An Architecture for Solving the Eigenvalue Problem on Embedded FPGAs

Burger, Alwyn; Urban, Patrick; Boubin, Jayson; Schiele, Gregor

Resource-limited embedded devices like Unmanned Aerial Vehicles (UAVs) often rely on offloading or simplified algorithms. Feature extraction such as Principle Component Analysis (PCA) can reduce transmission data without compromising accuracy, or even be used for applications like facial detection. This involves solving eigenvectors and values which is impractical on conventional embedded MCUs. We present a novel hardware architecture for embedded FPGAs that performs eigendecomposition using previously unused techniques like squared Givens rotations. That leads to a 3x performance improvement for 16 [Formula: see text] 16 covariance matrices over similar approaches that use much larger FPGAs. Offering higher than 30 fps at only 68.61 [Formula: see text]J per frame, our architecture creates exciting new possibilities for intelligent mobile devices.