ScooterID: Posture-Based Continuous User Identification From Mobility Scooter Rides

Abstract

Mobility scooters serve as a powerful last-mile transportation tool for people with mobility challenges. Given the unique riding behavior and posture of mobility scooter riders, such user-specific mobility scooter ride data has tremendous potential towards the design of continuous user identification and authentication mechanisms. However, there have been no prior research efforts in the literature exploring this unique modality for the design of continuous user identification techniques. To address this gap, this paper proposes ScooterID , the first framework which employs rider posture data collected from cameras on mobility scooters to continuously identify (and authenticate) users/riders. As part of this framework, a machine learning based model comprising of a spatio-temporal Graph Convolutional Network and a body-part-informed encoder is designed to effectively capture a user’s subtle upper-body movements during mobility scooter rides into discriminating embedding vectors. These embeddings can then be used to reliably and continuously identify and authenticate users/riders. Experiments with real-world mobility scooter ride data show that ScooterID achieves high levels of authentication accuracy with few enrollment video samples. ScooterID also performs efficiently on resource-constrained devices (e.g., Raspberry Pis) and is robust against adversarial perturbations to authentication inputs.