Estimation of Gait Kinematics and Kinetics from Inertial Sensor Data Using Ooptimal Control of Musculoskeletal Models

Eva Dorschky, Friedrich-Alexander University of Erlangen-Nürnberg (FAU)
Marlies Nitschke, Friedrich-Alexander University of Erlangen-Nürnberg (FAU)
Ann-Kristin Seifer, Friedrich-Alexander University of Erlangen-Nürnberg (FAU)
Antonie J. van den Bogert, Cleveland State University
Bjoern M. Eskofier, Friedrich-Alexander University of Erlangen-Nürnberg (FAU)

This research was supported by the Bavarian Ministry of Economic Affairs, Regional Development and Energy within the Embedded Systems Initiative. Bjoern Eskofier gratefully acknowledges the support of the German Research Foundation (DFG) within the framework of the Heisenberg professorship program (Grant No. ES 434/8-1).


Inertial sensing enables field studies of human movement and ambulant assessment of patients. However, the challenge is to obtain a comprehensive analysis from low-quality data and sparse measurements. In this paper, we present a method to estimate gait kinematics and kinetics directly from raw inertial sensor data performing a single dynamic optimization. We formulated an optimal control problem to track accelerometer and gyroscope data with a planar musculoskeletal model. In addition, we minimized muscular effort to ensure a unique solution and to prevent the model from tracking noisy measurements too closely. For evaluation, we recorded data of ten subjects walking and running at six different speeds using seven inertial measurement units (IMUs). Results were compared to a conventional analysis using optical motion capture and a force plate. High correlations were achieved for gait kinematics (ρ⩾0.93" role="presentation">) and kinetics (ρ⩾0.90" role="presentation">). In contrast to existing IMU processing methods, a dynamically consistent simulation was obtained and we were able to estimate running kinetics. Besides kinematics and kinetics, further metrics such as muscle activations and metabolic cost can be directly obtained from simulated model movements. In summary, the method is insensitive to sensor noise and drift and provides a detailed analysis solely based on inertial sensor data.