Dividing an image blob of two connected people using shape information

In many automated visual surveillance applications, humans are important targets. As humans often move together, occlusions between them occur frequently, and it brings difficulty into image analysis. In this paper, two novel techniques are proposed to divide an image blob where two people are connected due to partial occlusion between them. The first technique uses a simple human size model to distinguish the occluder and the occluded. In the second, a probabilistic neural network is employed to learn the pattern of good dividing position along the top pixels of a blob. Since both techniques are shape-based, they do not need temporal information, and can be implemented in real time. The two techniques proved their usefulness when they were tested in experiments with various occlusion cases.