A neural network based vision system for 3D motion estimations

An active vision approach is proposed for 3D object motion estimation. The motion estimation problem is formulated as the problem of planning the poses of a moving vision system so as to minimize the estimation uncertainties. A Kalman filter is employed to estimate the object motion parameters. The Riccati equation of the filter is developed as a function of the vision system control parameters, namely, position, orientation, velocity, and acceleration. This allows for the estimation uncertainties to be treated as an evolutionary process which is controlled by the vision system parameters. An objective function is formulated based on the solution of the Riccati equation to map the sensor parameters into an index of uncertainty performance. A genetic algorithm is used to search for the optimum parameters which minimize the objective function. To achieve real-time motion estimation performance an artificial neural network is proposed to relax the computational demands associated with solving the Riccati equation. Experiments to demonstrate the speed and accuracy of object motion estimation achieved by a vision system using this control scheme is discussed