聋人健美操学习眼动追踪数据集构建与分析

Construction and analysis of eye-tracking dataset for deaf aerobics learning

摘要:

现有的公开眼动数据集中，针对聋人群体的采集与构建几乎处于空白状态，且眼动类型分布不均衡，规模较大及标签覆盖完整的数据集较少。为解决这些问题，本文提出并构建了聋人健美操学习眼动数据集（Eye Tracking for Deaf Aerobics, ETFDA）。首先，借助眼动追踪技术采集112位聋人、54位听人的数据，填补聋人数据的缺失。其次，通过录制健美操视频作为刺激材料，诱发扫视和平滑追踪眼动类型，改善眼动类型分布不均衡的问题。最后，经过预处理、特征提取以及标注等步骤，构建了包含约133万条记录、覆盖四种眼动类型标签的数据集，以丰富现有眼动数据集资源。基于此数据集，一方面通过对比聋人与听人的眼动特性，结果揭示了二者间具有显著性差异，为理解聋人视觉认知机制提供了新的视角；另一方面，通过与其他公开数据集比较分析与算法验证，结果表明该数据集眼动类型分布是更为均衡的，标签是精细有效的，具有重要的应用价值。

Abstract:

Existing public eye-tracking datasets almost entirely lack data collection for the deaf population, exhibit unbalanced eye movement type distributions, and rarely include large-scale datasets with comprehensive labeling. To address these issues, this paper proposes and constructs the Eye Tracking for Deaf Aerobics (ETFDA) dataset. First, using eye-tracking technology, data from 112 deaf and 54 hearing individuals are collected to fill the gap in deaf data. Second, aerobics videos are recorded as stimulus materials to induce saccadic and smooth pursuit eye movements, addressing the unbalanced distribution of eye movement types. Finally, through preprocessing, feature extraction, and annotation, a dataset containing approximately 1.33 million records covering four types of eye movement labels is constructed to enrich existing eye-tracking dataset resources. Based on this dataset, comparing the eye movement characteristics of deaf and hearing individuals reveals significant differences, providing new insights into understanding the visual cognitive mechanisms of the deaf. Additionally, through comparative analysis and algorithm validation with other public datasets, the results show that this dataset has a more balanced distribution of eye movement types and more refined and effective labels, offering significant application value.