Object tracking with a moving camera

Abstract

The authors describe the implementation of the local and focal levels of a dynamic-motion-analysis framework. Dynamic motion analysis achieves efficiency through sequential decomposition of a complex analysis task into simpler tasks, by 'peeling off complexity', and by directing analysis to portions of a scene that are most critical to the vision task. The authors describe four basic techniques for implementing dynamic analysis: foveation, two-stage motion computation, tracking, and one-component-at-a-time segmentation. Each process entails several iterations of a basic operation but convergence is fast and the computations themselves can be relatively crude. By way of illustration, the dynamic motion analysis technique was applied to a number of image sequences. Particular attention is given to an actual video sequence of a helicopter flying over a terrain. The sequence was obtained from a camera moving relative to the helicopter. It is concluded that the dynamic approach to motion analysis holds the promise of performing real-time processing to obtain precise, robust results, using practical hardware.<>