Vision-based Human Posture Classification and Fall Detection using Convolutional Neural Network

In this paper, we propose a novel computer vision based system for detecting falls on the basis of a monocular camera that can be applied for assisting independent living of an older adult living alone at home. From the recorded video data, mask regional convolutional neural network (mask R-CNN) is first applied to detect and extract the human silhouettes. The low-dimensionality representative features of extracted silhouettes are then extracted by convolutional neural network (CNN), which is applied for classification of different human postures (e.g., sit, bend, stand and lie) and a detection of a fall event (i.e., lying posture detected on the ground). For performance comparison, several machine learning algorithms such as support vector machine (SVM), random forest (RF), k-means clustering and artificial neural network (ANN) are trained with the same dataset. The proposed framework proved to be stable in human posture detection and achieved an overall accuracy of 97.5% and F-measure of 97% with small false alarms. On the same dataset the accuracies of other competing models, i.e., ANN, SVM, RF, and k-means are; 96.56%, 86.95%, 91.92% and 40.06%, respectively.