| dc.description.abstract | Falls represent a significant risk to the elderly population, often leading to severe injuries or fatalities.
Automatic fall detection systems (FDS) are critical for mitigating these risks; however,
existing solutions, despite reporting accuracies in controlled environments, often fail to generalize
to real-world conditions. This performance gap stems from limitations in existing datasets,
overfitted models, and a lack of standardization. To address these challenges, this thesis presents
a comprehensive framework for fall detection, leveraging privacy-preserving thermal imaging to
develop deployable, real-world solutions.
The research is conducted in three progressive phases. The first phase explores a novel hybrid
architecture that combines supervised and unsupervised learning paradigms through a stacking ensemble
of 3D Convolutional Neural Networks (3D CNNs) and Autoencoders (AEs). This hybrid
approach demonstrates significant performance improvements on constrained datasets, highlighting
its potential in scenarios where fully supervised methods fall short. Ablation studies validate
the architecture’s utility while underscoring the critical need for a more robust dataset to achieve
true generalizability.
In the second phase, the thesis introduces Thermal Fall 66 (TF-66), the most diverse and comprehensive
thermal fall detection dataset to date. Designed to address the limitations identified
in Phase 1, TF-66 encompasses varied environments, participant demographics, and fall scenarios.
Accompanied by targeted subsets and flexible data generators, TF-66 serves as a benchmark
for meaningful comparisons and standardized evaluations, advancing the field toward real-world
applicability.
The third phase synthesizes insights from the hybrid approach and TF-66 dataset to refine a
supervised 3D CNN model. Enhanced with innovative features such as optical flow integration and
attention mechanisms, this model achieves state-of-the-art performance on TF-66 and the widely
used Thermal Simulated Fall (TSF) dataset. By bridging the gap between lab-optimized systems
and real-world demands, this final phase establishes a transformative approach to fall detection,
redefining the state of fall detection research, with a focus on generalizable systems that can operate
in real-time. The findings provide a clear path for developing accurate, privacy-preserving, and
scalable fall detection systems, ultimately aiming to enhance the safety and save lives of seniors
worldwide. | en_US |
