A fast, robust and low bit-rate representation for SIFT and SURF features

SIFT features have become extremely popular in computer vision because of their reliable matching qualities under changing lighting conditions. In Robotics they are ubiquitious in self localisation and mapping (SLAM), object tracking and recognition. However, the length of the descriptor is a major obstacle for real-time applications and mobile platforms where the computation time and storage capacity is limited. It has also been shown that the high-dimensional SIFT descriptors suffer from a numerical instability known as the curse of dimensionality. Therefore, we investigate low bit-rate representations based on a binarisation of SIFT and SURF features. Our method is able to reduce the descriptor length by a factor of 8-32 and lifts the curse of dimensionality. The new descriptor is tested in a series of robotics related experiments, namely stereo analysis and object tracking under outdoor conditions, as well as image recognition. Our experiments show that the new representation is not only extremely compact, but also yields faster, more accurate and more robust results.