Elevation-Based Stereo
This page describes our novel formulation of MRF stereo in terms of pixel elevation (relative to a dominant ground plane), rather than in terms of the usual disparity. Rather than imposing a smoothness prior on disparities, this formulation imposes smoothness on elevation. As a result, the fronto-parallel bias that disparity-based stereo methods impose, which is inappropriate when the dominant ground plane in the scene is viewed at a slanted angle (see Fig. 1 below), is eliminated. The result is a stereo algorithm that better distinguishes small bumps, curbs and other obstacles from the ground plane they are superimposed on, which is important when these features are small or far away from the camera (see Fig. 2). Our algorithm was motivated by the application of using stereo to detect terrain hazards for blind wheelchair users, who need to detect these hazards from some distance, and as soon as possible.
Figure 1. Typical scene in which dominant ground plane is viewed at a slant.
Elevation-based MRF stereo uses a prior that is more suitable for scenes that are mostly flat and tilted (the scenes we are concerned with) than the usual disparity prior. This is because disparity jumps are very common on a tilted plane (even when perfectly flat, since disparity is linear in pixel row/column coordinates), but elevation rarely changes except when an obstacle is present. In other words, the elevation prior is useful for our application because it assumes uniform elevation unless there is evidence to the contrary.
Figure 2. Top: original image, with region of interest outlined in red. Middle: reconstruction from disparity model. Bottom: reconstruction from elevation model. The obstacle is more clearly discriminable from the flat background in the elevation model.
Click here for information on our fast GPU implementation of the algorithm, including code.
Click here for presentation slides covering this work.