Simplifying SURF Feature Descriptor to Achieve Real-Time Performance

The detection and matching of interest points serves as the base for many computer vision algorithms, such as visual odometry, structure from motion, tracking or simultaneous localization and mapping. The accuracy of matching is therefore of very high importance. This requirement is however often irreconciliable with the requirement of real-time performance, especially on resource constrained architectures. In this paper, we analyze a few possible simplifications to the recently developed SURF feature description and matching scheme, enabling it to shorten the processing time on virtually all computing platforms. The introduced simplifications do not introduce any significant matching performance penalty when compared with the full SURF implementation in the aforementioned applications.