Feasibility and validity of a single camera CNN driven musculoskeletal model for muscle force estimation during upper extremity strength exercises: Proof-of-concept

Abstract

Muscle force analysis can be essential for injury risk estimation and performance enhancement in sports like strength training. However, current methods to record muscle forces including electromyography or marker-based measurements combined with a musculoskeletal model are time-consuming and restrict the athlete's natural movement due to equipment attachment. Therefore, the feasibility and validity of a more applicable method, requiring only a single standard camera for the recordings, combined with a deep-learning model and musculoskeletal model is evaluated in the present study during upper-body strength exercises performed by five athletes. Comparison of muscle forces obtained by the single camera driven model against those obtained from a state-of-the art marker-based driven musculoskeletal model revealed strong to excellent correlations and reasonable RMSD's of 0.4-2.1% of the maximum force (Fmax) for prime movers, and weak to strong correlations with RMSD's of 0.4-0.7% Fmax for stabilizing and secondary muscles. In conclusion, a single camera deep-learning driven model is a feasible method for muscle force analysis in a strength training environment, and first validity results show reasonable accuracies, especially for prime mover muscle forces. However, it is evident that future research should investigate this method for a larger sample size and for multiple exercises.

Keywords: artificial intelligence; fitness; markerless motion capture; musculoskeletal modeling; strength training; video-based motion capture.

Figure 1

Pictures of the start (left) and mid (right) pose of the biceps curl (top) and lateral fly (bottom) exercise.

Figure 2

Workflow data collection and analysis from the marker-based optolectronic measurement system (OMS) (top) and the markerless camera (bottom) methods. The OMS consisted of 8 infrared cameras that captured the 3D trajectories of reflective markers placed on bony landmarks of the participant (1.A). Additionally, joint center trajectories were estimated from bony landmarks and added to the data (1.B), before importing in OpenSim. The generic thoracoscapular model with virtual markers (pink) that corresponded to locations of the collected bony landmarks and joint centers (experimental markers) (blue and red) was imported in OpenSim (1.C). The camera captured standard 2D videos of the participant from a frontal view (2.A). A deep-learning model (8) was employed to obtain 3D joint center trajectories from these videos (2.B), which were subsequently imported into OpenSim. The generic thoracoscapular model was also imported in OpenSim with virtual markers (pink) (difficult to see in the figure as these are located within the joints) that corresponded to locations of the joint centers (experimental markers) (blue) (2.C). The OpenSim pipeline was almost the same for both methods (1.D and 2.D), including the steps: scaling, inverse kinematics (IK) and static optimization (SO) (with the external dumbbell force as additional input). The only difference was that for scaling and IK, experimental markers from bony landmarks and joint centers were used for the marker-based method, whereas only joint centers were used for the marker-less camera method.

Figure 3

Musculoskeletal model with (A) scapular degrees-of-freedom and (B) shoulder muscles that control the scapula. Reprinted from Muscle contributions to upper-extremity movement and work from a musculoskeletal model of the human shoulder by Seth et al. (17).

Figure 4

Results for the middle (left), anterior (middle) and posterior (right) deltoid muscle forces during the lateral fly exercise. Results obtained from OMS (blue) and a single camera (red) are presented as mean ± 1SD (shaded) over all participants.

Figure 5

Results for the trapezius scapula superior (left), trapezius scapula middle (middle), and trapezius scapula inferior (right) muscle forces during the lateral fly exercise. Results obtained from OMS (blue) and camera (red) are presented as mean ± 1SD (shaded) over all participants.

Figure 6

Results for the rotator cuff muscle forces during the lateral fly exercise. Results obtained from OMS (blue) and camera (red) are presented as mean ± 1SD (shaded) over all participants.

Figure 7

Results for the biceps brevis (left), biceps long (middle) and triceps (right) muscle forces during the biceps curl exercise. Results obtained from OMS (blue) and camera (red) are presented as mean ± 1SD (shaded) over all participants.

Figure 8

Results for the rotator cuff muscle forces during the biceps curl exercise. Results obtained from OMS (blue) and camera (red) are presented as mean ± 1SD (shaded) over all participants.