TWIST: Tension from Wire-image Strand Tracking

Vision-based Automated Cable Tension Monitoring Using Pixel Tracking

Dongyoung Ko, Minsoo Park∗, Soojin Jin, Pa Pa Win Aung, Seunghee Park∗, * : Coressponding Authors

Automation in Construction

Vol. 179, No.106488, pp. 1-19, Nov 2025, DOI:https://doi.org/10.1016/j.autcon.2025.106488

Abstract

Traditional methods for monitoring cable tension rely on indirect measurements such as cable vibrations and often require specialized calibration. These approaches limit the efficiency, and non-contact capability of tension monitoring across various structures. This paper presents a vision-based framework for automated cable tension monitoring, which directly captures image data of internal steel strands. By leveraging advanced computer vision techniques such as zero-shot segmentation, depth estimation, edge detection, and dense pixel tracking critical geometric parameters are extracted and integrated into a kinematic-based model for tension estimation. A calibration-free method for estimating real-world pixel size, derived from the helical geometry of the strands, enables field deployment without the need for camera setup information. Experimental results show strong correlation with reference data, achieving a mean absolute error of 4.94% under elastic conditions. These findings pave the way for a promising alternative in vision-based structural health monitoring for prestressed structures.

System Requirements

All experiments were conducted under the following hardware and software configuration:

Hardware Environment

• Workstation: Dell Precision 7920 Rack
• CPU: Intel Xeon Silver 4210R (10 cores, 20 threads, 2.4–3.2 GHz)
• Memory: 128 GB DDR4-3200 ECC
• GPU: NVIDIA RTX A6000 (48 GB GDDR6 ECC)
• Storage: 1 TB NVMe SSD
• Operating System: Ubuntu 20.04 LTS

Software Environment

• Python: 3.10
• PyTorch: 2.1.0 (with CUDA 12.1)
• TensorFlow: 2.15 (with CUDA 11.7)
• scikit-learn: 1.4

1. Preparation

• Clone or download the code package from this repository.
• Set the working directory:

MainFolder = "/your/custom/path"

Experimental video of the tensioning strand should be placed in the Input_Video/ folder.

Depth-Anything

• Download the depth_anything_v2_vitl.pth checkpoint file and place it in the Depth-Anything-V2/checkpoints/ directory.
• The pre-trained checkpoint can be downloaded from:
  https://github.com/DepthAnything/Depth-Anything-V2#pre-trained-models

SAM (Segment Anything Model)

• Download the sam2.1_hiera_large.pt checkpoint file and place it in the sam2_repo/checkpoints/ directory.
• The pre-trained checkpoint can be downloaded from:
  https://github.com/facebookresearch/sam2

Cotracker

• Download the scaled_offline.pth checkpoint file and place it in the co-tracker/cotracker/checkpoints/ directory.
• The pre-trained checkpoint can be downloaded from:
  https://github.com/facebookresearch/co-tracker

	Module	GitHub	Version / Commit	License
1	SAM 2 – Segment Anything Model V2	facebookresearch/sam2	sam2.1	Apache 2.0
2	Depth Anything V2	DepthAnything/Depth-Anything-V2	vitl	MIT
3	CoTracker	facebookresearch/co-tracker	CoTracker V3	Apache 2.0

LDC (Lightweight Dense CNN for Edge Detection)

• Download code from https://github.com/xavysp/LDC
• For custom usage:

• [main.py] Set input image directory:
  --input_val_dir=/your/image/folder

• [main.py] Set output directory for results:
  --output_dir=/your/output/folder

• [main.py] Match image resolution to your data:
  --img_width=YOUR_WIDTH and --img_height=YOUR_HEIGHT
  (e.g., 1080×710)

• [img_processing.py] Adjust edge map threshold:

  tensor = np.where(tensor >= 0.70, tensor, 0)  # threshold for edge map

2. Run the Main Pipeline

Open and execute the notebook:

Run.ipynb

This notebook will guide you through the full processing pipeline.

Processing Steps

Step 1: Semantic Segmentation (Sengmentation Anything Model V2)

• Segment the strands using the SAM2 framework.
• Output: Preliminary masks of individual strands.

Step 2: Depth Estimation (Depth Anything V2)

• Use a monocular depth estimator to generate depth maps.
• Compute the ROI by intersecting:
  • SAM2 segmentation mask
  • Valid depth region
• This ensures a more precise ROI for subsequent wire tracking and analysis.

Step 3: Feature Tracking (CoTracker V3)

• Track the displacement of strands across frames using the CoTracker algorithm.
• Output: Time-series of strand displacements.

Lay Angle Estimation

• The lay angle is estimated using LDC edge detection applied only to the first frame of the video.
• Run Edge(LDC).ipynb to estimate the lay angle for all frames.

Citation

@article{ko2025vision,
  title={Vision-based automated cable tension monitoring using pixel tracking},
  author={Ko, Dongyoung and Park, Minsoo and Jin, Sujin and Aung, Pa Pa Win and Park, Seunghee},
  journal={Automation in Construction},
  volume={179},
  pages={106488},
  year={2025},
  publisher={Elsevier}
}

Contributors

Dongyoung Ko∗

Minsoo Park∗

Taebum Lee

Sujin Jin

Pa Pa Win Aung

Seunghee Park∗