A novel computer vision-based data driven modelling approach for person specific fall detection

Liyun Gong, Lu Zhang, Ming Zhu, Miao Yu, Ross Clifford, Carol Duff, Xujiong Ye, Stefanos D. Kollias

Abstract: In this paper, we propose a novel person specific fall detection system based on a monocular camera, which can be applied for assisting the independent living of an older adult living alone at home. A single camera covering the living area is used for video recordings of an elderly person’s normal daily activities. From the recorded video data, the human silhouette regions in every frame are then extracted based on the codebook background subtraction technique. Low-dimensionality representative features of extracted silhouetted are then extracted by convolutional neural network-based autoencoder (CNN-AE). Features obtained from the CNN-AE are applied to construct an one class support vector machine (OCSVM) model, which is a data driven model based on the video recordings and can be applied for fall detection. From the comprehensive experimental evaluations on different people in a real home environment, it is shown that the proposed fall detection system can successfully detect different types of falls (falls towards different orientations at different positions in a real home environment) with small false alarms.