Hierarchical feature scheme for object recognition in visual sensor networks

Visual sensor networks are the meeting point of two significantly different technologies: one is image processing with high computation and storage demands, and the other is distributed sensor approach with low power, low computational and storage capabilities. We propose a framework for hierarchical feature encoding scheme for a frequent computer vision task - object recognition. The key of our approach is the principle that individual nodes in the network hold only a small amount of information about objects seen by the network. However, this information is sufficient to efficiently route network query, when a new, unknown object is encountered. A set of criteria has to be fulfilled by the object recognition method to qualify for use in our framework. The paper provides examples of three widely known object recognition approaches that can be easily adapted for use in such hierarchical feature encoding scheme.