The analysis of human motion is a crucial component of many digital health applications, including gait analysis, therapeutic training, and daily assistance. In this realm, paradigm shifts have been made by the integration of deep/machine learning with advanced sensing techniques, providing valuable insights into human health conditions through pervasive sensing informatics. However, the success of existing approaches is mostly driven by statistical fitting between input data and given labels, relying on conventional supervised learning methodologies that assume well-balanced and clean datasets with similar distributions between training and testing data. Unfortunately, this assumption rarely holds true for sensing informatics collected in real-world healthcare applications, resulting in poor generalization capabilities.

This PhD thesis investigates the limitations of existing deep learning approaches regarding pervasive human motion analysis. First of all, the impact of sensor position heterogeneity as well as the absence of ground truth data for quantitative markerless movement measurement, based on mobile depth cameras, is explored. To tackle these challenges, a self-supervised framework is proposed to leverage synthetic human models and it successfully addresses the heterogeneity between both sensor positions and real-synthetic domains.

Furthermore, for movement disorder detection, this thesis examines the impact of the noise as well as individual differences with wearable/ambient sensor data. A cascaded disentangled representation learning framework is proposed to extract clean and clinically meaningful representations from raw sensor data, for cross-subject generalization in abnormal gait recognition.

In addition, this thesis investigates two different paradigms for movement intention recognition, egocentric camera based action anticipation and EEG based motor imagery classification. The former work delves into the co-occurrence issue of the cross-subject heterogeneity and class imbalance problem associated with real-world sensor data in the specific egocentric action anticipation task. Several key strategies are proposed to learn unbiased and meaningful representations across different subjects and classes.

On the other hand, the issue of data/feature dimensionality heterogeneity with sensor data in motor imagery EEG classification is identified and a novel learning framework is proposed to address this. The proposed framework is targeted at the data-dimensionality variations, and subsequently, implicitly promotes the generalization capability when aggregating data from multiple datasets with heterogeneous dimensions.