An Integrated Framework for Feature Extraction, Object Recognition and Stereo Vision with GPU support

This paper investigates the integration of feature extraction, object recognition and 3D reconstruction by stereo vision into a unified framework. In doing so, stereo vision can be made more robust by applying feature extraction results to the stereo matching process, and object recognition can be extended through the integration of depth information as another feature of the scene. In this work a hierarchical feature extraction algorithm using Gabor filters is combined with a multi-path dynamic programming stereo algorithm. Subsequently, from this combination a new matching cost measured is proposed. The design is suitable for implementation on the graphics card (GPU), making it a target for real-time computer vision. This paper focuses on the framework’s application to stereo reconstruction and experiments show its ability to robustly match regions and preserve object depth boundaries through the use of feature analysis.