A 1.5nJ/pixel super-resolution enhanced FAST corner detection processor for high accuracy AR

Abstract

Most vision applications such as object recognition and augmented reality require a high resolution image because their performance is heavily dependent on a local feature point like an edge and a corner. Unfortunately, the vulnerability of correct feature detection always exists in vision applications. Moreover, it is hard to increase image resolution because there is the trade-off between the image resolution and the system power consumption in a wearable device. To resolve this, we present an energy-efficient Features from Accelerated Segment Test (FAST) corner detection processor with a high-throughput super-resolution 4-core cluster for low-power and high accuracy AR applications. To perform high throughput super-resolution, the hardware is proposed with an adaptive multi-issue multiply-accumulate (AMMAC) unit and a shift register (SHR) based angle integrator. Finally, a proposed super-resolution enhanced FAST corner detection processor performs 13.51% detection accuracy enhanced FAST corner detection on up to a 16× super-resolution image with only 1.5nJ/pixel energy efficiency.