A Transfer Learning Approach for an Automatic Fall Detection System Based on the Combined Use of a Smartwatch and a Smartphone

Fall is a critical event (mainly for the elderly and dependent subjects) that requires prompt assistance. In the past, many commercial wearable sensors have been used to detect falls, with limitations in terms of cost and usability. Consequently, there is still a growing interest in the scientific community studying issues related to active and healthy aging. The integration of fall detection functionality into commercial smartwatches represents a solution of clear advantage. It is generally related to the application of threshold values to the accelerometer signals acquired by the wearable device but with important limitations in terms of accuracy. More recently, Machine Learning and Deep Learning techniques were widely investigated, but again, adequate classification results were not obtained mainly due to the lack of specific fall datasets and the imbalance within them of events to be classified. The proposed work describes the design and implementation of a fall detection system based on a smartwatch and a smartphone. Firstly, we developed an application that collects the raw acceleration sensor data acquired from the smartwatch and sends them to the smartphone through a Bluetooth connection. Then, we apply a Transfer Learning approach to scalogram images generated, using the Continuous Wavelet Transform, from the raw accelerometric signals. In this way, we try to overcome the small dataset training problem for fall detection, especially for elderly subjects. Preliminary results obtained on a benchmark dataset show the effectiveness of our proposed pipeline using DenseNet201 deep architecture for Transfer Learning, reaching an accuracy of about 97.6%.