Article Preview
TopIntroduction
With the continuous advancement of technology, smoking detection methods have also been continuously improved. Traditional smoking detection methods are usually detected by physical means such as smoke sensors and wearable devices. Mobile health technologies are being developed for personal lifestyle and medical healthcare support, of which a growing number are designed to assist smokers to quit (Ortis et al., 2020). However, these methods have many limitations: one is that the concentration of smoke in outdoor scenes is greatly diluted and cannot be sensed by the smoke sensor; the other is that wearable devices are expensive to perform detection and need to be owned by everyone. In addition, the movement trajectory and speed of multiple parts of the limbs are judged in this method, the pattern is match with the smoking behavior, and then the matching degree is judged through machine learning classification methods such as support vector machine (SVM). The detection accuracy and efficiency of this type method are relatively low (Senyurek et al.,2019).
In addition to using physical equipment to detect smoking, some scholars detect smoking by using traditional graphics object detection methods. This type of method is divided into three steps (Wu et al., 2010): First, different sizes and step length sliding windows are set, and then all the windows are slided in each position on the image. For each window, the feature of the object to be measured is extracted through the histogram of oriented gradient (HOG) or scale-invariant feature transform (SIFT) method, and finally the classification algorithm is used for each sliding window to perform classification, such as SVM, Adaboost, etc., and the sliding window with the highest score is selected as the detection result. However, this type of method has the following disadvantages: firstly, the detection effect is not ideal, it is easy to be interfered by other objects, and the positioning is not accurate, relying on the preset sliding window size and sliding step length; secondly, there is a large amount of calculation in this method, and it needs to perform feature processing and classification judgment for each sliding window; finally, the method and process of manually extracting features are more complicated and do not have generalization.