ATTACH数据集

ATTACH数据集是由Ilmenau工业大学神经信息学和认知机器人实验室创建的，专注于人类在装配过程中的双手机械动作理解。该数据集包含51.6小时的装配视频，共95.2万条精细动作标注，通过三个摄像头捕捉，模拟协作机器人的潜在观察视角。数据集特别关注装配环境中工人倾向于用双手同时执行不同动作的特点，因此超过68%的标注与其他标注重叠。此外，数据集还使用了Azure Kinect身体追踪SDK来估计工人的3D骨骼，以增强背景的泛化能力。ATTACH数据集旨在为协作机器人在工业制造中的人类动作理解提供支持，解决实时动作识别和检测的问题。

The ATTACH dataset was created by the Neuroinformatics and Cognitive Robotics Lab at Technische Universität Ilmenau, focusing on the understanding of human two-handed manual assembly actions. This dataset includes 51.6 hours of assembly video footage, with a total of 952,000 fine-grained action annotations, captured by three cameras to simulate the potential observation viewpoints of collaborative robots. The dataset specifically targets the scenario where workers tend to execute distinct actions simultaneously with both hands in assembly environments, resulting in over 68% of annotations overlapping with one another. Additionally, the dataset leverages the Azure Kinect Body Tracking SDK to estimate the 3D skeletons of workers, thereby enhancing the generalization capability across various background contexts. The ATTACH dataset aims to support human action understanding for collaborative robots in industrial manufacturing, addressing the challenges of real-time action recognition and detection.