Improved detection and reidentification algorithm for natural gas pipeline excavator dynamic tracking based on deep learning

• Published in: Measurement : journal of the International Measurement Confederation Vol. 257; p. 118567
• Main Authors: Wang, Weiqiang, He, Wei, Bai, Mingxing, Zhu, Baikang, Lv, Xiaofang, Ji, Shouhong, Hong, Bingyuan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 15.01.2026
• Subjects: Deep learning; Natural gas pipeline; Pan-Tilt-Zoom camera; Target tracking; Visual servo; Deep learning; Visual servo; Target tracking; Natural gas pipeline; Pan-Tilt-Zoom camera
• ISSN: 0263-2241
• DOI: 10.1016/j.measurement.2025.118567

•A target tracking visual servo Pan-Tilt-Zoom control system framework is proposed.•The framework considers an improved attention mechanism module.•The framework employs a fuzzy control strategy to adjust the camera parameters.•The proposed method can monitor the real-time dynamic tracking of the excavator. The unpredictable nature of excavator operations in natural gas pipeline areas creates critical monitoring challenges, particularly regarding insufficient feature learning in complex environments and inadequate Pan-Tilt-Zoom (PTZ) control strategies for dynamic tracking. Existing methods struggle with target occlusion and lighting variations, necessitating an integrated vision-control solution. We propose an integrated deep learning framework combining an enhanced Fairness in Detection and Re-Identification for Multi-Object Tracking (FairMOT) tracker with PTZ visual servo control. Three key innovations drive the solution: 1) A Convolutional Block Attention Module (CBAM)-augmented YOLOv5 detection backbone improves feature discrimination through channel-spatial attention, 2) A fuzzy hybrid controller enables adaptive pan-tilt adjustments using real-time centroid deviations, and 3) Dynamic zoom optimization maintains target prominence through area-proportional focal tuning. The FairMOT algorithm uses the You Only Look Once (YOLO) v5 target detection model, compared with the improved YOLOv5m recognition model, the precision, accuracy, recall rate, and F1 score of the four indicators reach 85.3 %, 86.9 %, 88.6 %, and 82.2 %, respectively. Finally, a target tracking visual servo PTZ control system is designed, effectively integrating computer vision and mechanical control. This research advances intelligent pipeline monitoring by establishing a closed-loop vision-control integration framework. Practical deployments in provincial pipeline networks confirm the solution’s robustness against environmental interference, providing a scalable approach for critical infrastructure protection.