Acute dizziness and vertigo are common symptoms in the emergency departments, among which central vertigo, especially stroke, has dangerous characteristics such as rapid onset, irreversible results, and difficulty in distinguishing from common vertigo. It is still challenging to differentiate between peripheral vertigo and central vertigo via the HINTS (Head impulse test, Nystagmus test, Test of skew) examination. In this work, we develop a nystagmus recognition system for stroke preliminary screening in the emergency room based on a hand-held device. Specifically, an effective and lightweight pupil segmentation network is proposed by designing Multi-scale Feature Input (MSFI) module and Multi-scale Feature Fusion (MSFF) module for pupil tracking. Extensive experiments show the proposed segmentation model can obtain competitive performance while maintaining low parameters and low computational burden, and the proposed system is effective and available to distinguish between peripheral acute vestibular syndrome and central acute vestibular syndrome.