Automated Detection and Tracking of Sewer Pipeline Structural Elements

Fabian Wientjes, Bram Ton, Henry Maathuis

Track: Type A (Regular Papers)

Keywords: Object Detection, Computer Vision, Civil Infrastructure, Sewer Pipe Inspection, YOLO, Tracking, Class Imbalance

Abstract: Regular inspection of sewer systems is essential to ensure reliable operation and prevent costly failures or environmental harm. Current practice relies on manual review of camera-based inspection footage, a process that is labor-intensive, subjective, and difficult to scale. While object detection methods offer a promising alternative, conventional frame-based approaches cannot reliably determine whether detections across consecutive frames represent the same object. This limits both detection reliability and the ability to map unique sewer components which are critical steps for condition assessment and maintenance planning. In this work, we present a method for detecting and counting unique sewer objects, specifically inlets and joints, across frames in 360° inspection imagery. Unlike prior approaches focused solely on defect detection, our framework addresses the challenge of object-level consistency across video data. The method achieves high accuracy (96.8% mAP@50 overall, 96.5% for inlets, 97% for joints on the validation set), bridging the gap between frame-level recognition and pipeline-level evaluation. This advances automated sewer inspection toward industry-standard reporting, enabling more efficient, objective, and scalable maintenance planning.

Serve As Reviewer: ~Bram_Ton1

Submission Number: 32