Enhancing active vision by a neural movement predictor

We present an application of a neural network predictor for an active vision system. A short sequence of the objects behaviour is analyzed by the neural network to calculate an estimate of the forthcoming position. This result is fed into the pan-tilt-unit movement control, to steer the camera directly onto the prospective object position. By this means a predictive tracking system is realized, keeping the moving object of interest within the center of the visual field. Even a non-predictive tracking algorithm, always limping after the object, can be exploited to generate training data suitable for teaching the neural predictor. Implementing the neural movement predictor into the control loop enhanced the tracking capabilities of the active vision system substantially. The results, demonstrating the capabilities of the approach, are believed to be the basis for enabling a variety of further industrial applications with active vision systems.