SWIT-Dataset

Scaffolding Worker IMU time-series dataset for deep learning-based construction-site behavior recognition
Minsoo Park*, Seongwoo Son, Yuntae Jeon, Dongyoung Ko, Mingeon Cho, Seunghee Park*, Advanced Engineering Informatics
Vol. 65 Part B, No. 103232, pp. 1–17, March 2025, DOI: https://doi.org/10.1016/j.aei.2025.103232

Download the dataset

LINK (google drive) to the dataset

Introduction

Scaffolding Worker IMU Time-series (SWIT) Dataset for Deep Learning-based Construction Site Behavior Recognition

Unsafe Behavior Categories

• 4 behaviors are related to hazardous construction environments (Obstacle; Slip; Hole; Unstable)
• 2 behaviors are related to worker-compliance (Climb; Step-off)
• 2 behaviors are related to musculoskeletal risk postures (Squat; Kneel)
• 1 behavior is related to emergency situations (Fall-down)
• 1 behavior is related to general situations (Walk) [involved walking forward, walking sideways, walking backward, briefly stopping, and working in a safe manner]

Experimental Setup for dataset acquisition

• Participants: 27 construction engineering expert
• IMU (Inertial Measurement Unit) Sensor: EBIMU24GV5 wireless IMU sensor, 100Hz sampling rate
• Sensor placement: Attached to the lower back of participants
• Camera: GoPro Hero 12 black, recording at 25 FPS
• Data collection method: Each participant repeated 10 behaviors 10 times each
• Data synchronization: Time synchronization between IMU sensor data and video data
• Pose estimation: 17 keypoints extracted from video using YOLOv7-pose

Pre-proccesing dataset for encoding (6 x 200 x 200)

Use 'Datatset_Image_Encoding.py' file:

• Data from 6 channels of the IMU sensor: acceleration(g) x, y, z and angular velocity(degree/second) x, y, z
• Defined 200 data points over 2 seconds as one behavior data instance
• Applied a sliding window with 0.1-second (10 data points) intervals to extract data
• Used Gramian Angular Difference Field (GAF) to transform time-series data into images
• Encoded each of the 6 channels into images and stacked them. Finally, generated data with dimensions of 6 x 200 x 200

Download the SWIT-Dataset

• It will be appeared soon
• `IMU_Sensor_Raw.zip' is compressed file containing raw data of time series dataset containing IMU experiment data for a total of 27 subjects across 10 different tasks (Task 1 to Task 10).
• The dataset includes x, y, and z axis readings from accelerometers, gyroscopes, and magnetometers, all recorded at a frequency of 100 Hz.
• `Keypoints_Raw.zip' is compressed file containing keypoint estimation data.
• `keypoints_(200).zip' is compressed file containg segmented into 2-second intervals, synchronized with the IMU.

Citation

• (https://doi.org/10.1016/j.aei.2025.103232)

@article{park2025scaffolding,
  title={Scaffolding worker IMU time-series dataset for deep learning-based construction site behavior recognition},
  author={Park, Minsoo and Son, Seongwoo and Jeon, Yuntae and Ko, Dongyoung and Cho, Mingeon and Park, Seunghee},
  journal={Advanced Engineering Informatics},
  volume={65},
  pages={103232},
  year={2025},
  publisher={Elsevier}
}