Keywords: Human activity recognition, deep learning, self-supervised learning
 

"Feature Representation Learning for Human Activity Recognition"

Human movements and activities are crucial elements of nonverbal communication. They transmit rich information providing meanings and context explaining human behavior. Besides, our movements can serve as valuable markers for identifying hidden information, such as motor symptoms associated with certain diseases. The use cases for AI-based Human Activity Recognition (HAR) modules can be found in ambient assisted living, smart manufacturing, sports analytics, and surveillance systems. In this work, the emphasis is placed on addressing the challenges of recognizing human activities from various sensors (wearable devices, skeletal joints obtained through cameras) faced by contemporary AI models, namely Neural Networks. In particular, the work aims to provide a set of techniques that can be used to effectively encode multimodal sensorial data into compact and homogeneous representations. Exploring Self-Supervised Learning (SSL) for HAR, which relies on unlabeled data to build deep feature representations, is a prominent aspect of this work. Furthermore, in this dissertation, novel frameworks that address the limitations of existing baselines are compared with state-of-the-art methods, and evaluated in different scenarios.

Click here for the live stream.