On the Multi-Modal Human Motion Analysis for Edge Computing

In the era of pervasive computing, integrating human motion analysis with edge computing systems is an effective approach to overcome challenges in healthcare, industry and intelligent video analytics. This thesis investigates multi-modal human motion analysis frameworks that leverage edge computing for real-time, privacy-preserving, and resource-efficient solutions. As a foundation for this research, the thesis presents a preliminary edge-based platform for real-time human pose estimation. This platform implements three-dimensional human pose estimation (3D HPE) on resource-constraints computing devices, achieving real-time performance and privacy compliance, while maintaining high accuracy levels. The work compares the 3D HPE platform to an infrared marker-based motion capture system, demonstrating its applicability to remote and portable gait analysis in clinical applications. In addition, the thesis introduces a novel sensor fusion approach that combines HPE data from cameras with Inertial Measurement Unit (IMU) data from wearable devices. This fusion employs geometric and deep-learning models to enhance identification and tracking of individuals in dynamic and crowded scenarios. In addition, the complementary strengths of HPE and IMU data enable precise motion analysis even in complex scenarios. To address the occlusion limitations inherent in single-camera systems, the thesis proposes a distributed real-time 3D HPE framework that leverages multiple synchronized edge devices. By deploying redundant 3D pose estimation nodes, the system mitigates occlusions and aggregates pose information through a central aggregator node. Advanced synchronization and clustering mechanisms ensure accuracy, enabling robust pose estimation in multi-person and occluded environments. Finally, the thesis addresses the challenge of adapting deep learning models for motion analysis to different domains and dynamic environments. The thesis proposes an online active learning framework that integrates knowledge distillation and singular value decomposition to enable efficient retraining of edge-deployed models. Case studies in human pose estimation and object detection show significant improvements in model adaptability and inference efficiency on edge devices. The thesis provides an in-depth evaluation of the proposed solutions across various application domains, such as healthcare telemedicine for remote motion analysis and intelligent manufacturing lines for human-machine interaction. Results demonstrate significant advancements in accuracy, scalability, and efficiency compared to state-of-the-art approaches, setting the way for widespread adoption of edge-based human motion analysis systems.