Vision-based Fall Detection with Multi-task Hourglass Convolutional Auto-encoder

Fall detection is a hot research issue in intelligent video surveillance. Falls can generate physical and psychological damage, especially for the elderly. Different from most conventional vision-based fall detection methods typically relying on hand-crafted features, fall detection methods based on deep learning techniques can automatically learn features and hence have got widespread concern recently. However, as deep networks are increasingly applied to fall detection, the problem of information loss in the deep networks can not be ignored, because this will ultimately affect the performance of fall detection. To solve the above problem, we propose a vision-based fall detection method using multi-task hourglass convolutional auto-encoder (HCAE). In this method, hourglass residual units (HRUs) are introduced into the encoder of the HCAE to extract multiscale features by expanding receptive fields of neurons. A multi-task mechanism is presented to enhance the feature representativeness of the network by completing an auxiliary task of frame reconstruction while realizing the main task of fall detection. Experimental results demonstrate that, the proposed method can effectively achieve accurate fall detection with the shallow-layer network, and outperforms several state-of-the-art methods.