基于深度学习的办公室吸烟行为检测
Office Smoking Behavior Detection Based on Deep Learning

吸烟有害健康，为优化办公环境以及保证办公室人员身心健康。随着深度学习中卷积神经网络(Convolutional Neural Network, CNN)在目标检测领域上的发展，其目标检测方法有：单阶段检测(YOLO、SSD、RetinaNet等)、双阶段检测(Fast RCNN、Faster RCNN、Cascade RCNN等)。相比于传统的手工设计特征算法，基于深度学习的方法通过学习大量标注数据来自行进行特征的学习和提取，并预测或识别出结果，基于深度神经网络的目标检测方法具有更好的特征提取能力和分类识别效果。本文设计了采用YOLO深度学习算法的办公室吸烟行为检测方法。通过网络公开数据集收集的吸烟数据集，经过对数据集的整合与调整形成最终进行实验的吸烟行为检测数据集。用吸烟行为检测数据集分别训练YOLOv5，YOLOv6，YOLOv7，YOLOx四个模型，通过对比训练产生的结果得到办公室吸烟行为检测的最佳训练模型。实验结果表明，在办公室吸烟行为检测实验中，YOLOv5为检测效果最优异的模型，其精确度均值(mAP)：76.6%，平均推理时间：17.1 ms。
Smoking is harmful to health, in order to optimize the office environment and ensure the physical and mental health of office staff. With the development of Convolutional Neural Network (CNN) in the field of target detection, the methods of Convolutional Neural Network detection include one-stage detection (Yolo, SSD, RetinaNet, etc.) and two-stage detection (Fast RCNN, Faster RCNN, Cascade RCNN, etc.). Compared with the traditional hand-designed feature algorithm, the method based on deep learning can learn and extract feature by learning a lot of labeled data, and predict or recognize the result, the object detection method based on deep neural network has better feature extraction ability and classification recognition effect. In this paper, Yolo deep learning algorithm is used to detect office smoking behavior. The smoking data set collected through the web-based open data set was integrated and adjusted to form the smoking behavior detection data set for the final experiment. Four models, Yolov5, YOLOV6, Yolov7 and Yolox, were trained with smoking behavior detection data set. The best training model of office smoking behavior detection was obtained by comparing the results of training. The results showed that YOLOV5 was the best model in the detection of office smoking behavior. The mean accuracy (mAP) was 76.6% and the mean reasoning time was 17.1 ms.