Fast depth estimation using semi-global matching and adaptive stripe-based optimization

Depth estimation is one of the essential components in many computer vision applications such as 3D scene reconstruction and stereo-based object detection. In such applications, the overall quality of the system highly depends on the quality of depth maps. Recently, several depth estimation methods based on semi-global matching (SGM) have been proposed because SGM provides a good trade-off between runtime and accuracy in depth estimation. In addition, non-parametric matching costs have drawn a lot of attention because they tolerate all radiometric distortions. In this paper, a depth estimation method based on the Census transform with adaptive window patterns and stripe-based optimization has been proposed. For finding the most accurate depth value, adaptive length optimization paths via multiple stripes are used. A modified cross-based cost aggregation technique is proposed which adaptively creates the shape of the cross for each pixel distinctly. In addition, a depth refinement algorithm is proposed which fills the holes of the estimated depth map using the surrounding background depth pixels and sharpens the object boundaries by applying a trilateral filter to the generated depth map. The trilateral filter uses the curvature of pixels as well as texture and depth information to sharpen the edges. Simulation results indicate that the proposed methods enhance the quality of depth maps while reducing the computational complexity compared to the existing SGM-based depth estimation methods.