Sitting Posture Recognition Based on the Computer's Camera

The popularity of computers has made sitting a normal part of people's lives. Since unhealthy sitting postures not only harm the eyes but also increase the risk of various musculoskeletal diseases, sitting posture detecting and warning is very necessary to help people develop healthy sitting habits. Thus, a cheaper and more convenient method using an ordinary camera is proposed in this paper for 8 kinds of sitting posture. Restricted by the view of camera, the positions of nose, eyes, ears and shoulders are extracted using MoveNet, as the input of a Multi-layer Perceptron (MLP) classifier. To train the classifier, the user needs to record a video for each sitting posture. In the process of recording, guidelines are provided to improve the quality of training data. For this study, a small dataset including 12 groups of videos recorded using 2 sets of table and chair was collected. The relatively good MLP with (64,64) as hidden layer sizes and 0.162378 as alpha are found by GridSearchCV, using Leave P Group(s) Out cross-validator as the cross-validation splitting strategy. Overall, the average accuracy of this method for the dataset is 95.8%, while 97.3% for the same set of table and chair and 94.1% for the different set. These results mean that is feasible to recognize the sitting postures using ordinary camera, which implies a more easily available sitting posture warning technology.