Abstract:

Smoking detection has become an important measure for smoking bans in public places. Video image-based smoking action recognition has been widely used in smoking detection. Deep learning method for action recognition requires a large number of data sets to train. The accuracy and real-time of the existing smoking action recognition are weak, and smoking action recognition is for a single person. To address these problems, this paper proposes a method for recognizing smoking actions by detecting periodic actions. Through a large number of experiments, the smoking behavior is rhythmic and repetitive is found. This paper analyzes periodicity of smoking actions and formulates smoking action norms. This work uses information of human joint points to track action trajectory of the human body. Smoking action recognition is realized by detecting periodicity action. Recognition of smoking actions of multiple people is realized by tracking information of multi-person. Experimental results show that the method achieve an accuracy rate of 91%, and maintains high accuracy and robustness in various scenarios.

Key words: human joint point, periodicity, multi-person smoking action recognition

摘要：

吸烟检测已成为公共场所禁烟的重要措施，基于视频图像的吸烟动作识别已广泛用于吸烟检测中。使用深度学习的方法进行图像处理，需要大量数据集训练模型。现有的吸烟动作识别方法的准确率和实时性不够理想，且多只针对一个人进行动作识别。为解决这些问题，提出了一种通过检测周期性动作来识别多人吸烟动作的方法。在进行了大量的实验后发现吸烟行为是有节奏和周期性的，对此具体分析了吸烟行为的周期性并制定了吸烟行为规范；利用人体关节点信息，关注关节点的运动轨迹，检测运动轨迹是否符合周期性规律从而实现吸烟动作识别；同时跟踪多人关节点的信息，以实现多个人实时吸烟行为的识别。实验结果表明，该方法可以达到91%的准确率，在各种情况下都可以保持较高准确率和鲁棒性。

关键词: 人体关节点, 周期性, 多人吸烟动作识别