This paper presents a natural feature tracking system for object recognition in real-life environments. The system is based on a local keypoint descriptor method optimized and adapted to extract salient regions within the image. Each object in the gallery is characterized by keypoints and corresponding local descriptors. The method first identifies gallery object features in new images using nearest neighbor classification. It then estimates camera pose and augments the image with registered synthetic graphics. We describe the optimizations necessary to enable real-time performance on a mobile tablet. An experimental evaluation of the system in real environments demonstrates that the method is accurate and robust.