Comparison of Wearable Sensor-based Fall Event Detection by 1-D and 2-D Convolutional Neural Networks

Monitoring of the elderly people for their safety and well-being is crucial to help them to age independently in their own homes. Ubiquitous wearable sensors such as accelerometers and gyroscope have been used for monitoring activities of daily life (ADL) and fall. Traditional machine learning algorithms are currently replaced by the emerging deep learning methods to detect fall events and classify ADL. In this paper, fall event detection abilities of two recently proposed CNN architectures for fall event detection are compared. These two architectures have different receptive fields and therefore, discover different number of features for fall event detection. Performance of 1-D and 2-D versions of each of these architectures are compared using the data from two 3-axis accelerometers (large and small dynamic range), and a 3-axis gyroscopes available in a public dataset. 1-D CNN uses raw time signal while 2-D CNN uses the spectrograms of the raw signal as input. Effect of noise on the fall event detection using small dynamic range accelerometer is also studied. It is observed that 2-D CNN provides better fall event detection than 1-D CNN for all the three sensors and fall event detection is slightly better with accelerometers than with gyroscopes.