AOW-YOLO: An efficient and lightweight model for smoking behavior detection on construction sites

Abstract

We propose AOW-YOLO, a novel model designed to address the challenge of detecting smoking behavior in complex construction site environments, which is particularly difficult due to the typically small size of objects and cluttered backgrounds. AOW-YOLO is built upon YOLO11n through the following steps. First, we develop a novel loss function, Adaptive Occlusion and Weighting IoU (AOWIoU), which dynamically optimizes sample quality gradient allocation using multiple mechanisms, thereby overcoming the reduced sensitivity of CIoU to spatial scale errors. Second, we introduce the Spatial Grouped-Pointwise (SGP) convolution module. By integrating a channel-adaptive grouping mechanism, it minimizes information loss during downsampling, alleviating issues related to low resolution and feature degradation. Finally, our SGP module replaces the original backbone in LCNet. Through pointwise convolution, this method enables cross-group information fusion and expands channel dimensions, effectively solving feature integration problems caused by excessive channel segmentation. Experimental results show that on the smoking detection dataset, the AOW-YOLO model surpasses existing lightweight models (e.g., PP-LCNet) in both mAP50 and mAP50:95 metrics, while achieving roughly 31.6% faster inference speed than YOLO11n. This approach provides new insights for designing lightweight detection models and has broad potential applications in construction site safety management.