Seismic Intensity Prediction with a Low-Computational-Cost Transformer-Based Tracking Method

Honglei Wang, Zhixuan Bai, Ruxue Bai, Liang Zhao, Lin Ma, Yamin HanSensors 2025 · DOI

TL;DR

This paper introduces LCCTV, a novel, low-computational-cost transformer-based visual tracking method for predicting seismic intensity from surveillance videos. By efficiently tracking calibration targets and analyzing their motion, LCCTV offers a cost-effective and real-time alternative to traditional seismograph equipment for disaster relief and early warning systems.

What Problem Does This Paper Address?

Accurate and timely prediction of seismic intensity is crucial for effective disaster relief and early warning systems. Traditional methods rely on seismograph equipment, which can be slow to respond and expensive to deploy widely. While advancements in computer vision have opened possibilities for using surveillance videos, existing methods often require manual feature extraction or are computationally intensive, limiting their real-time applicability, especially on resource-constrained devices. This paper aims to bridge this gap by developing a computationally efficient visual tracking method that can directly infer seismic intensity from video data.

Methods

The core of this work is the Low-Computational-Cost Transformer-Based (LCCTV) Visual Tracking method. LCCTV employs a hybrid, asymmetric Siamese network architecture. Key innovations include:

  • Two-Stream Design with Shared Weights: This approach separates template and search region feature extraction initially, reducing redundant computations.
  • Skip Multi-head Self-Attention (SMSA) Module: Inspired by the high similarity of self-attention in adjacent transformer layers, SMSA intelligently skips redundant computations by leveraging features from neighboring layers. This significantly reduces computational cost and model size without sacrificing performance.
  • Global Interaction Stage: After initial feature extraction, template and search region features are combined for global interaction, enhancing the modeling of relationships.
  • Post-processing: A Butterworth filter is used to smooth the extracted movement trajectories, removing low-frequency interference.
  • Intensity Prediction: Seismic intensity is then predicted based on the derived velocity and acceleration from the smoothed trajectories.

Datasets

  • Newly Introduced Dataset: The authors created a novel earthquake video dataset captured in a laboratory environment. This dataset was generated using a shaking table to simulate seismic events, with seismic levels obtained from a seismic station. It includes approximately 18 hours of video data, resulting in over 700,000 images covering earthquake intensities from degree II to VI.
  • Public Datasets for Training: The model was trained using four public datasets: GOT-10k, COCO, LaSOT, and TrackingNet.

Key Insights and Ideas

  • Vision as a Viable Seismic Sensor: The research strongly suggests that standard surveillance cameras, when paired with sophisticated tracking algorithms like LCCTV, can serve as effective sensors for seismic intensity estimation, offering a low-cost alternative to specialized equipment.
  • Efficiency through Intelligent Skipping: The SMSA module's ability to selectively skip redundant self-attention computations is a critical insight. It demonstrates that not all attention operations are equally valuable, and intelligent pruning can lead to significant computational gains.
  • Real-time Performance on Edge Devices: The successful deployment and real-time performance of LCCTV on an edge device (Nvidia Jetson Orin NX) is a major breakthrough, paving the way for practical, on-site seismic monitoring.
  • Robustness Across Intensity Levels: LCCTV shows remarkable accuracy, especially in higher intensity scenarios, indicating its ability to capture pronounced structural vibrations.

Key Results

  • Seismic Dataset Performance: On their newly created seismic dataset, LCCTV achieved an average prediction accuracy of 80.5%. It significantly outperformed state-of-the-art trackers like HiT-Base (72.2%) and AsymTrack-B (70.2%).
  • High Intensity Accuracy: The method demonstrated particularly high accuracy in medium to high-intensity scenarios, reaching 89.5% for degree V and 95.0% for degree VI earthquakes.
  • Real-time Capability: LCCTV achieved real-time performance, running at over 20 FPS on the Nvidia Jetson Orin NX edge device, meeting the standard for real-time applications.
  • Public Benchmark Performance: On the LaSOT dataset, LCCTV achieved the highest performance among compared methods, with 65.3% AUC, 75.4% PNorm, and 69.9% P.
  • Real-world Validation: In a real-world Lushan earthquake scenario, LCCTV's estimated intensity of 6.0 closely matched the official intensity degree VI for that location. Similarly, for the Luding earthquake, its estimations were nearly identical to instrumental measurements.

Limitations and Future Work

  • Low-Intensity Sensitivity: The paper acknowledges that LCCTV exhibits lower prediction accuracy for low-intensity earthquakes (degrees II-III). This is attributed to the subtle nature of vibrations, making them difficult to distinguish from background noise and challenging for precise scale estimation.
  • Future Work: The authors plan to address these limitations by:
  • Incorporating sub-pixel motion estimation to enhance sensitivity to small vibrations.
  • Adopting more precise scale estimation techniques to reduce accumulated errors.
  • Further developing the framework to improve accuracy in low-intensity scenarios.

Summary updated

Abstract

The prediction of seismic intensity in an accurate and timely manner is needed to provide scientific guidance for disaster relief. Traditional seismic intensity prediction methods rely on seismograph equipment, which is limited by slow response times and high equipment costs. In this study, we introduce a low-computational-cost transformer-based (LCCTV) visual tracking method to predict seismic intensity in surveillance videos. To this end, an earthquake video dataset is proposed. It is captured in the laboratory environment, where the seismic level is obtained through seismic station simulation. With the proposed dataset, a low-computational-cost transformer-based visual tracking method is first proposed to estimate the movement trajectory of the calibration board target in videos in real time. In order to further improve the recognition accuracy, we then utilize a Butterworth filter to smooth the generated movement trajectory so as to remove low-frequency interference signals. Finally, the seismic intensity is predicted based on the velocity and acceleration derived from the smoothed movement trajectory. Experimental results demonstrated that the LCCTV outperformed other state-of-the-art approaches. The findings confirm that the proposed LCCTV can serve as a low-cost, scalable solution for seismic intensity analysis.
BibTeX
@inproceedings{wang2025seismic,
  author = {Honglei Wang and Zhixuan Bai and Ruxue Bai and Liang Zhao and Lin Ma and Yamin Han},
  title = {{Seismic Intensity Prediction with a Low-Computational-Cost Transformer-Based Tracking Method}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25206269},
}

Loading related content...

Figure