Abstract
This thesis proposes simpler rotation-invariant features to be implemented on hardware and to rectify the stereo camera with active head. The invariant features combined the detector of features from accelerated segment test (FAST) and the feature descriptor of pattern-based binary robust independent elementary features (BRIEF). Comparing with popular invariant features such as scale-invariant feature transform (SIFT), the simpler rotation-invariant features we proposed are computationally faster and still reliable. Therefore, it is more suitable for the application on stereo rectification. Besides, due to the feature descriptor of binary string, it could promote the efficiency on feature matching. By minimizing the cost function over the obtained sparse set of correspondences between left and right views, we could estimate the essential matrix and achieve the goal of stereo rectification. In the system of hardware, we design a pipeline flow and make the usage of hardware more efficient. A parallel architecture speeds up the parts of detector and descriptor. The time-consuming feature matching is also integrated into the system. It makes the system more complete. From the experimental result, its frame rate is up to 100 frames per second with the clock rate of 47.9 MHz and performance is comparable to SIFT. In conclusion, the proposed system is not only faster and simpler, but also still reliable for the application.