Enhancing scene structure in prosthetic vision using iso-disparity contour perturbance maps

We present a novel approach for enhancing structurally significant features in a scene to facilitate safe mobility with prosthetic vision. Previous approaches rely on visually salient features (e.g., intensity gradients, size, texture), or surface fitting (e.g., ground plane extraction), to determine and convey regions of structural change in the scene. Such approaches can be costly to compute, and/or are not guaranteed to detect all features relevant to the needs of safe mobility (e.g., small, low-contrast trip hazards). Assuming a dense disparity image, we propose a novel feature using iso-disparity contours. Regions of significant structural change are detected via a cost function based on local comparisons of iso-disparity contour orientations. Through this, structurally interesting features such as surface boundaries and general clutter are extracted and emphasised in the output visual representation. Our approach is real-time, and requires no surface fitting. Experimental results quantitatively and qualitatively validate our approach.