Lightweight and Efficient Kinetics Estimation With Sparse IMUs via Multimodal Sensor Distillation

Gait kinetics, including joint moments and ground reaction forces (GRFs), are derived from optical motion capture and force plate data. However, these systems are costly, time-consuming to process, and location-restricted. Biomechanical and deep learning models using wearables such as inertial measurement units (IMUs) can estimate these parameters. However, numerous sensors complicate positioning, attachment, and calibration while reducing user comfort. Although sparse IMU sets with flexible placement are more practical, they lack critical information available in full sensor configurations, reducing performance. Furthermore, existing deep learning models for kinetics estimation are either computationally expensive or exhibit suboptimal performance due to generic architectures, making them unsuitable for compact, low-cost devices. To address these limitations, we propose Kinetics-MFFM-Net, a lightweight model incorporating two novel components: (i) a gated dual encoder with a multi-modal fusion module, and (ii) a sensor distillation (SD) framework that transfers knowledge from full-sensor supervision (e.g., full set of IMUs with sEMG or joint positions) to sparse-IMU configurations. We validate the proposed approach on two publicly available gait datasets: a lower-limb IMU–sEMG dataset spanning treadmill, stairs, ramps, and level-ground ambulation, and a dataset combining IMUs with dual-camera recordings under diverse treadmill walking conditions. Kinetics-MFFM-Net+ SD achieves up to ∼30% lower NRMSE compared to state-of-the-art models with similar or smaller parameter counts. In contrast, larger baseline models achieve comparable accuracy at higher computational cost, requiring up to ∼130× more parameters. Overall, Kinetics-MFFM-Net+SD demonstrates an effective accuracy–efficiency trade-off for gait kinetics estimation, facilitating reliable performance even with sparse IMU configurations.