A single octave SIFT algorithm for image feature extraction in resource limited hardware systems

With the availability and rapid advancement of low-cost, low-power, and high-performance processors, machine vision is gaining popularity in various fields, including that of autonomous navigation systems. Applying feature extraction techniques on the captured images provides rich information about the surrounding environment that can be used to accurately determine the position, velocity, and orientation of a vehicle. To extract these features in such an application, we developed the Single Octave Scale Invariant Feature Transform (Single Octave SIFT). This solution drastically reduces the computational load and memory bandwidth requirements while providing an accurate image-based terrain referenced navigation system for micro- and small-sized Unmanned Aerial Vehicles (UAVs). The Gaussian filtering and Keypoint extraction stages are the most computationally intensive parts of the Single Octave SIFT. The main focus of this paper is the design of this modified SIFT algorithm and the basic building blocks needed to implement these two stages within a low-cost, low-power and small footprint Xilinx Spartan-6 LX150 FPGA. Simulation results show that the number of memory accesses is reduced by 99.7% for Full-HD (1920×1080) images¹. The operation cycles of the Gaussian filtering and Keypoint extraction stages are reduced by 90.2% and 95% respectively, compared with the single instruction multiple data (SIMD) architecture.