Real-time stereo matching using extended binary weighted aggregation

Abstract

This paper presents an accurate real-time stereo matching method, which is based on the extended binary weighted aggregation. The accuracy of the proposed stereo matching method was significantly enhanced by extending its binary weighted aggregation so that remote connections of support regions can be allowed for aggregation. The extended binary weighted aggregation is based on the following two new ideas. First, the extended binary weighted aggregation connects distant regions over color boundaries, making them one large support region for a given pixel. This approach induces more aggregation targets, and, thus, makes the aggregation step more robust. Second, it excludes cost outliers in the support region to prevent them from being propagated during the aggregation step, making a quality support region. With the extended binary weighted aggregation, the proposed stereo matching method obtains more accurate disparity maps than existing stereo matching methods using binary weighted aggregation methods, while maintaining the speed advantage of binary weighted aggregation. Experimental results illustrated that the proposed stereo matching method outperformed all existing real-time stereo matching methods in terms of accuracy, providing the average bad pixel rate of 5.12%, for the Middlebury stereo test images. The proposed stereo matching method was implemented on a CUDA platform with a high-end GPU. The implemented system operated at up to 300 fps for the stereo images with 320×240 pixel resolution and a disparity range of 32 pixels.