Citation

Guerrero JCE, España EM, Añasco MM, Lopera JEP. Data Brief 2022; 45: e108610.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.dib.2022.108610

PMID

36164302

PMCID

PMC9508401

Abstract

This article presents a dataset (CAUCAFall) with ten subjects, which simulates five types of falls and five types of activities of daily living (ADLs). Specifically, the data include forward falls, backward falls, lateral falls left, lateral falls right, and falls arising from sitting. The participants performed the following ADLs: walking, hopping, picking up an object, sitting, and kneeling. The dataset considers individuals of different ages, weights, heights, and dominant legs. The data were acquired using an RGB camera in a home environment. This environment was intentionally realistic and included uncontrolled features, such as occlusions, lighting changes (natural, artificial, and night), participants different clothing, movement in the background, different textures on the floor and in the room, and a variety in fall angles and different distances from the camera to the fall. The dataset consists of 10 folders, one for each subject, and each folder includes ten subfolders with the performed activities. Each folder contains the video of the action and all the images of that action. CAUCAFall is the only database that contains details of the lighting lux of the scenarios, the distances from the human fall to the camera and the angles of the different falls with reference to the camera. The dataset is also the only one that contains labels for each image. Frames including human falls recorded were labeled as ''fall'', and ADL activities were marked ''nofall". This dataset is useful for developing and evaluating modern fall recognition algorithms, such as those that apply feature extraction, convolutional neural networks with YOLOv3-v4 detectors, and camera location and resolution increase the performance of algorithms such as OPENPOSE. Thus, the dataset enables knowledge of the real progress of research in this area since existing datasets are used in strictly controlled environments. The authors intend to contribute a dataset with real-world housing environments characteristics.

Language: en

Keywords

Feature extraction; Activities of daily living; Fall detection; Openpose; Uncontrolled environment; YOLO