Fast Feature Detection with a Graphics Processing Unit Implementation

We present a fast feature detector that uses the Graphics Processing Unit (GPU) to extract distinctive features from an image, which are suitable for reliable matching between different views of an object or a scene. This detector is a particularly useful tool for the computer vision community because it is: (i) simple to implement; (ii) extremely fast; (iii) operates stand alone requires no post-processing or filtering of detected features. We compare the fast feature detector to other similar published algorithms and demonstrate that it extracts features in desirable/salient locations, repeatedly and more rapidly than other scale-space detectors. We further describe implementations of two demanding real-time application scenarios: (i) real-time detection and tracking of multiple visual targets; (ii) real-time visual SLAM, with the results of each confirming the efficacy of the new features. In this latter application we make a comparison between full frame feature detection and Davison’s [3] active search technique and discuss the trade off between the two.