Characterization of SURF interest point distribution for visual processing in sensor networks

We study the statistical characteristics of SURF interest points and descriptors, with the aim of supporting the design of distributed processing across sensor nodes in a resource constrained visual sensor network. We consider a sensor network with a single camera node and four schemes of delegating processing tasks to the sensor nodes. We discuss the potential and the challenges of the different schemes in light of the results of the statistical analysis. Our results show that the distribution of the number of interest points per image exhibits a heavy tail. The interest point locations are almost uniformly distributed along the axes of the images, but their X and Y coordinates are slightly correlated. Most interest points are found in the lowest octave layers, and the number of interest points decreases exponentially with scale. Our analysis suggests that for a wireless broadcast channel delegating subareas of images to processing nodes would lead to a more even allocation than delegating by octave layers. For directional wireless channels the efficiency can be significantly improved by performing some of the feature extraction tasks at the camera node.