River Surface Velocity Measurement for Rapid Levee Breach Emergency Response Based on DFP-P-LK Algorithm

Zhao-Dong Xu, Zhi-Wei Zhang, Ying-Qing Guo, Yan Zhang, Yang Zhan2024 · DOI

TL;DR

This paper introduces a novel Dynamic Feature Point Pyramid Lucas-Kanade (DFP-P-LK) algorithm for accurately measuring river surface velocity using video imagery. This method is crucial for rapid levee breach emergency response, offering a non-invasive, cost-effective, and robust solution compared to traditional techniques.

What Problem Does This Paper Address?

Climate change is increasing the frequency and intensity of extreme weather events, leading to a higher risk of levee breaches and subsequent large-scale floods. Rapid and accurate measurement of river surface velocity is critical for effective emergency response and mitigation efforts. Traditional methods, such as buoy systems or velocimetry poles, are often impractical, dangerous, and unable to provide real-time, high-frequency data, especially in complex river segments or during emergencies. While video image velocimetry offers a promising non-contact alternative, existing methods struggle with robustness and accuracy in dynamic and challenging environments like turbulent water, varying lighting, and reflections.

Methods

The core of this work is the Dynamic Feature Point Pyramid Lucas-Kanade (DFP-P-LK) algorithm. It builds upon the Pyramid-Based LK optical flow method by incorporating two key innovations:

  • Feature Point Fusion Detection Strategy: This strategy combines the strengths of SURF (for efficiency and robustness in detecting many candidate points) and Shi-Tomasi (for evaluating corner quality). It aims to identify more stable and reliable feature points by generating descriptors and calculating minimum eigenvalues.
  • Dynamic Feature Point Update Strategy: Unlike traditional LK methods that only re-detect points when tracking fails, DFP-P-LK actively manages feature points. It uses a maximum consecutive processing frame limit (f_max) to force re-detection and a minimum feature point threshold (P_min) to ensure a sufficient number of points for robust tracking, preventing degradation over time.

The DFP-P-LK algorithm is then integrated into a river surface velocity measurement model. This model involves:

  1. Image Preprocessing: Grayscale conversion, filtering, and enhancement to improve image quality.
  2. Optical Flow Estimation: Using the DFP-P-LK algorithm to track feature point motion.
  3. Optical Flow-Velocity Conversion: A model that translates the estimated optical flow (pixel movement) into actual water velocity, taking into account camera calibration parameters and video information.

Datasets

The proposed DFP-P-LK algorithm and the river surface velocimetry method were validated using two datasets:

  • Flying Chairs Dataset: A standard dataset for optical flow estimation, containing synthesized image pairs and their corresponding ground truth optical flow fields. This dataset was used to evaluate the algorithm's accuracy in optical flow estimation.
  • Homemade Water Flow Dataset: This dataset comprises 82 videos captured from different sections of the Sancha River and Zihu Creek in Nanjing, China. These videos represent various conditions (sunny, overcast, after rain; clear, turbid, choppy) and flow velocities ranging from 0.43 m/s to 2.06 m/s. This dataset was specifically created to test the river surface velocity measurement model in realistic scenarios.

Key Insights and Ideas

  • Active Feature Point Management is Crucial: The paper strongly suggests that passively waiting for feature point failure is insufficient for dynamic environments like water surfaces. An active, dynamic update strategy is essential for maintaining tracking robustness and accuracy over time.
  • Fusion of Detection Methods Yields Better Points: Combining SURF and Shi-Tomasi corner detection provides a more robust set of initial feature points compared to using either method alone.
  • Pyramid Structure Enhances Robustness: The use of an image pyramid in the LK algorithm allows for processing at different resolutions, which helps in handling large displacements and improving overall stability.
  • Real-World Performance Matters: While standard optical flow datasets are useful, the homemade water flow dataset highlights the challenges of real-world conditions (turbidity, reflections, varying light) and the need for algorithms specifically tuned for these scenarios.

Key Results

  • Improved Optical Flow Accuracy: On the Flying Chairs dataset, the DFP-P-LK method achieved a mean angular error (MAE) of 16.49 degrees and a mean endpoint error (APE) of 6.91 pixels. This represents a significant reduction of 43.39% in MAE and 12.97% in APE compared to the traditional LK algorithm.
  • Accurate River Velocity Measurement: For the homemade water flow dataset, the DFP-P-LK method achieved an average relative error of less than 15% in river surface velocity measurement. This is significantly better than traditional LK (average 22.24%), Farneback (average 21.48%), and Template Matching (average 34.885%) methods, especially in challenging conditions.
  • Robustness in Dynamic Conditions: The dynamic update strategy for feature points demonstrated a significant improvement in maintaining the number of tracked feature points over time, preventing rapid decline and ensuring continuous tracking.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Flowmeter Measurement Error: In low-speed flow conditions, the accuracy of the reference flowmeter itself can be a limiting factor, affecting the evaluation of the algorithm's performance.
  • Difficulty in Clear/Gentle Water: In very clear and gentle water, finding suitable and reliable feature points for tracking is challenging, leading to increased errors.
  • Reflection Interference: Reflections from trees or other objects on the riverbank can interfere with feature point detection and matching.
  • Future Work:
    • Optimized Feature Point Selection: Developing more robust strategies for feature point detection and tracking, especially under varying lighting and reflection conditions.
  • Drone-Assisted Feature Point Generation: Using drones to scatter materials (straw, leaves) onto the water surface to create artificial feature points in challenging scenarios.
  • Deep Learning Integration: Incorporating deep learning techniques to automatically learn more robust features and improve optical flow estimation accuracy and stability in complex environments.
  • Handling Light Change and Occlusion: Further improving the algorithm's performance in scenarios with significant light variations and occlusions.

Summary updated

BibTeX
@misc{xu2024river,
  author = {Zhao-Dong Xu and Zhi-Wei Zhang and Ying-Qing Guo and Yan Zhang and Yang Zhan},
  title = {{River Surface Velocity Measurement for Rapid Levee Breach Emergency Response Based on DFP-P-LK Algorithm}},
  year = {2024},
  doi = {10.3390/s24165249},
}

Loading related content...

Figure