A novel feature fusion-based and complex contextual model for smoking detection

Detecting smoking behavior is crucial for advancing public health and preventing fire-related incidents. To address persistent issues in multi-scale feature integration, fusion efficiency, and precise identification of key objects within complex environments, this study introduces a novel and high-performance algorithm named S-YOLO. The proposed framework is designed to overcome these challenges and elevate overall detection capability. First, BiFPN is utilized to achieve effective multi-level feature integration, addressing the limitations of prior detection algorithms in feature extraction and target recognition. Next, the ECA module is incorporated to enhance the precision and efficiency of key object extraction in intricate data scenarios. Additionally, CBAM is employed to further refine feature representation and improve fusion quality. Experimental results reveal a 5.2% increase in mAP compared to the baseline model, underscoring the S-YOLO model’s superior accuracy in complex smoking detection scenarios.