Space Variant Feature Extraction for Omni-directional Images

In recent years, the use of omni-directional cameras has become increasingly more popular in vision systems and robotics. To date, most of the research relating to omnidirectional cameras has focussed on the design of the camera or the way in which to project the omni-directional image to a panoramic view rather than on how to process these images after capture. Typically images obtained from omni-directional cameras are transformed to sparse panoramic images that are interpolated to obtain a complete panoramic view prior to low level image processing. This interpolation presents a significant computational overhead with respect to real-time vision. We present an approach to real-time vision that projects an omni-directional image to a sparse panoramic image and directly processes this sparse image. Feature extraction operators previously designed by the authors are used in this approach but this paper highlights the reduction of the computational overheads of processing images arising from omni-directional cameras through efficient coding and storage, whilst retaining accuracy sufficient for application to real-time robot vision.