Criterion validity of neural networks to assess lower limb motion during cycling

The use of marker-less methods to automatically obtain kinematics of movement is expanding but validity to high-velocity tasks such as cycling with the presence of the bicycle on the field of view is needed when standard video footage is obtained. The purpose of this study was to assess if pre-trained neural networks are valid for calculations of lower limb joint kinematics during cycling. Motion of twenty-six cyclists pedalling on a cycle trainer was captured by a video camera capturing frames from the sagittal plane whilst reflective markers were attached to their lower limb. The marker-tracking method was compared to two established deep learning-based approaches (Microsoft Research Asia-MSRA and OpenPose) to estimate hip, knee and ankle joint angles. Poor to moderate agreement was found for both methods, with OpenPose differing from the criterion by 4-8 degrees for the hip and knee joints. Larger errors were observed for the ankle joint (15-22 degrees) but no significant differences between methods throughout the crank cycle when assessed using Statistical Parametric Mapping were observed for any of the joints. OpenPose presented stronger agreement with marker-tracking (criterion) than the MSRA for the hip and knee joints but resulted in poor agreement for the ankle joint.