Self-Supervised Few-Shot Time-Series Segmentation for Activity Recognition

Extracting valuable activity segments from continuously received sensor data is a key step for many downstream applications such as activity recognition, trajectory prediction, and gesture recognition. Numerous unsupervised and supervised approaches have been proposed for activity segmentation. However, current unsupervised methods generally suffer from subject and environment-dependent problems, and supervised methods require a great many labeled data which is time-consuming and expensive to be collected. To address these issues, we propose a Self-supervised Few-shot Time-series Segmentation framework called SFTSeg, which introduces few-shot learning to conduct activity segmentation only relying on several labeled target samples. To be applicable to time-series data, we design a line-level data augmentation method to build a consistency regularization for the few-shot learning framework, which can augment limited labeled target samples to enhance generalization capacity of the model. Also, we devise a time series-specific pretext task to construct a self-supervised loss with adaptive weighting, which can adopt unlabeled target data to enable the model to learn characteristics of the target data and further improve segmentation performance. The experiments illustrated that SFTSeg achieves obvious gains compared to state-of-the-art methods.