可穿戴设备场景数据集

UESTC-MMEA-CL由电子科技大学智能视觉信息处理与通信实验室（IVIPC Lab）构建并发布。随着GoPro、智能眼镜等可穿戴设备的普及，第一人称视觉成为计算机视觉领域的重要研究方向。然而，现有的深度神经网络在处理动态变化的现实场景时，往往面临“灾难性遗忘”（Catastrophic Forgetting）问题，即在学习新任务时会遗忘旧知识。现有的数据集大多侧重于第三人称视角或缺乏同步的多模态传感器数据，无法满足连续学习算法的研究需求。作为首个专门针对多模态第一人称行为识别连续学习设计的数据集，UESTC-MMEA-CL填补了该领域的空白，对于推动人工智能终身学习（Lifelong Learning）、智能可穿戴助手、健康监测及人机交互系统的研发具有重要的学术价值和应用意义。该数据集基于电子科技大学自研的智能眼镜观测产生。该设备集成了第一人称RGB摄像头与惯性测量单元（IMU）传感器。数据采集过程由10名志愿者参与，在办公楼、校园、家庭等自然场景下进行。主要记录了同步的第一人称视频流（640×480分辨率，25FPS）、三轴加速度（Acceleration）、三轴角速度（Gyroscope）等多模态观测值。为确保数据质量，所有传感器数据均经过中值滤波处理以去除噪声，并完成了严格的时间同步校准。数据集涵盖了32类日常交互活动，包括基本的身体运动（如上下楼梯、行走）、手部操作（如洗碗、切水果、打字）、以及休闲活动（如看电视、聊天）等。为了适应连续学习任务的评估，这32类活动被划分为不同的增量学习步骤（如16-8-4的任务划分设置）。数据样本总时长为30.4小时，数据大小约28.3GB。共包含约6400个样本片段（每类约200个样本），所有数据均已划分为训练集、验证集和测试集（比例7:2:1）供研究使用。

UESTC-MMEA-CL is constructed and released by the Intelligent Visual Information Processing and Communication Laboratory (IVIPC Lab) at the University of Electronic Science and Technology of China (UESTC). With the popularization of wearable devices such as GoPro and smart glasses, first-person vision has emerged as a critical research direction in computer vision. However, existing deep neural networks often encounter the "Catastrophic Forgetting" issue when handling dynamically changing real-world scenarios, i.e., they tend to erase previously acquired knowledge while learning new tasks. Most existing datasets either prioritize third-person perspectives or lack synchronized multi-modal sensor data, thus failing to meet the research demands of continual learning algorithms. As the first dataset specifically tailored for multi-modal first-person activity recognition in continual learning settings, UESTC-MMEA-CL fills a critical gap in this domain, holding substantial academic value and practical significance for advancing the research and development of artificial intelligence lifelong learning, intelligent wearable assistants, health monitoring, and human-computer interaction systems.

This dataset is generated using UESTC’s self-developed smart glasses, which integrate a first-person RGB camera and an Inertial Measurement Unit (IMU) sensor. The data collection process involved 10 volunteers, and was conducted across natural scenarios including office buildings, campuses, and homes. It primarily records synchronized multi-modal observations, such as first-person video streams (640×480 resolution, 25 FPS), triaxial acceleration, triaxial angular velocity (Gyroscope), and other related measurements. To ensure data quality, all sensor data were processed with median filtering to eliminate noise, and strict time synchronization calibration was performed. The dataset encompasses 32 categories of daily interactive activities, including basic body movements (e.g., ascending/descending stairs, walking), hand manipulations (e.g., dishwashing, fruit cutting, typing), and leisure activities (e.g., watching television, chatting), among others. To support the evaluation of continual learning tasks, these 32 activity categories are partitioned into distinct incremental learning steps (e.g., the 16-8-4 task division setup). The total duration of all data samples amounts to 30.4 hours, with an overall data size of approximately 28.3 GB. It contains roughly 6400 sample segments (about 200 samples per category), and all data have been split into training, validation, and test sets at a ratio of 7:2:1 for research purposes.

搜集汇总

背景与挑战

背景概述

该数据集是由电子科技大学构建的多模态第一人称行为识别连续学习数据集，包含32类日常交互活动的同步视频流和惯性测量单元数据，总时长30.4小时，数据量28.22GB。该数据集填补了第一人称视角连续学习研究的空白，适用于人工智能终身学习和智能可穿戴设备研发。