Human Activity Recognition Using Smartphone Sensors: Focusing on Fall Detection with the UNIBA Dataset

Human Activity Recognition (HAR) identifies many techniques for recognizing daily life activities. This discipline has taken significant steps forward in recent years and has used several data sources, including smartphone sensors. Smartphones can be considered a simple but effective strategy due to their popularity across different generations and people. Some of the activities usually analyzed in HAR can be very relevant. This paper introduces a new dataset including 19 users performing three “neutral” activities (Walking, Jumping, and Sitting) and five activities “at-risk” (Falling Forward, Falling Backward, Falling Left, Falling Right, and Running). Moreover, the differentiation of types of falls has also been studied. Finally, highly effective Machine Learning Models have been considered for comparison purposes. These models have been trained and tested on accelerometer data collected through the smartphone.