Robust Real-Time Stereo Edge Matching by Confidence-Based Refinement

Edges are important features for tasks like object detection and vision-based navigation. In this paper, a novel real-time capable stereo edge refinement technique is presented. It propagates confidence and consistency along the detected edges, which reduces false matches significantly. Unmatched pixels are safely recovered by interpolation. We also investigate suitable support regions for edge-based matching. In the proposed solution, depth discontinuities are specifically accounted for. All approaches are extensively tested with the Middlebury benchmark datasets and compared to a sparse and several popular dense stereo algorithms.