Abstract

An active visual target tracking system is an automatic feedback control system that can track a moving target by controlling the movement of a camera or sensor array. This kind of system is often used in applications such as automatic surveillance and human-computer interaction. The design of an effective target tracking system is challenging because the system should be able to precisely detect the fine movements of a target while still being able to detect a large range of target velocities. Achieving this in a computationally efficient manner is difficult with a conventional system architecture. This thesis presents an architecture for an active visual target tracking system based on the idea of space-variant motion detection. In general, space-variant imaging involves the use of a non-uniform distribution of sensing elements across a sensor array, similar to how the photoreceptors in the human eye are not evenly distributed. In the proposed architecture, space-variant imaging is used to design an array of elementary motion detectors (EMDs). The EMDs are tuned in such a way as to make it possible to detect motion both precisely and over a wide range of velocities in a computationally efficient manner. The increased ranges are achieved without additional computational costs beyond the basic mechanism of motion detection. The technique is general in that it can be used with different motion detection mechanisms and the overall space-variant structure can be varied to suit a particular application. The design of a tracking system based on a space-variant motion detection array is a difficult task. This thesis presents a method of analysis and design for such a tracking system. The method of analysis consists of superimposing a phase-plane plot of the continuous-time dynamics of the tracking system onto a map of the detection capabilities of the array of EMDs. With the help of this 'sensory-motor' plot, a simple optimization algorithm is used to design a tracking system to meet particular objectives for settling time, steady-state error and overshoot. Several simulations demonstrate the effectiveness of the method. A complete active vision system is implemented and a set of target tracking experiments are performed. Experimental results support the effectiveness of the approach