Dilated spatial–temporal convolutional auto-encoders for human fall detection in surveillance videos

Abstract

Although methods based on supervised learning have demonstrated remarkable performance on fall detection, these existing fall detection algorithms require a substantial quantity of manually labeled training data. In this paper, we combine dilated convolution and LSTM based on auto-encoder, which can be trained on unlabeled data, further saving time and resources, and a novel fall score is computed based on the high-quality reconstructed frame to detect falls. Extensive experimental results indicate that the proposed method further boosts the performance, achieving recognition rate of 97.1%, sensitivity rate of 93.9% and precision rate of 95.1% on the UR dataset.