REAL-TIME CORNER DETECTION ON MOBILE PLATFORMS USING CUDA

Abstract

Corner detection is a widespread task in many high-end applications such as autonomous driving systems and augmented reality. However, corner detection algorithms are computationally expensive and thus are not suitable for real-time applications on mobile devices. The Tegra device series, which incorporates Graphics Processing Units (GPUs), is aimed at mobile computing and accelerates computations for mobile applications. In this paper, a GPU realization of a color-adapted Harris corner detector with high-level extensions is proposed and tested in Tegra-based platforms Jetson TK1, TX1 and TX2. The proposed realization achieves good detection quality, real-time processing for grayscale and color images up to full HD resolution (1080p) in all platforms and frame rates of 14.02 ~ 18.71, 27.76 ~ 39.44 and 34.82 ~ 51.32 for images of resolution 4K UHD (2160p) in Jetson TK1, TX1 and TX2 respectively.