Research on recognition of indoor fall behaviors based on video monitoring

In this paper, we propose an indoor video-based feature recognition method to detect the fall behaviors of people. We firstly establish and update the video background using Gaussian mixture model, and apply background subtraction to extract the moving targets. To remove the shadow interference on these extracted moving targets, we eliminate these shadows by integrating color and gradient features. Then for the current frame, the four features of Hu's invariant moments, aspect ratio, attitude rate and velocity are used to make up a 7-dimensional feature vector. Totally, we extract such 15 frames by interval sampling per motion cycle, and form 105-dimension features to describe a behavior. Based on the feature representations, we employ support vector machines to classify six daily activities, i.e., walking, jogging, sitting down, squatting, bending, as well as falling. Numerical experiments demonstrate our method reaches an average correct recognition rate of 92%, with high sensitivities and specificities. It is able to distinguish falling with the other different behaviors and has many potential applications such as old people remote nursing.