Feature Extraction Method Based on Content-Aware Tiling on Mobile Devices

The computing power of mobile devices makes it possible to transplant the applications which run on PCs to mobile devices. However, the power is still too weak to meet the requirements for some kinds of applications which are highly in real-time. In augmented reality applications, feature extraction, as well as feature matching is critical. In this paper .we focus on the part of feature extraction, and modify SURF as an example to fit the hardware characteristics of mobile devices. To solve the mismatch between the small cache size and the data access pattern. We proposed the content-aware tiling based SURF according to the small cache capacity. Experiments show that the accelerated SURF achieves a 1.5x-1.7x speedup without sacrificing recognition accuracy.