ACLGuard: Physics-Aware Knee Loading Monitoring System for Anterior Cruciate Ligament Injury Prevention Training

Anterior cruciate ligament (ACL) injuries are common in sports and significantly affect athletes' health and performance. Integrating knee adduction moment (KAM) biofeedback into ACL injury prevention training has been shown to effectively reduce injury risk and enable athletes to safely engage in high-risk activities. However, current motion capture-based monitoring methods are impractical for on-field use due to their bulky setups and limited coverage. While Inertial Motion Unit (IMU)-based methods address some of these issues, their poor performance during high-risk tasks limits their applicability in real-world scenarios. This paper presents ACLGuard, a novel physics-aware KAM monitoring system designed for out-of-lab ACL injury prevention training. ACLGuard utilizes a combination of continuous monitoring with a set of IMUs and a one-time body capture with RGB-D camera. We identify key limitations in existing approaches, including insufficient body information and inadequate encoding of biomechanical principles. To overcome these challenges, we introduce a one-time RGB-D registration scheme to capture comprehensive body information and develop an inverse dynamics (ID)-guided modeling algorithm to incorporate biomechanical principles into the system. However, extracting kinematic features under high-risk conditions and obtaining representative body features with respect to ID principle are challenging. Even worse, these extracted imperfect features increase the ID-guided modeling difficulty for KAM estimation. To derive meaningful physical features, we propose a hybrid deep learning model referring to motion patterns and physical priors. For ID-guided modeling, we introduce an attention-enhanced multi-task learning framework to establish hidden physical mappings from imperfect features to KAM. We collect a dataset from 10 athletes and 9 non-athlete subjects, containing four main high-risk tasks in real-world ACL injury prevention. Evaluations show that ACLGuard achieves an average root mean square error of 0.176 Nm/kg and a normalized root mean square error of 11.5% in KAM estimation, comparable to existing markerless motion capture solutions but offers an on-field monitoring potential with a significantly simpler setup.