Binary Histogrammed Intensity Patches for Efficient and Robust Matching

This paper describes a method for feature-based matching which offers very fast runtime performance due to the simple quantised patches used for matching and a tree-based lookup scheme which prevents the need for exhaustively comparing each query patch against the entire feature database. The method enables seven independently moving targets in a test sequence to be localised in an average total processing time of 6.03 ms per frame.A training phase is employed to identify the most repeatable features from a particular range of viewpoints and to learn a model for the patches corresponding to each feature. Feature models consist of independent histograms of quantised intensity for each pixel in the patch, which we refer to as Histogrammed Intensity Patches (HIPs). The histogram values are thresholded and the feature model is stored in a compact binary representation which requires under 60 bytes of memory per feature and permits the rapid computation of a matching score using bitwise operations.The method achieves better matching robustness than the state-of-the-art fast localisation schemes introduced by Wagner et al. (IEEE International Symposium on Mixed and Augmented Reality, 2008). Additionally both the runtime memory usage and computation time are reduced by a factor of more than four.