Fall recognition system using feature selection and SVM: an empirical study

In this paper we are interested in improving the accuracy of recognition when a person is fallen and when he is lying on the floor. It represents some difficulties, because the same postures exist in both actions and they happen in seconds. In our proposal the visual information acquired by the Kinect device is used to evaluate if a non-intrusive system is capable of detecting falls with high precision. A set of features, which describe the human shape and its orientation are used to be selected by a Genetic Algorithm (GA) and Principal Component Analysis (PCA). The two set of features are used for detecting the postures when the person is closing to the floor and Support Vector Machine (SVM) is used as a classifier. After detecting the falling posture, the velocity and acceleration are used, in order to distinguish when the person is fallen or lying on the floor. The obtained results suggest that the success of detection of fall depends on the selected features, and the genetic algorithm is a good technique to select them, when compared with PCA. The main contribution of this paper is to improve the accuracy of the fall detection event selecting features using a GA and PCA and incorporating acceleration and speed measurement obtained only through visual means. The public database TST Fall detection data set v2 [1] was used used in the experiments.