Abstract: Long-time incorrect sitting posture is seriously harmful to human health. The existing computer vision-based method of judging whether the sitting posture is healthy or not mainly relies on the detection of the state of the human body itself, in disregard of the interaction between the human body and the screen, resulting in a failure to accurately detect a number of unhealthy sitting postures. We proposed a method of judging the healthiness of screen-reading sitting posture based on multi-relevant features. In consideration of the constraints imposed by the human body and the binding force between the human and screen, the method extracts the features that are strongly related to the sitting posture healthiness according to the spatial orientation of the target in a comprehensive way after detecting the human body and the screen. Subsequently, the sitting posture feature sequence is input to the convolutional neural network for analysis and classification in order to judge whether it is healthy or not. The experimental results show that the method can effectively identify a variety of unhealthy sitting behaviors during screen reading. Compared with other existing methods, this method is characterized with better recognition effects and application value.

Key words: sitting posture detection, healthiness judgment, multi-correlation features, convolutional neural network