Self Occlusions and Graph Based Edge Measurement Schemes for Visual Tracking Applications

Abstract

The success of visual tracking systems is highly dependent upon the effectiveness of the measurement function used to evaluate the likelihood of a hypothesized object state. Generative tracking algorithms attempt to find the global and other local maxima of these measurement functions. As such, designing measurement functions which have a small number of local maxima is highly desirable. Edge based measurements are an integral component of most measurement functions. Graph based methods are commonly used for image segmentation, and more recently have been applied to visual tracking problems. When self occlusions are present, it is necessary to find the shortest path across a graph when the weights of some graph vertices are unknown. In this paper, treatments are given for handling object self occlusions in graph based edge measurement methods. Experiments are performed to test the effect that each of these treatments has on the accuracy and number of modes in the observational likelihood.