Dynamic vision via deterministic learning

The recovery of three dimensional structure and motion from time vary images with the aid of CCD camera(s) is usually performed using a nonlinear dynamic system, often referred to as a perspective dynamic system, where the major task is formulated as the problem of state estimation and parameter estimation. A Luenberger-type observer can be used to measure the constant motion parameter system states when only the output of the perspective system is measurable. In this paper, based on the recent results on deterministic learning theory, when the system states are periodic or recurrent, RBF neural networks can satisfy the partial PE condition along the states, the system dynamics will be learned by RBF neural networks and saved in a way of constant RBF neural networks, and the learning error converges exponentially to a small neighborhood of zero. Take the constant RBF neural networks achieved as training pattern to form the bank of system dynamical patterns, and before that a similarity definition is given. When meeting new system dynamics which are considered as test patterns, it can be used to achieve rapid recognition of system dynamical patterns between the test and training dynamical patterns.