Research of CUDA in intelligent visual surveillance algorithms

Abstract

When used in practical applications, the speed of intelligent visual surveillance algorithms may decline dramatically due to massive data. Thus the computing speed of algorithms can be a crucial factor in the practical applications. In addition to excellent parallel computing capability, a modern GPU also has large bandwidth and powerful floating-point computing capability. These features make GPU an appropriate device for doing general-purpose computing. This paper accelerates Gaussian Mixture Model and HLSIFT (Harris-like Scale Invariant Feature Detector) using CUDA. The former algorithm gets more than 45 times accelerating and the latter one gets more than 35 times accelerating. The acceleration result is impressive.