BACKGROUND

Falls pose a serious health risk for the elderly, particular for those who are living alone. The utilization of WiFi-based fall detection, employing Channel State Information (CSI), emerges as a promising solution due to its non-intrusive nature and privacy preservation. Despite these advantages, the challenge lies in optimizing cross-individual performance for CSI-based methods.

OBJECTIVE

This study aimed to develop a resilient real-time fall detection system across individuals utilizing CSI, named TCS-Fall. This method was designed to offer continuous monitoring of activities over an extended timeframe, ensuring accurate and prompt detection of falls.

METHODS

Extensive CSI data on 1800 falls and 2400 daily activities was collected from 20 volunteers. The grouped coefficient of variation of CSI amplitudes were utilized as input features. These features capture signal fluctuations and are input to a convolutional neural network classifier. Cross-individual performance was extensively evaluated using various train/test participant splits. Additionally, a user-friendly CSI data collection and detection tool was developed using PyQT. To achieve real-time performance, data parsing and pre-processing computations were optimized using Numba's just-in-time compilation.

RESULTS

The proposed TCS-Fall method achieved excellent performance in cross-individual fall detection. On the test set, AUC reached 0.999, no error warning ratio score reached 0. 955 and correct warning ratio score reached of 0.975 when trained with data from only two volunteers. Performance can be further improved to 1.00 when 10 volunteers were included in training data. The optimized data parsing/pre-processing achieved over 20× speedup compared to previous method. The PyQT tool parsed and detected the fall within 100 ms.

CONCLUSIONS

TCS-Fall method enables excellent real-time cross-individual fall detection utilizing WiFi CSI, promising swift alerts and timely assistance to elderly. Additionally, the optimized data processing led to a significant speedup. These results highlight the potential of our approach in enhancing real-time fall detection systems.