Extracting depth information from stereo images using a fast correlation matching algorithm

Stereo matching algorithms are essential for recovering depth information of objects in many computer vision applications including 3D reconstruction, robot navigation, autonomous driving and so on. Most of the stereo algorithms generally rely on two types of matching technique: global and local matching. The state-of-the-art stereo algorithms that measure disparity or depth with high accuracy are generally based on global methods. However, they are not suitable for real-time applications because of high computational costs. The local algorithms, on the other hand, are very fast but they provide less computational accuracy compared to the global methods. To make a tradeoff between computation speed and accuracy, this paper proposes an efficient local correlation approach for depth estimation using a pruning proposal. This paper also evaluates the performance of different matching cost functions/algorithms for disparity or dense estimation. Experimental evaluation confirms that our proposed pruning method for point correspondence is able to achieve a significant accuracy with high computational speed that can be very useful for real-time environments.