Stereo matching architecture for 3D pose/gesture recognition and distance-measuring application

Abstract

Stereo matching technique has been extensively investigated for depth map extraction, while most depth-map applications (such as 3D gaming, vehicle collision detection, etc.) adopt active scan systems to measure the distance. It is challenging to design an efficient hardware architecture of stereo matching which meets the real-time/high-resolution requirement. Moreover, external factors such as lighting condition, variance of lens distortion, and stereo image misalignment may impact the accuracy of the depth measurement significantly. To address these issues, we propose a real-time stereo matching architecture which is optimized for the accuracy of pose/gesture recognition and vehicle collision detection. The proposed architecture features sub-pixel estimation and programmable features in lens distortion, misalignment, lighting factor, working range, and refinement parameters. FPGA implementation of the proposed architecture produces stable depth map stream, reaching 1920×1080 image resolution at 60fps.