AUTOMATED SMOKING DETECTION IN RESTRICTED ZONES: A YOLOv8-BASED APPROACH WITH GENDER IDENTIFICATION

This study investigates the application of YOLOv8 for smoking detection, utilizing a custom dataset and evaluating different optimization parameters, including optimizers AdamW, Adam, and SGD, with learning rates of 0.1, 0.01, and 0.001. The evaluation highlights that the Adam optimizer with a learning rate of 0.001 delivers the best overall performance, achieving an mAP@50-95 of 0.713, along with high precision (93.1%) and recall (85.9%). In comparison, AdamW at the same learning rate shows slightly higher precision (93.9%) but marginally lower recall (86.4%), resulting in similar mAP values. On the other hand, SGD performs well in some cases but falls short overall, particularly at higher learning rates, which lead to unstable convergence and reduced accuracy. While the model exhibits robust detection capabilities in close-range, low-angle scenarios, it struggles with high-angle or long-distance conditions due to dataset diversity limitations. This research contributes to restricted-zone monitoring systems by offering insights into real-time behavioral tracking and informing future applications in public health and policy enforcement. The findings emphasize Adam's superiority in balancing precision, recall, and stability, providing an effective framework for real-time smoking detection systems. Future work should focus on diversifying and augmenting the dataset to enhance generalization across challenging conditions.