Exploiting image formation knowledge for motion analysis

This paper deals with scene sequences showing objects in motion. Gray values of an object surface element, traced through the sequence, will exhibit systematic variations due to changing lighting and viewing conditions. Much of today's understanding of the underlying image formation process is due to Horn, who showed in particular that given sufficient knowledge about imaging conditions, gray values may be used to infer object shape. Assuming certain idealized but unknown object surface properties and light source characteristics it is shown that in general rotation angle increments and projected object surface angles but no absolute angle values may be computed from the observed data. Given additional knowledge about unknown illumination parameters, however, object shape may be determined. Computational procedures are proposed including a solution for the special case of constant rotation rate where rotation angles may be inferred from viewing a single surface element at least four times. In general, more than one element must be traced, e.g., two elements in five images. Simulations have been performed which show that many observations should be combined to obtain useful results from noisy intensity values.