Detection of Gait-Related Activities from Accelerometer Signals Obtained from the Ankle

Accurate activity monitoring systems are essential to differentiate between everyday activities and events that could potentially endanger physical safety. This work analyzes the performance in terms of effectiveness of Deep Learning models, specifically Dense Neural Networks and Recurrent Neural Networks models, in the identification of different activities collected in the AnkFall dataset. This dataset contains different gait-related scenarios, including daily living activities, risk events and falls simulations, recorded with a triaxial accelerometer placed on the ankle. Different classification problems were considered, using different groupings of the activities contemplated in the dataset. To train the models, data augmentation techniques were employed to create modified versions of the original signals without causing excessive distortion. The results indicate difficulty in distinguishing between similar daily activities with minimal leg movement, as well as in recognizing fall risks in the event of a trip. However, promising performance was observed in distinguishing when the user walks while dizzy, which increases the risk of losing balance and falling.