The integral image method for fisheye images

Abstract

Because of strong radial distortion, fisheye images have to be rectified before use in most computer vision algorithms. We investigate three ways to rectify fisheye images: the rectangular local perspective projection, local perspective projection bounded by longitude and latitude lines, and cylindrical projection. The traditional integral image method doesn't work for local projections, because they produce patches with curved edges. We find the integral over any local projection patch by traversing the patch edges, rather than using just the corner values. This is slower than the conventional method, but still fast enough for real time applications. We can compute integrals for 5000 patches in 10 ms on a fisheye image of 640 × 480 pixels. Computing local projections themselves is slow (120 ms for 5000 patches), while the cylindrical projection needs only 15 ms for the whole image.