A Study On Vision System In Front For Tele-presence And Tele-control System Based On Virtual Reality

In the tele-presence and tele-control system based on virtual reality (VR), the remote robot and the local control task by operators need interaction in real time. Because of an enormous amount of data and narrow band of channel, transmitting the whole image often makes a long time delay and difficult to synchronize the robot with the control task. According to the modern image coding and compressing technology, a transmitting approach is proposed in this paper. Only essential elements of visual objects in an image, i.e. objects' class, number, location, distance, pose and velocity etc., are transmitted instead of the whole image. On the basis of the above elements, the original object and the whole image can reconstruct with the object model database and VR techniques. The operator immersing in the virtual scene can give valid instructions to the robot or vehicle in distance plane to complete tele-operation. If the operator wants to survey the real scene and to correct the object in virtual image, he can order the remote to send out the compressed origin image. Then the time delay for transmitting object information in an image will be largely decreased and the tele-interation might be synchronous. So the visual system on the robot should complete corresponding visual function of producing the essential elements of an object. For this purpose, the remote robots must accomplish 3D object recognition, object tracking and feature point correspondence. How to obtain the essential elements of objects is the major task of the paper, i.e. one of requirements for the tele-presence and tele-control system based on VR.In order to make a deep study on machine vision for VR, the visual attention and fusion are explained in some detail and given the idea of a multiple model vision system. That is the fusion of the visual mechanism model, the visual function model and the visual behavior model. The multiple model vision can not only extract the essential elements of the objects form sequence images for VR remote control, but also give an exploration to machine vision research. Moreconcretely, the fusion of three models is supported by the polar-exponential grid technology. The visual mechanism is roughly analogized with polar-exponential grid sampling and polar-log coordinate transforming, the visual function is represented by visual invariants, and by the visual behavior we generally mean object recognition and tracking. The nice character of the polar-exponential grid technology is similar to human vision at some degree, which must be of interest to machine vision research.Centroid is a very important characteristic of the object, especially to the polar-exponential grid technology. The traditional algorithm of searching centroid has two weaknesses. An improved fast algorithm for searching object is proposed in this paper. The new algorithm can ingeniously find the starting dynamic centroid and adaptively select searching step. Searching time and region are reduced greatly. The practice data show the high efficiency of the algorithm.To accomplish 3D moving objects recognition, a method is used with the help of polar-exponential grid technology and aspect graphs. The scale and rotation invariants of objects are obtained in the polar-log coordinate FFT algorithm. These invariants are clustered by ART neural network, and finally the 3D objects are recognized with the model database of objects. The experiment verifies feasibility of the method.A new approach for feature point correspondence in arising a large displacement due to object moving is presented in this paper. The correspondence method has two procedures. First, the approximate location of feature points of moving targets can be obtained using log-polar coordinate mapping. The rough state parameters of a moving object, including centroid position, translation, rotation and scale, are gotten in log-polar and Cartesian coordinate, respectively. Then finely corresponding feature points can be realized with available traditional means, because the feature point displacements...