Lying human activity recognition based on shape characteristics

This paper proposes a markerless video analytic system for quantifying body parts movement while lying. These movements include: hand, leg, both hand & leg and turning to left or right movements. Combination of pixel intensity and area difference of both segmented and the whole parts of each silhouette compared with the following silhouettes would provide a useful cue for detection of different body parts movement while lying. Extracted feature vectors after applying PCA algorithm for dimension reduction are finally fed to a multiclass support vector machine for precise classification of proposed movements. Unlike most of the existent human action detection systems that only deal with human movements while standing, we have considered movements that a person does while lying, which has a wide range of application in sport and medical science. Reliable recognition rate of experimental results underlies satisfactory performance of our system.