harvardairobotics/IMU-HAR

IMU-HAR（IMU人类活动识别）是一个基于Ego4D自我中心视频语料库构建的约160K样本的行为活动识别数据集。它配对了来自Ego4D的头戴式IMU信号（6轴加速度计和陀螺仪，50Hz）与行为级动作标签——关注用户的功能性行为，而不仅仅是物理运动。与之前针对运动原语（如行走、站立）的IMU数据集不同，该数据集关注运动背后的功能意图，涵盖五个可直接应用于AR助手系统的行为类别。标签通过LLM-人类反馈标注循环生成：Qwen3-8B在355K个叙述上生成了初始标签和推理链，然后12名人类标注者经过两轮验证了27K个黄金子集，形成了一个四层质量框架。经验证的标签通过近重复叙述传播，最终扩展到约160K个样本。注意：该数据集提供注释标签和元数据，原始IMU信号需从Ego4D单独获取，并通过video_uid和timestamp_sec匹配。IMU-HAR (IMU Human Activity Recognition) is a ~160K-sample behavioral activity recognition dataset built from the Ego4D egocentric video corpus. It pairs head-mounted IMU signals (6-axis accelerometer + gyroscope at 50 Hz) from Ego4D with behavioral-level action labels — what the user is functionally doing, not just how they are physically moving. Unlike prior IMU datasets targeting motion primitives (walking, standing), this dataset targets the functional intent behind motion, spanning five behavioral categories directly applicable to AR assistant systems. Labels were produced through an LLM–human backfeed annotation loop: Qwen3-8B generated initial labels and reasoning chains over 355K narrations, then 12 human annotators verified a 27K gold subset across two rounds, yielding a four-tier quality framework. Verified labels were propagated to near-duplicate narrations for a 5.8× expansion to the final ~160K samples. Note: This dataset provides the annotation labels and metadata. The raw IMU signals must be obtained separately from Ego4D and matched by video_uid and timestamp_sec.