Human Activity Recognition (HAR) focuses on developing automated systems to identify andcategorize human activities. This field is dominated by two approaches: vision-based and sensor-based
HAR. Vision-based HAR utilizes deep learning algorithms to analyze video frames and interpret
human actions and gestures. A key challenge here is the temporal dependency problem, which addresses
the evolving nature of human actions over time. This thesis presents several research questions
on improving video-level representation learning, leading to the novel F4D CNN architecture.
This architecture demonstrates significant performance enhancements over existing models across five
benchmark datasets.
Onthe other hand, sensorbasedHARuseswearable sensors to directly capture humanmovementsand physiological states, generating high-quality data suited for deep learning applications. Research in this area often involves benchmark datasets to study sensor based HAR and develop advanced recognition methods. However, these benchmark datasets have their own limitations such as the limited number of classes presented. This Thesis studies the generalization of models for real-world scenarios and how they perform for less common activities. Moreover, explores the potential of using pre-trained vision-based models to help with enhancing the recognition. This will lead to the proposal of JS-Siamese architecture, a generalized zero shot learning paradigm that have the ability to recognize unseen classes with the help of vision-based models.
While benchmark datasets allowed researchers to focus more on enhancingHARrecognition, privacyissues have emerged due to the sensitive nature of data from sensors, raising several research questions. These include the potential de-anonymization of data and risks associated with data breaches. This study investigates these concerns by employing an LSTM model that was able to match a significant proportion of anonymized data samples to the correct user. To enhance privacy, this thesis introduces a differentially private GAN model (HDP-GAN) to produce synthetic data without compromising individual privacy.
In conclusion, this thesis provides an in-depth exploration of key issues in human activity recognition,outlining challenges that could affect recognition accuracy and proposing innovative solutions to address these challenges.
Onthe other hand, sensorbasedHARuseswearable sensors to directly capture humanmovementsand physiological states, generating high-quality data suited for deep learning applications. Research in this area often involves benchmark datasets to study sensor based HAR and develop advanced recognition methods. However, these benchmark datasets have their own limitations such as the limited number of classes presented. This Thesis studies the generalization of models for real-world scenarios and how they perform for less common activities. Moreover, explores the potential of using pre-trained vision-based models to help with enhancing the recognition. This will lead to the proposal of JS-Siamese architecture, a generalized zero shot learning paradigm that have the ability to recognize unseen classes with the help of vision-based models.
While benchmark datasets allowed researchers to focus more on enhancingHARrecognition, privacyissues have emerged due to the sensitive nature of data from sensors, raising several research questions. These include the potential de-anonymization of data and risks associated with data breaches. This study investigates these concerns by employing an LSTM model that was able to match a significant proportion of anonymized data samples to the correct user. To enhance privacy, this thesis introduces a differentially private GAN model (HDP-GAN) to produce synthetic data without compromising individual privacy.
In conclusion, this thesis provides an in-depth exploration of key issues in human activity recognition,outlining challenges that could affect recognition accuracy and proposing innovative solutions to address these challenges.