M. Yamamoto1,2,3, Y. Ishige1, K. Shimatani4, M. Hasegawa4, Y. Kurita2, H. Takemura1
1Tokyo University of Science, Department of Mechanical Engineering, Faculty of Science and Technology, Noda, Japan, 2Hiroshima University, Graduate School of Engineering, Higashi-Hiroshima, Japan, 3Fukuyama Memorial Hospital, Department of Rehabilitation, Fukuyama, Japan, 4Prefectural University of Hiroshima, Faculty of Health and Welfare, Mihara, Japan
Background: Gait analysis is a crucial assessment in clinical site. Motion capture (MC) systems are useful for assessment because they can be used to accurately measure body kinematics such as joint angles. However, their use is mainly limited to laboratories because of high costs and low portability. Recently, many video-based pose estimation systems have been developed, which are also used to assess spatiotemporal gait parameters. However, the accuracy of joint angle estimation by video-based pose estimation systems remains unclear.
Purpose: We aimed to investigate the joint angle accuracy of 2-dimensional pose estimation during gait by single side camera.
Methods: Seven healthy adults (22.8 ± 1.1 years) participated in this study. Two treadmill walking conditions were performed by the participants: normal and external rotation gait conditions. The participants wore ankle foot orthoses, and the gait speed was set to 0.60 m/s, assuming a clinical situation. A video camera placed on the right side of the treadmill recorded gait for pose estimation. OpenPose (OP), an open source human pose system, was used to estimate body pose. The OP estimates joint locations by two branch multi-stage convolution neural network. We calculated the right hip, knee, and ankle joint angles during gait based on OP joint locations data. Simultaneously, an MC system was used to measure joint angles. The frame rate of the camera and MC were set at 100 Hz. The coefficient of determination (R2) was used to compare the OP and MC joint angles. In addition, the absolute error between the OP and MC peak joint angles during gait was also measured. The Wilcoxon signed-rank test was used to compare the absolute error in each gait condition. Statistical significance was set at p < 0.05.
Results: The R2 between the OP and MC joint angles in the normal gait condition were as follows: hip flexion-extension, 0.94; knee flexion-extension, 0.96; and ankle dorsiflexion-plantarflexion, 0.56. The R2 between the OP and MC joint angles in the external rotation gait condition were 0.93, 0.97, and 0.21, respectively. The absolute error between the OP and MC peak plantarflexion angles in the external rotation gait condition increased significantly compared with normal condition (p < 0.05). No statistically significant difference was found in the absolute error of peak hip flexion, hip extension, knee flexion, knee extension, and dorsiflexion angle (p = 0.81, p = 0.81, p = 0.16, p = 0.21, and p = 0.47, respectively).
Conclusion(s): Although the accuracy of ankle joint angle measurements using 2-dimensional pose estimation system is not high, hip and knee joint angles are measured with approximately the same accuracy as those measured using the MC system. Peak hip and knee joint angle measurements using OP with calculation methods may be used to assess peak joint angles, even in the external rotation gait condition.
Implications: This method has great potential as a low-cost and portable device to assess gait kinematics.
Funding, acknowledgements: This work was supported in part by JSPS KAKENHI under Grant number JP19K20748.
Keywords: Pose estimation, Gait analysis, Kinematics
Topic: Disability & rehabilitation
Did this work require ethics approval? Yes
Institution: Prefectural University of Hiroshima
Committee: the ethics committee of the Faculty of Health and Welfare
Ethics number: 15MH036
All authors, affiliations and abstracts have been published as submitted.
