Smoking behavior detection algorithm based on YOLOv8-MNC.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Detecting smoking behavior is challenging due to small, occluded objects like cigarette butts.
• Existing deep learning models struggle with accuracy and robustness for such tasks.

Purpose of the Study:

• To introduce a novel smoking detection algorithm, YOLOv8-MNC, to overcome limitations in current deep learning methods.
• To enhance the accuracy and robustness of detecting small, occluded objects in smoking behavior analysis.

Main Methods:

• Developed YOLOv8-MNC, building on YOLOv8 with a specialized small target detection layer.
• Incorporated NWD Loss to improve training accuracy by reducing sensitivity to minor positional deviations.
• Integrated Multi-head Self-Attention Mechanism (MHSA) for enhanced global feature learning and CARAFE for efficient up-sampling to minimize feature loss.

Main Results:

• The YOLOv8-MNC model achieved a detection accuracy of 85.887% on a customized smoking behavior dataset.
• Demonstrated a significant 5.7% increase in mean Average Precision (mAP@0.5) compared to previous algorithms.
• Showcased improved detection accuracy and model robustness for small, occluded objects.

Conclusions:

• YOLOv8-MNC offers a significant advancement in smoking behavior detection, addressing key challenges in accuracy and robustness.
• The algorithm's enhanced performance suggests potential applicability in related object detection fields.
• Future work will focus on refining the YOLOv8-MNC technique and exploring its broader applications.