YOLO-GL for efficient multi-target detection in large FOV via hierarchical feature fusion

Area of Science:

• Computer Vision
• Artificial Intelligence
• Industrial Safety Engineering

Background:

• Industrial construction sites present complex safety inspection challenges, including large-scale variations and multi-object detection difficulties.
• Accurate detection of critical safety indicators such as flames, smoke, personnel attire, and operational behaviors is crucial but often limited by current technologies.

Purpose of the Study:

• To propose YOLO-GL, an enhanced object detection network designed to overcome the limitations of current safety inspection methods in industrial environments.
• To improve the accuracy and robustness of detecting critical safety indicators in complex industrial settings.

Main Methods:

• Development of YOLO-GL, featuring a Parallelized Local-Global Multi-Level Fusion Module (C2f_gl) with attention mechanisms for multi-scale feature representation.
• Implementation of a hierarchical feature fusion architecture and an adaptive feature fusion shuffling module (MSF&ACS) for dynamic cross-scale semantic optimization.
• Extensive experimentation using composite datasets combining public benchmarks and industrial site collections.

Main Results:

• YOLO-GL achieved state-of-the-art performance in flame detection, with an increase in mAP@0.5 by 3.5% (70.8% to 74.3%) and mAP@0.5:0.95 by 2.9% (38.7% to 41.6%).
• The model maintains real-time processing capabilities at 80.59 frames per second (FPS).
• Demonstrated superior robustness in complex industrial environments compared to existing methods.

Conclusions:

• YOLO-GL offers a significant advancement in object detection for industrial safety monitoring.
• The proposed architecture provides an effective and robust solution for real-time safety inspections on construction sites.
• The innovations in feature fusion and attention mechanisms contribute to improved performance in challenging scenarios.