Validity of Neural Networks to Determine Body Position on the Bicycle

Abstract

With the increased access to neural networks trained to estimate body segments from images and videos, this study assessed the validity of some of these networks in enabling the assessment of body position on the bicycle. Fourteen cyclists pedalled stationarily in one session on their own bicycles whilst video was recorded from their sagittal plane. Reflective markers attached to key bony landmarks were used to manually digitise joint angles at two positions of the crank (3 o'clock and 6 o'clock) extracted from the videos (Reference method). These angles were compared to measurements taken from videos generated by two deep learning-based approaches designed to automatically estimate human joints (Microsoft Research Asia-MSRA and OpenPose). Mean bias for OpenPose ranged between 0.03-1.81 whilst the MSRA method presented errors between 2.29-12.15. Correlation coefficients were stronger for OpenPose than for the MSRA method in relation to the Reference method for the torso (r = 0.94 vs. 0.92), hip (r = 0.69 vs. 0.60), knee (r = 0.80 vs. 0.71) and ankle (r = 0.23 vs. 0.

Validity of neural networks in determining lower limb kinematics in stationary cycling

“…In prior studies involving walking gait, marker-less methods presented differences between < 1° [9] and 6° [10], which is comparable to ndings from the current study for the hip and knee joints. For cycling, Bini et al observed 3-12° of difference between the MSRA and the reference data [11], which suggests that TransPose and MediaPipe may perform better than the MSRA. It is also important to highlight that these methods seem to perform well when tracking the hip and knee joints but struggled to track foot markers.…”

“…Even though studies have explored the validity of marker-less methods in determining joint angles [9,10], only one study explored the validity of a pre-trained neural network for cycling movement [11]. This preliminary data suggests that a popular convolutional neural network (CNN) method for pose estimation proposed by Microsoft Research Asia [12] produced errors between 3-12° in relation to a criterion measure [11]. These errors would be potentially larger than the range proposed to determine body position on the bicycle [i.e.…”