A Convolutional Neural Network for Smoking Activity Recognition

Smoking is linked to more than two million preventable deaths yearly. The widespread use of sensors embedded in everyday devices provides novel means for research on smoking. Smartphones and smartwatches can monitor smoking behavior, which could lead to the development of new methods for smoking reduction or cessation. However, smoking often co-occurs with other activities, such as drinking and eating, which makes the recognition of concurrent and overlapping smoking activities from wearable sensors challenging. In this paper, we proposed for the first time to use deep learning for the automatic detection of smoking activities. A Convolutional Neural Network (CNN) architecture was proposed, and this improved on previously reported performance results. We investigated the impact of various data preprocessing approaches that influence the CNN classification results with statistical features and raw sensor data. We also considered the individual performance of the smartwatch vs. the smartphone and the gyroscope vs. accelerometer sensors for smoking activity recognition. Considering a dataset of concurrent activities such as drinking, eating, smoking while sitting, standing, walking, and partaking in a group conversation, our CNN approach obtained an F1-score of 92-96% in person-independent evaluation.