Real-Time Water Surface Object Detection Based on Improved Faster R-CNN

Area of Science:

• Computer Vision
• Environmental Monitoring
• Deep Learning

Background:

• Traditional methods like background subtraction and image segmentation struggle with variable conditions (sunlight, waves) and feature sensitivity, limiting their effectiveness for water surface object detection.
• The River Chief System in China requires timely detection of water surface floats, highlighting the need for robust and efficient detection solutions.
• Existing deep learning methods show promise but require optimization for real-time performance and accuracy in complex natural environments.

Purpose of the Study:

• To propose a novel, real-time water surface object detection method tailored for complex natural scenes.
• To address the limitations of classical and existing deep learning approaches in detecting water surface objects, particularly floating debris.
• To develop a system that meets the accuracy and speed requirements for environmental monitoring applications like the River Chief System.

Main Methods:

• A real-time water surface object detection method based on the Faster R-CNN architecture is proposed.
• The network integrates low-level and high-level features through a two-module design to enhance detection accuracy.
• Customized anchor scales and aspect ratios were determined by analyzing object scale distribution in the dataset, improving robustness for multi-scale objects.

Main Results:

• The proposed method achieved a mean average precision (MAP) of 83.7% in detecting water surface floats.
• The system demonstrated a detection speed of 13 frames per second, suitable for real-time applications.
• Validation was performed using a three-day video surveillance stream of the North Canal in Beijing, confirming performance in a real-world scenario.

Conclusions:

• The developed Faster R-CNN based method offers high accuracy and robustness for detecting multi-scale objects in complex natural water scenes.
• The method's performance meets the accuracy and speed requirements for online water surface object detection, particularly for floating debris monitoring.
• This approach provides a viable solution for environmental monitoring systems requiring efficient and precise detection of water surface contaminants.