Multi-Purpose Security Camera for E-Cigarette Smoke Detection with Semantic Segmentation

Smoke detection with alarm sensors is often less effective because it has a limited range and cannot distinguish cigarette smoke, vaping, or other types of smoke. At the same time, semantic segmentation offers a more comprehensive solution for detecting cigarette and e-cigarette smoke. However, detecting e-cigarette smoke specifically using semantic segmentation presents a research opportunity. To address this issue, this study proposes a new approach for detecting cigarette and e-cigarette smoke based on the shape of the smoke using semantic segmentation with the lightweight DeepLabv3 architecture and MobileNetV2 backbone. First, we downloaded the smoke semantic segmentation dataset from Kaggle called “GOES16 Wildfires Smoke Plumes Dataset”. Then, we designed and tested the semantic segmentation with the DeepLabv3 model and compared J-statistics and G-mean optimization. In the next step, we developed a multi-purpose security camera design that can perform surveillance while detecting e-cigarette smoke in a certain area. Finally, we tested our semantic segmentation model on several e-cigarette smoke images and several images that did not contain e-cigarette smoke. We use several metrics, such as sensitivity, specificity, dice, and pixel accuracy. The test results show that the ROC from the training results of the semantic segmentation model on the smoke dataset has an AUC of 0.90. J-statistics and G-mean produce two different optimum threshold values, 0.52 and 0.55, respectively. It turns out that the optimized model with a threshold from G-mean has better sensitivity, dice, and pixel accuracy values than J-statistics, with values of 0.80, 0.85, 0.86, and 0.77, respectively.