Text

Introduction: Advancements in computer vision technology are quickly changing how we can conduct musculoskeletal disorder (MSD) risk assessments in the field by using single handheld video cameras (i.e., phones). Given the quick emergence of these computer vision-based MSD risk assessment tools, there is limited knowledge and understanding of how to best use these tools to capture insightful data. Therefore, the purpose of this work was to determine how the angle of video recording and computer vision approach influenced estimates of trunk inclination and shoulder elevation angles relative to joint angles estimated from a validated lab-grade markerless motion capture approach.

Methods: Forty participants completed a series of simulated work tasks (i.e., lifting, palletizing, above shoulder work). Videos of the participants’ body motion was captured using eight synchronized 2D video cameras oriented in ~45° increments surrounding the participant. Video data from all eight cameras were processed through Theia Markerless software to produce 3D human pose data from which trunk inclination and shoulder elevation angles were computed. Video data from each individual camera angle was processed through 4 different single camera computer vision-based pose estimation models yielding trunk inclination and shoulder elevation angles for comparison. The mean absolute difference (MAD) in time series trunk and shoulder angles between the lab-grade Theia Markerless data and each computer vision-based pose estimation method was calculated. Repeated measures ANOVAs were completed to determine the effects of camera angle and computer vision approach on trunk inclination and shoulder elevation angle estimates for each task.

Results: Camera views perpendicular to the major plane of motion and on the same side as the joint of interest (i.e., left shoulder elevation captured best from the left camera view) typically resulted in lower MAD values (e.g., Figure 1 [Abb. 1]). However, interaction effects between camera angle and computer vision approach we also observed.

Discussion/Conclusions: Although our data suggests that camera views perpendicular to the major plane of motion generally produce most accurate trunk inclination and shoulder elevation angle estimates, it also suggests that effectively implementing these technologies is not as simple as “point and shoot” to capture and assess whole-body MSD risks. For example, to accurately assess risk for both left and right shoulder, best practice may be to video capture the movement/task from both left and right side of the body, respectively. Since each single camera computer vision-based MSD risk assessment tool is developed on their own proprietary models, when making a purchase decision, practitioners should ask about which directions of capture were used to train their underlying model. Single camera pose estimation continues to show promise, but continued efforts to understand the strengths and limitations of these technologies remain important to inform and guide use.