Research on a Lightweight Smoking-Behavior Detection Algorithm Based on an Improved YOLOv5

As the demand for automated smoking behavior monitoring in public places grows, this paper proposes a lightweight smoking behavior detection model for edge deployment. This method addresses the challenges of limited computational resources, small object detection, and complex background interference on edge devices. It replaces the backbone in the YOLOv5 framework with MobileNetV3-Small and embeds an Efficient Channel Attention (ECA) mechanism at three feature pyramid outputs. This enhances key feature representation with minimal added computational and parameter costs. Experiments based on the Roboflow smoking dataset show that the improved YOLOv5 model improves overall accuracy by approximately 3.5 percentage points. Additionally, it maintains low computational complexity, making it suitable for deployment on edge cameras and embedded systems.