NeuroMobi: User-Independent Human Mobility Analysis using Brain Signals

Abstract

Brain-Computer Interfaces (BCIs) have primarily focused on using brain signals for device control. We instead propose leveraging brain signals for passive sensing of human activities (e.g., mobility analysis). A fundamental obstacle to this vision is the high variability of brain signals between different users and even within the same user over time, severely degrading performance of conventional models on new, unseen individuals. This paper presents NeuroMobi, a novel deep learning framework designed specifically to learn robust, user-invariant features from brain signal data, thereby addressing the critical challenge of inter-user variability. Our approach employs a denoising autoencoder that learns to reconstruct clean electroencephalography (EEG) signals from noisy inputs simulating inter-subject variability, forcing the model to extract fundamental movement-related neural patterns. This representation learning is then fine-tuned for classification tasks (e.g., mobility analysis), enabling the model to generalize across unseen users for these tasks. We validate NeuroMobi on human mobility detection and speed estimation tasks, showing performance improvements of up to 19% in F1 score compared to state-of-the-art models when tested on unseen users, a step toward practical and user-independent brain signal-based sensing systems.