Cross-modal learning from visual information for activity recognition on inertial sensors

Thesis

The lack of large-scale, labeled datasets impedes progress in developing robust and generalized predictive models for human activity recognition (HAR) from wearable inertial sensor data. Labeled data is scarce as sensor data collection is expensive, and their annotation is time-consuming and error-prone. As a result, public inertial HAR datasets are small in terms of number of subjects, activity classes, hours of recorded data, and variation in recorded environments. Machine learning models, developed using these small datasets, are effectively blind to the diverse expressions of activities performed by wide-ranging populations in the real world, and progress in wearable inertial sensing is held back by this bottleneck for activity understanding. .

But just as Internet-scale text, image and audio data have pushed their respective pattern recognition fields to systems reliable enough for everyday use, easy access to large quantities of data can push forward the field of inertial HAR, and by extension wearable sensing. To this end, this thesis pioneers the idea of exploiting the visual modality as a source domain for cross-modal learning, such that data and knowledge can be transferred across to benefit the target domain of inertial HAR. .

This thesis makes three contributions to inertial HAR through cross-modal approaches. First, to overcome the barrier of expensive inertial data collection and annotation, we contribute a novel pipeline that automatically extracts virtual accelerometer data from videos of human activities, which are readily annotated and accessible in large quantities. Second, we propose acquiring transferable representations about activities, from HAR models trained using large quantities of visual data to enrich the development of inertial HAR models. Finally, the third contribution exposes HAR models to the challenging setting of zero-shot learning; we propose mechanisms that leverage cross-modal correspondence to enable inference on previously unseen classes. .

Unlike prior approaches, this body of work pushes forward the state of the art in HAR not by exhausting resources concentrated in the inertial domain, but by exploiting an existing, resourceful, intuitive, and informative source, the visual domain. These contributions represent a new line of cross-modal thinking in inertial HAR, and suggest important future directions for inertial-based wearable sensing research.

Authors: Tong, EGC

• DOI: 10.5287/ora-4jqroa50k
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: wearable devices; machine learning; transfer learning; human activity recognition; cross-modal learning; multi-modal learning
• Subjects: Wearable technology; Transfer learning (Machine learning); Deep learning (Machine learning)