A Data Augmentation Strategy for Skeleton-Based Fall Detection

Fall is the main cause of injuries in elderly people. With the development of monitor techniques and deep-learning, vision-based behavior recognition algorithms have a promising future in the fall detection task. Recently, graph convolution net-works (GCN) for skeleton-based action recognition have achieved significant development, thus fall detection can be realized by training such networks in fall datasets. However, it is impossible for a network to learn enough features due to the less behavior categories and the small scale of existing fall datasets. In addition, although current skeleton-based algorithms are robust in the face of complex background, multi-scale human, multi-perspective and multi-motion speed, it is still difficult to solve the occlusion problem. In this paper, a novel real fall dataset is proposed by fusion strategy, and it contains a large number of fall and fall-like behaviors. Moreover, we propose a data augmentation mechanism to simulate the case of occlusion. We create an occlusion matrix by data-driven method to get the occlusion skeleton dataset. We experimentally show that the proposed methods effectively improve detecion accuracy and anti-occlusion ability of the fall detection system.