X4Gait: Cross-Modal Supervision for Practical IMU-based Dual-Task Gait Recognition
Abstract
Gait is widely recognized as a sensitive marker of neurological health, yet most studies are limited to controlled clinical settings, reducing their applicability to real-world screening. Free-living gait analysis offers a scalable, natural, and ecologically valid approach, particularly beneficial in urban areas where healthcare access is limited. However, dual-task gait, i.e., walking while performing concurrent activities, imposes varying levels of cognitive load, subtly altering gait parameters and potentially confounding the analysis of gait patterns in free-living settings, where the dual-task context is dynamic and typically unknown. Inertial measurement units (IMUs), while ubiquitous and practical for everyday monitoring, capture limited kinematic information compared to skeletal representations from motion capture, constraining dual-task detection accuracy. To address these challenges, we introduce X4Gait, a cross-modal supervision framework that leverages a high-information (MoCap) modality to guide representation learning for IMU-based dual-task gait recognition. X4Gait transfers MoCap representations to the IMU encoder through cross-modal supervision and incorporates auxiliary objectives, including kinematic regression and identity-adversarial learning, to promote kinematics-relevant and person-independent representations. We rigorously evaluate X4Gait through a progressive experimental methodology that examines cross-subject generalization and deployment-time adaptation. Results show that cross-modal representation supervision substantially improves cross-subject generalization, increasing subject-independent accuracy by 23.1% relative to conventional label supervision. Moreover, it enables highly effective local adaptation. Specifically, adaptation using cross-modal representations improves accuracy from 58.2% to 81.8%, a 40.5% relative gain over direct deployment without local adaptation. These findings demonstrate that X4Gait enables both robust generalization and annotation-efficient personalization, providing a practical pathway toward scalable in-home dual-task gait monitoring.