AOW-YOLO: An efficient and lightweight model for smoking behavior detection on construction sites

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Construction sites present unique challenges for object detection, including small object sizes and cluttered backgrounds.
• Detecting smoking behavior in these environments is crucial for safety management but technically difficult.

Purpose of the Study:

• To develop a novel, lightweight object detection model (AOW-YOLO) for accurately identifying smoking behavior in complex construction site settings.
• To improve detection performance and inference speed compared to existing models.

Main Methods:

• Introduced Adaptive Occlusion and Weighting IoU (AOWIoU) loss function to optimize sample quality gradient allocation.
• Developed a Spatial Grouped-Pointwise (SGP) convolution module to minimize information loss during downsampling and enhance feature integration.
• Integrated the SGP module into the LCNet backbone, replacing the original structure.

Main Results:

• AOW-YOLO demonstrated superior performance over existing lightweight models on a smoking detection dataset, achieving higher mAP50 and mAP50:95 metrics.
• The model achieved a 31.6% faster inference speed compared to YOLO11n.
• The SGP module effectively addressed feature integration issues caused by channel segmentation.

Conclusions:

• AOW-YOLO offers a promising solution for detecting smoking behavior in challenging construction environments.
• The proposed methods provide valuable insights for designing efficient and effective lightweight detection models.
• This technology has significant potential for enhancing safety management in construction sites.