Cross-modal contrastive pre-training between natural language and other modalities, e.g., vision and audio, has demonstrated astonishing performance and effectiveness across a diverse variety of tasks and domains. In this paper, we investigate whether such natural language supervision can be used for wearable sensor based Human Activity Recognition (HAR), and discover that--surprisingly--it performs substantially worse than standard end-to-end training and self-supervision. We identify the primary causes for this as: sensor heterogeneity and the lack of rich, diverse text descriptions of activities. To mitigate their impact, we also develop strategies and assess their effectiveness through an extensive experimental evaluation. These strategies lead to significant increases in activity recognition, bringing performance closer to supervised and self-supervised training, while also enabling the recognition of unseen activities and cross modal retrieval of videos. Overall, our work paves the way for better sensor-language learning, ultimately leading to the development of foundational models for HAR using wearables.

Anticipating future actions is a highly challenging task due to the diversity and scale of potential future actions; yet, additional and complementary information from different modalities help narrow down plausible action choices. Going beyond typical sources such as video and audio, we primarily explore how text descriptions of actions and objects leads to more accurate action anticipation, as they provide additional contextual cues, e.g., about the environment and its contents. We propose Multi-modal Contrastive Anticipative Transformer (M-CAT), which is trained in two stages, where the model first learns to align video and other modalities with descriptions of future actions, and is subsequently fine-tuned to predict future actions. Through extensive experimental evaluation, we demonstrate that M-CAT outperforms baselines on the EpicKitchens datasets, and show that explicit incorporation of object and action information via their text descriptions leads to more effective action anticipation. Code available at https://github.com/ApoorvaBeedu/M-CAT.

The ubiquitous availability of smartphones and smartwatches with integrated inertial measurement units (IMUs) enables straightforward capturing of human activities through collecting movement data. For specific applications of sensor-based human activity recognition (HAR), however, logistical challenges and burgeoning costs render especially the ground-truth annotation of such data a difficult endeavor, resulting in limited scale and diversity of datasets available for deriving effective HAR systems and less than ideal recognition capabilities. Transfer learning, i.e., leveraging publicly available labeled datasets to first learn useful representations that can then be fine-tuned using limited amounts of labeled data from a target domain, can alleviate some of the performance issues of contemporary HAR systems. Yet they can fail when the differences between source and target conditions are too large and/or only few samples from a target application domain are available—each of which are typical challenges in real-world human activity recognition scenarios. In this article, we present an approach for economic use of publicly available labeled HAR datasets for effective transfer learning. We introduce a novel transfer learning framework—Cross-Domain HAR—which follows the teacher-student self-training paradigm to more effectively recognize activities with very limited label information. It bridges conceptual gaps between source and target domains, including sensor locations and type of activities. Cross-Domain HAR enables substantial performance improvements over the state-of-the-art in sensor-based HAR scenarios. Through our extensive experimental evaluation on a range of benchmark datasets we specifically demonstrate the effectiveness of our approach for practically relevant few-shot activity recognition scenarios. We also present a detailed analysis into how the individual components of our framework affect downstream performance and provide practical suggestions for using the framework in real-world applications.

The dichotomy between the challenging nature of obtaining annotations for activities, and the more straightforward nature of data collection from wearables, has resulted in significant interest in the development of techniques that utilize large quantities of unlabeled data for learning representations. Contrastive Predictive Coding (CPC) is one such method, learning effective representations by leveraging properties of time-series data to setup a contrastive future timestep prediction task. In this work, we propose enhancements to CPC, by systematically investigating the encoder architecture, the aggregator network, and the future timestep prediction, resulting in a fully con-volutional architecture. Across sensor positions and activities, our method shows substantial improvements on four of six target datasets, demonstrating its ability to empower a wide range of application scenarios. Further, in the presence of very limited labeled data, our technique significantly outperforms both supervised and self-supervised baselines, positively impacting situations where collecting only a few seconds of labeled data may be possible. This is promising, as CPC does not require specialized data transformations or reconstructions for learning effective representations.

Traditionally, the sliding window based activity recognition chain (ARC) has been dominating practical applications, in which features are carefully optimized towards scenario specifics. Recently, end-to-end, deep learning methods, that do not discriminate between representation learning and classifier optimization, have become very popular also for HAR using wearables, promising "out-of-the-box" modeling with superior recognition capabilities. In this paper, we revisit and analyze specifically the role feature representations play in HAR using wearables. In a systematic exploration we evaluate eight different feature extraction methods, including conventional heuristics and recent representation learning methods, and assess their capabilities for effective activity recognition on five benchmarks. Optimized feature learning integrated into the conventional ARC leads to comparable if not better recognition results as if using end-to-end learning methods, while at the same time offering practitioners more flexibility to optimize their systems towards specifics of wearables and their constraints and limitations.