Recognizing Seatbelt-Fastening Activity Using Wearable Sensor Technology

Abstract

Many fatal car accidents involve victims who were not wearing a seatbelt, even though systems for detecting such behavior and intervening to correct it already exist. Activity recognition using wearable sensors has been previously applied to many health-related fields with high accuracy. In this paper, activity recognition is used to generate an algorithm for real-time recognition of putting on a seatbelt, using a smartwatch. Initial data was collected from twelve participants to determine the validity of the approach. Novel features were extracted from the data and used to classify the action, with a final accuracy of 1.000 and an F-measure of 1.000 using the MultilayerPerceptron classifier using laboratory collected data. Then, an iterative real-time recognition user study was conducted to investigate classification accuracy in a naturalistic setting. The F-measure of naturalistic classification was 0.825 with MultilayerPerceptron. This work forms the basis for further studies which will aim to provide user feedback to increase seatbelt use.