DAHD-YOLO: A New High Robustness and Real-Time Method for Smoking Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are increasingly used for human behavior recognition.
• Existing object detection models for smoking behavior suffer from low accuracy and poor real-time performance, especially in complex settings.

Purpose of the Study:

• To develop an advanced deep learning model for precise and real-time smoking behavior detection.
• To address the limitations of existing models in terms of accuracy and efficiency.

Main Methods:

• Introduced DAHD-YOLO, a novel model based on YOLOv8.
• Incorporated the DBCA module for enhanced feature extraction and adaptive fine-grained channel attention (AFGCA) for better feature fusion.
• Utilized an improved feature pyramid network (ECA-FPN) and a decoupled detection head with Wise-PIoU loss for bounding box regression.

Main Results:

• Achieved superior results on a self-constructed smoking detection dataset compared to existing models.
• Reduced computational complexity by 23.20% and model parameters by 33.95%.
• Increased mAP50 by 5.1% to 86.0% and achieved 50.2 fps detection rate on RK3588 after optimizations.

Conclusions:

• DAHD-YOLO significantly improves smoking behavior detection accuracy and real-time performance.
• The model's efficiency and effectiveness make it suitable for practical applications requiring precise and instantaneous identification of smoking activities.