Horizon Detection

The first step is to segment the image into sky and ground regions. We have found that thresholding directly the images received by the camera is not sufficient, due to the influence of the light source (the Sun). The solution is to estimate a quadratic relation between average brightness and column position by fixing an initial upper section of the image. The image is then ``equalized'': the inverse of the quadratic function is applied to all pixels, so that the whole image is referenced equally to the middle upper pixels. Sky equalization is done automatically for each incoming image; a threshold is then applied to the equalized image to segment the sky. The threshold is fixed at the beginning of a sequence.

The border between the sky and ground regions must be detected and analyzed. The first step is to scan the segmented image and mark all pixels that belong to the boundary between regions. The second step is to link all the edgels and to obtain line descriptions of the boundaries. Each image line is represented as a parametric line: $x\; =\; a\; +\; b\; t$ and $y\; =\; c\; +\; d\; t$. This representation is complete and linear; linear regression is adequate for line fitting. Lines are terminated using a $\chi$-square test, and merged when necessary using a test based on the Akaike Information Criterion [2].