Transfer learning approach for fall detection with the FARSEEING real-world dataset and simulated falls

Area of Science:

• Biomedical Engineering
• Computer Science

Background:

• Falls are rare in natural settings, complicating data acquisition for fall detection systems.
• Existing systems often rely on simulated fall data, but their real-world validation is uncertain.
• Real-world fall datasets are typically imbalanced due to the sporadic nature of falls.

Purpose of the Study:

• To investigate the effectiveness of transfer learning by combining simulated and real-world fall datasets.
• To evaluate methods for handling imbalanced datasets in fall detection.
• To improve the accuracy and generalizability of fall detection classifiers.

Main Methods:

• Utilized a transfer learning approach to combine simulated fall/non-fall data with the real-world FARSEEING dataset.
• Employed imbalance learning techniques: SMOTE, Balance Cascade, and Ranking models.
• Trained supervised classifiers to discriminate between fall and non-fall events.

Main Results:

• The Balance Cascade method demonstrated fewer misclassifications on the validation set.
• Mixing real falls with simulated non-falls improved model performance compared to using only simulated falls.
• Models trained with simulated falls generalized better to real-world fall data than vice-versa.
• Overall accuracy exceeding 95% was achieved when combining datasets.

Conclusions:

• Combining simulated and real fall data enhances fall detection model performance.
• Transfer learning and appropriate imbalance handling are crucial for robust fall detection systems.
• Models trained on mixed datasets show superior generalization capabilities for real-world fall events.