| Sumario: | Several open-source platforms for marker-less motion capture offer the ability to develop deep learning models for 2-dimensional (2D) pose estimation. These models can then be used to track 2D kinematics using simple digital cameras. However, very few studies have examined the accuracy of these platforms for tracking human motion. PURPOSE: To establish the accuracy of a marker-less motion capture platform, compared to manual digitization, for 2D foot and tibia segment angles during running METHODS: 60 runners who had presented to a running clinic and had sagittal plane videos recorded of their left lower leg (125 fps, res= 644x486) participated. 50 participants were used to train a deep learning model for 2D pose estimation of the foot and tibia segments (final training set= 1,840 frames, training iterations= 200,000). The trained model was used to process novel videos from 10 participants for continuous 2D coordinate data. Foot and tibia angles at initial contact were manually digitized for 5 strides per participant using a video-processing software. Marker-less angles for the same strides were calculated from 2D coordinate data. Independent t-tests were used to compare mean angles between methods. Absolute and relative errors for marker-less angles were calculated for each stride. Mean absolute error was used to assess the accuracy of the marker-less method. Bland-Altman plots, using relative errors, were used to assess systematic bias. RESULTS: There were no significant differences in mean angles between measurement methods (Figure 1; p= 0.19-0.54). Mean absolute errors were 0.83° (0.44-1.22) and 0.80° (0.40-1.20) for the foot and tibia. Bland-Altman plots (omitted) did not reveal the presence of systematic bias. CONCLUSION: These results demonstrate excellent accuracy for the marker-less method for estimating 2D kinematics, reducing the burden and potential inter-rater/session errors with digitization.
|
|---|