Construction Motion Data Library: An Integrated Motion Dataset for On-Site Activity Recognition

Through collecting 16 relatively small-scale motion datasets and conducting a series of in-lab expreiment, we established a 3D skeleton dataset for recognizing construction worker actions. All skeleton data were processed in four major steps, including uniform data extraction, skeleton structure alignment, resampling, and coordination transformation. Then all the aligned skeleton data will be manually annotated into four activity categories and assigned with labels. Experiment version: It contains over 61,275 samples (10 million frames) from 73 classes performed by about 300 different subjects.The dataset includes four fundamental categories of activities, including Production Activities(12), Unsafe Activities(38), Awkward Activities(10), and Common Activities(13). However, We have carefully reviewed the licenses of all the current datasets. We found more than half of the datasets did not specify their licenses and usage policy. Therefore, in this version, we only shared the tagged and processed dataset that clearly allows redistribution and modification. For the rest of the datasets, we highlighted their URL and doi (all of them are publicly accessible and free for use). Instead of providing the processed data, we public the full preprocess codes on GitHub, which could be used to retag and process (such as converting to predefined .bvh files). All readers and users could process the source dataset by themselves. Public version： Construction Motion Data Library(CML) contains 6131 samples(ALL_DATA); among them, and 4333 samples are highly related to construction activities ( Construction_Related_Data). GitHub: https://github.com/YUANYUAN2222/Integrated-public-3D-skeleton-form-CML-library.

本研究通过收集16个小规模运动数据集并开展一系列实验室实验，构建了用于建筑工人动作识别的三维骨骼数据集（3D skeleton dataset）。所有骨骼数据均经过四大核心步骤处理，包括统一数据提取、骨骼结构对齐、重采样与坐标变换。随后，所有对齐后的骨骼数据将被人工标注为四大活动类别并赋予对应标签。
实验版本：该数据集包含约300名不同受试者完成的73个类别的共计61275余条样本（总计1000万帧数据），涵盖四大基础活动类别，分别为生产活动（Production Activities，12）、不安全行为（Unsafe Activities，38）、异常姿势活动（Awkward Activities，10）与日常活动（Common Activities，13）。
然而，本团队对本次收集的全部数据集的许可协议与使用政策进行了逐一核查，发现超过半数的数据集未明确标注其使用许可与使用规则。因此，在本版本中，我们仅共享明确允许二次分发与修改的已标注处理后数据集；对于其余数据集，我们已标注其公开可访问且可免费使用的URL与数字对象标识符（DOI，Digital Object Identifier）。
本数据集未直接提供处理后的成品数据，而是在GitHub平台公开了完整的预处理代码，可用于重新标注与处理（例如转换为预设的.bvh格式文件），所有读者与用户可自行对原始数据集进行处理。
公开版本：建筑运动数据库（Construction Motion Data Library, CML）包含6131条样本（ALL_DATA），其中4333条样本与建筑活动高度相关（Construction_Related_Data）。GitHub仓库地址：https://github.com/YUANYUAN2222/Integrated-public-3D-skeleton-form-CML-library.

搜集汇总

背景与挑战

背景概述

Construction Motion Data Library是一个用于建筑工人动作识别的3D骨架数据集，包含61,275个样本（约1000万帧），分为生产活动、不安全活动、尴尬活动和常见活动四大类。数据集经过统一数据提取、骨架结构对齐、重采样和坐标转换等处理步骤，并公开了预处理代码以便用户自行处理原始数据。