Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (2): 327-337.DOI: 10.3778/j.issn.1673-9418.2009069

• Graphics and Image • Previous Articles     Next Articles

Algorithm for Real-Time Smoking Detection Based on Deep Learning

CHEN Ruilong, LUO Lei, CAI Zhiping, MA Wentao   

  1. College of Computer Science, National University of Defense Technology, Changsha 410073, China
  • Online:2021-02-01 Published:2021-02-01

基于深度学习的实时吸烟检测算法

陈睿龙罗磊蔡志平马文涛   

  1. 国防科技大学 计算机学院,长沙 410073

Abstract:

In public, smoking behavior not only causes pathological harm to human health, but also exerts danger of fire hazards, etc. For the health and safety considerations, this paper designs a real-time smoking detection model based on deep learning for airports, gas stations, chemical warehouses and other places where smoking is strictly prohibited. This model uses a convolutional neural network to process the input frame from the video stream captured by the webcam. Through the process of image feature extraction, feature fusion, target classification and target posi-tioning, the coordinate of the cigarette is located, and then the smoking behaviors can be found out. Common object detection algorithms are not ideal for small target objects and the detection speed needs to be improved. This paper designs a series of convolutional neural network modules to reduce the amount of model parameters and pick up the inference speed to meet real-time requirements as well as improving the accuracy of small target object (cigarette) detection. This paper also comes up with some training skills to make the model more robust. Due to the lack of relevant dataset, this paper produces a dataset related to smoking behaviors. Through comparative experiments, it is proven that the algorithm proposed in this paper has better detection effects on proposed dataset and some public datasets.

Key words: computer vision, small object detection, real-time, smoking detection, robustness

摘要:

在公共场所内吸烟,不仅对自身、他人身体健康造成潜在的危害,还存在造成火灾等现象的隐患。因此,出于健康和安全方面的考虑,为机场、加油站、化工仓库等严禁吸烟的场所,设计了一种基于深度学习的能快速发现和警告吸烟行为的检测模型。该模型使用卷积神经网络对摄像头所拍摄的视频流输入帧进行处理,经过图像特征提取、特征融合、目标分类以及目标定位等过程,定位烟头的位置,进而判断出吸烟行为。常见的目标检测算法针对小目标物体检测效果不甚理想,检测速度亦有待提高。通过设计的一系列卷积神经网络模块,不但减少了模型计算量,加快了推演速度,满足实时性要求,而且提高了小目标物体(烟头)检测准确率。此外,运用了一些模型训练的技巧,提升了模型的鲁棒性。由于缺乏现有数据集,自制了一个与吸烟行为相关的数据集。对比实验证明了提出的算法在本数据集以及一些公开数据集上有着更好的检测效果。

关键词: 计算机视觉, 微型目标检测, 实时性, 吸烟检测, 鲁棒性