Fall Detection in the Real-World

In SnartFall Lab, we have experimented with various machine learning algorithms, especially the Deep Learning (DL) algorithm for learning accelerometer data sensed by smartwatches for fall detection. These fall detection models are aimed at real-world deployment on smartwatches paired with smartphones, specifically for use by older adults. DL models typically require large datasets for practical training, but large, annotated datasets are scarce in fall detection due to the rarity of fall events. Additionally, fall data is inherently noisy as wrist-worn smartwatches can mistake other motions for falls. We have performed comparative studies, including basic LSTM, ensemble LSTM techniques, and transformer-based DL approaches for fall detection using smartwatches. We are one of the few research groups that tested our fall detection model on a physical device using a carefully designed SmartFall app. Unfortunately, none of our DL-based fall detection models achieved a performance acceptable by older adults in the real world based on our small-scale user study. This highlights the limitations of using a single wrist-worn accelerometer data for fall detection in the real world. I will discuss a cross-modal learning approach leveraging Knowledge Distillation and Generative AI to address this.