FIR Human

This dataset contains video-clips of five volunteers developing daily life activities. Each video-clip is recorded with a Far InfraRed (FIR) camera and includes an associated file which contains the three-dimensional and two-dimensional coordinates of the main body joints in each frame of the clip. This way, it is possible to train human pose estimation networks using FIR imagery.It contains over 250.000 2D and 3D human poses and their corresponding FIR images. The dataset is recorded by 5 volunteers (4 males, 1 female) engaged in 27 different action classes, including falls of different kinds.The FIR video-clips are recorded at 23.98 frames per second with a resolution of 480 x 640 pixels. The annotations associated with them include accurate 3D positions of the 19 main body joints provided by a high-speed motion capture system and their projections onto the image plane.本数据集包含5名志愿者进行日常活动的视频片段。每个视频片段均由远红外（Far InfraRed, FIR）相机录制，并附带关联文件，该文件存储了视频片段每一帧中主要人体关节的三维与二维坐标。借此可基于远红外图像训练人体姿态估计网络。本数据集包含超过25万个二维与三维人体姿态及其对应的远红外图像。该数据集由5名志愿者（4男1女）录制，志愿者共完成27类不同动作，涵盖多种类型的跌倒行为。远红外视频片段的录制帧率为23.98帧/秒，分辨率为480×640像素。相关标注包含由高速运动捕捉系统获取的19处主要人体关节的精确三维位置，以及这些关节在图像平面上的投影坐标。

搜集汇总

背景与挑战

背景概述

FIR Human是一个用于人体稳定性评估和跌倒检测的远红外（FIR）图像数据集，包含五名志愿者执行27种动作（包括日常活动和不同类型跌倒）的视频片段，超过25万个人体姿态和对应图像，并附带身体关节的2D和3D坐标。数据集分为训练、验证和跌倒检测三个块，适用于计算机视觉和机器学习任务，如人体姿态估计和跌倒预防研究。