Lightweight detection model for safe wear at worksites using GPD-YOLOv8 algorithm

Area of Science:

• Computer Vision
• Artificial Intelligence
• Occupational Safety

Background:

• Construction sites have high safety risks due to improper use of personal protective equipment (PPE).
• Accurate and efficient detection of safety wear like helmets and reflective clothing is crucial for worker safety.

Purpose of the Study:

• To develop an enhanced YOLOv8 model for improved safety wear detection on construction sites.
• To increase the accuracy and speed of detecting helmets and reflective clothing.

Main Methods:

• Integration of a P2 detection layer to enrich semantic feature representation.
• Incorporation of a lightweight Ghost module to reduce computational load.
• Addition of a Dynamic Head (Dyhead) with an attention mechanism for feature extraction.
• Adoption of an Exponential Moving Average (EMA) SlideLoss function for enhanced accuracy and stability.

Main Results:

• The proposed model achieved a 6.2% improvement in mean Average Precision (mAP) over the baseline YOLOv8.
• Detection speed increased by 55.88% in frames per second (FPS).
• The enhanced model demonstrated superior performance in detecting safety wear.

Conclusions:

• The enhanced YOLOv8 model effectively addresses safety risks by accurately detecting safety wear.
• The model offers a significant improvement in both detection accuracy and speed for construction site safety applications.