Non-intrusive continuous user verification by care robots: MoveNet gait data

The growing ageing population demands for advanced care technologies, such as care robots, to support quality living. Ensuring the safety and privacy of these vulnerable users necessitates reliable and user-friendly authentication methods. Security features should not become a burden to the user experiences. However, it is vital to continuously verify the user by the care robot in communicating and delivering its services. To offload the burden of verification to the users, it should be the robot initiating the verification process. However, relying on biometric data like voice doesn’t guarantee the source and needs continual verbal input while face relies on the line of sight for continuous verification making it challenging. Thus, this study examined the use of MoveNet by the care robot for continuous identity verification for an authentication process of a user, leveraging the 17 data points of the gait data collected from body joints and face features along with the distances among the data points to verify user identity. The research evaluated the performance of various MoveNet models and machine learning algorithms to identify the most effective approach for continuous user authentication in care robots. The methodology involved collecting and analysing gait data from a controlled group of participants, implementing and testing with several MoveNet models and machine learning techniques, with a particular emphasis on neural networks. The results highlighted that integrating MoveNet with neural network models, especially the Thunder and Lightning f16 variants, achieved accurate user identification with an accuracy of 99.86 % (NN), 99.89 % (CNN),99.93 % (Random Forest) and KNN gives F1 score of 99.74 %, while SVM performs the worst with only 13.23 % F1 score. These findings provide an opportunity for the robot to seamlessly verify the user for authentication purpose using machine learning methods. A neural network is tested with all the MoveNet models (lightning, lightning int8, lightning f16, thunder, thunder int8, and thunder f16) and the paper proofs its usability in a ROS system, in average, prediction time takes between 0.99 to 1.06 s with an accuracy ranging from 99.64 % to 99.90 %. Lightning f16 and Thunder are the best performing models in terms of prediction time and accuracy.