Scene Understanding System of Underground Pipeline Corridors Under Characteristic Degradation Conditions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Underground utility tunnel inspections require accurate scene understanding for safe operations.
• Low-light conditions in tunnels pose significant challenges for current inspection technologies.

Purpose of the Study:

• To develop an object recognition method for low-light environments using enhanced image semantic segmentation.
• To fine-tune a visual language model for generating scene descriptions from tunnel imagery.
• To integrate these components for real-time scene understanding in utility tunnels.

Main Methods:

• Low-light image enhancement and semantic segmentation for object recognition.
• Fine-tuning a visual language model with real underground utility tunnel image data.
• System integration for real-time image processing and scene understanding.

Main Results:

• The improved object recognition model showed a nearly 1% increase in average accuracy.
• The fine-tuned visual-language model achieved over 70% improvement in accuracy and recall.
• The integrated system enables real-time scene understanding and description generation.

Conclusions:

• The proposed method significantly enhances object recognition and scene understanding in low-light tunnel environments.
• Fine-tuning visual language models improves their performance on specific domain data.
• The integrated system offers a practical solution for real-time, accurate underground utility tunnel inspections.