Adaptive binocular robotic visual tracking algorithm based on mixture kernel Support Vector Machine

A novel control scheme for binocular robotic visual tracking problem is proposed. First we obtain the parallel configuration using rotation matrix and translational vector which can be both derived by corresponding points. Second, the binocular vision system provide the position in space and the references are used to estimate the transformation matrix to predict the image motion of the virtually stationary target. A nonlinear visual mapping model for the pseudo Jacobian Matrix is first proposed based on a mixture kernel Support Vector Machine implementation. We abandon the complex iteration calculating procure in Asada. M's approach, and the drawback of ill-condition of pesuodo-Jacobian matrix is overcomed due to its unsuitable initial value. Besides, it is not necessary to know exactly the camera intrinsic parameters; instead, 3 references are enough for the algorithm. The primary advantage of our approach is that only a few number of learning data are need for offline training, and it turn out that this new method not only makes the system get a satisfied fitting output, but also effectively restrains the fluctuation of the prediction output caused by local kernels. Extensive simulations and experiments demonstrate that effectiveness of the proposed control scheme.