Neural Net Based Division of an Image Blob of People into Parts of Constituting Individuals

摘要

This paper presents an example-based learning approach to divide a foreground blob of people into its constituents on a surveillance video camera image. As people tend to walk and interact in groups with other people, occlusions frequently happen in camera images. They are detected in the same foreground image blob and dividing it into image parts of constituting individuals is a prerequisite for high-level vision processing like people tracking and activity understanding. The division is easy for a human observer but difficult in computer vision especially when the image resolution is low. We treat this task as a pattern classification problem by identifying partial outline shape patterns of a foreground blob, which can characterize the position where the blob can be well divided. When a probabilistic neural network was employed to identify the pattern, the network showed over 80% correct recognition rates in experiments.