Trackerbot: A robotic surveillance system based on stereo-vision and artificial neural networks

Abstract

Master-slave camera systems are ideal for detailed surveillance of desired targets in wide scenes. In this paper, we propose a fully automatic camera system that is structurally unconstrained and independent from both intrinsic and extrinsic camera parameters for detecting activities in indoor or outdoor environments. Unlike traditional models in our system, two wide cameras are used to reach the minimum error in estimating the pan, tilt and zoom parameters (PTZ) which then would be only affected by the resolution of the wide cameras or the PTZ motor system. After an initial automatic calibration, a feed-forward artificial neural network (ANN) takes charge of controlling the PTZ unit according to the information extracted from the frames of the wide cameras.