Vision Sensor with an Active Digital Fovea

Foveated image sensors have a variable spatial resolution enabling a significant reduction in image size and data volume. In this work, the requirements for a foveated sensor within an active vision system are analysed. Based on these requirements, the constraints on the resolution mapping function are determined, and a range of Cartesian based mapping schemes investigated. The results demonstrate that separable mappings, which independently map two orthogonal dimensions are efficient to implement using an FPGA. The computational requirements for an L2 or Euclidean radial mapping are significantly higher, although this yields more natural looking low-resolution images. A compromise is based on using an L∞ radial mapping. The requirement to process data as it is streamed from the camera necessitates implementing the warping using a forward mapping rather than the more common reverse mapping. The implications of this for FPGA implementation are described.