Automatic Counting of Chewing Actions in Eating Videos
Jun-Min Kim, Yuhyun Kim, Yumin Lee, Jeongwha Chun, Daehwan Kim · 2024
This paper proposes a method for automatically calculating the number of chewing actions from meal videos. The method uses a chewing cycle graph, which represents the periodic changes in the landmark positions of the nose and jaw. Chewing actions are detected by identifying local maxima in the graph. However, accurately detecting chewing actions can be challenging due to noise, which can obscure the identification of local maxima in the graph. A recent study attempted to reduce noise by eliminating maxima below a certain threshold, but this approach was limited in its effectiveness. In contrast, this study presents an improved method that enhances noise reduction and peak detection accuracy, leading to more precise identification of true local maxima. Specifically, a Savitzky-Golay filter is applied to smooth the graph, and the analysis includes a time constraint on the chewing cycle period. This approach reduces the error rate in counting chewing actions and enhances the detection of subtle movements.