Automatic Detection on Appearance Changes using Local Moment Invariants
Lee-Yeng Ong, Siong Hoe Lau, Voon Chet Koo · 2019
A reliable appearance template has to extract a set of invariant features that holds the blueprint of the target image, regardless of being affected by translation, scale, rotation, skew, reflection, contrast and blur. Invariant features are superior in describing a number of visually distinctive and outstanding characteristics of a target but inadequate to identify the appearance changes of the target at a particular surveillance time. Thus, this paper aims to present an automatic detection approach to discover the rate of appearance changes in the target by exploiting the locality property of moment invariants. Unlike the existing local feature descriptors which usually extract the salient features randomly from the target, the proposed approach examines the entire target and subsequently reveals a suspicious region that contains some features that are drifting away from the original template. When the variations of the target's size, shape and orientation are corrupting the features, the reliability of the appearance template will gradually deteriorate and affect the object tracking process. Experiments are conducted to show that a selection of orthogonal moments is applicable to identify the appearance changes of the target by using the generalization relation between geometric monomials and orthogonal polynomial functions. Other than locality property, the proposed approach also preserves the capabilities of distinctiveness and invariance.