Segmenting Images of Occluded Humans Using a Probabilistic Neural Network

摘要

When processing an image of multiple occluded humans, segmenting them is a prerequisite for higher-level tasks such as tracking and activity analysis. Although a human observer can easily segment target humans partly occluded among themselves in an image, automatic segmentation in computer vision is difficult. In this paper, the use of a probabilistic neural network is proposed to learn various outline shape patterns of a foreground image blob of occluded humans, and then to segment the blob into its constituents. The segmentation is here regarded as a two-class pattern recognition problem; seg-mentable positions constitute a class and other positions constitute the other. The technique proposed is useful particularly for low-resolution images where existing image analysis techniques are difficult to be applied.