Defect Detection and 3D Reconstruction of Complex Urban Underground Pipeline Scenes for Sewer Robots

Area of Science:

• Computer Vision
• Robotics
• Urban Infrastructure Monitoring

Background:

• Detecting sewer defects is vital for urban underground structure health.
• Current image-based models are often complex and resource-intensive.
• Detailed damage information and edge deployment are lacking in existing solutions.

Purpose of the Study:

• To develop an efficient, lightweight computer vision framework for robotic sewer defect detection.
• To enable real-time defect identification and 3D pipeline reconstruction on edge devices.
• To improve the accuracy and reduce the computational demand of sewer inspection.

Main Methods:

• Proposed a lightweight Sewer-YOLO-Slim model based on YOLOv7-tiny, incorporating channel pruning.
• Implemented a multiview reconstruction technique for 3D pipeline modeling.
• Deployed the model on edge devices using TensorRT for acceleration.

Main Results:

• Sewer-YOLO-Slim achieved a 60.2% reduction in model size and a 1.5% mAP increase to 93.5%.
• The pruned model size is only 4.9 MB, with a detection speed of 15.3 ms/image.
• 3D reconstruction yielded a maximum measurement error of 0.57 m.

Conclusions:

• The proposed framework effectively detects sewer defects with high accuracy and low computational cost.
• The lightweight model is suitable for edge device deployment, enabling real-time inspection.
• 3D modeling provides valuable insights for pipeline data visualization and defect measurement.