An Energy-Efficient Visual Object Tracking Processor Exploiting Domain-Specific Features

Abstract

Nowadays, visual object tracking (VOT) has become a key technology in many AI applications such as intelligent surveillance and mobile robots. In the past, general AI accelerators have been developed and used for accelerating VOT. However, as the domain-specific knowledge is not well utilized, the energy efficiency of general AI accelerators is limited when accelerating VOT. To address this issue, in this work, an energy-efficient VOT processor is proposed by exploiting diverse domain-specific features of VOT processing. Several techniques have been proposed to improve the energy efficiency and processing speed, including an efficient sparse computing architecture with channel-wise data and weight compression, a Siamese-core processing technique and a bounding box (bbox) early-drop technique. Implemented and fabricated with a 55nm CMOS technology, the proposed VOT processor achieves high energy efficiency (15.08 TOPS/W), high frame rate (69.8 fps) with high accuracy, outperforming several state-of-the-art designs.