiority in the aspects of processing speed and the accuracy of FOAs.The visual attention model for the robot for simulating the human eyes and neck absorbs the merits of the saliency map and SIFT algorithm. The study and memory function of human cerebrum are simulated by the target characteristic database obtained by the off-line extraction and construction; the bottom-up, image-driven component is based on the improvement saliency map method, namely weighted characteristic gragh fusion method; the top-down, object-driven component is based on SIFT algorithm; the extraction and dynamic control of FOAs from the vision of robot are realized. The contrast experiment shows that the FOAs extracted by this model have good robustness and accuracy, meanwhile the processing speed is obviously enhanced.Taking the mobile robot and the robot for simulating the human eyes and neck as the experiment platforms, the mapping relation between visual space and real space in mobile robot system is established. The localization algorithms with error compensation in view of "human eyes" and "human neck" are proposed. The kinematics model aiming at the real-time change of relative position between the camera and the tracked objects is constructed. The prototype applications of the two visual attention models are realized on the robots. The experimental results validate the validity of the visual attention model in the robot object tracking systems.