The deployments of current object trackers on mobile or embedded devices are hindered due to high computation complexity and high resource footprint. To address these issues, this paper proposes an ultra-fast and energy-efficient object tracker, to perform accurate and robust object tracking at a speed of more than 56fps with a memory cost of less than 40KB and a clock rate of 480MHz via a unified feature representation and matching. First, we propose a compact feature representation for describing the object, which utilizes the response distribution to provide sparse representation. In this manner, the object can be represented with a low-dimensional and fixed-point vector. Further, we propose a fast feature matching method to jointly perform object verification and regression. Experiments on the public OTB benchmark and our collected weak object tracking (WTB) dataset demonstrate the effectiveness of our proposed tracker.
展开▼
