Is the Diagramming Neural Network OpenPose Ready for Widespread Use?

Jason Plants* and Sergiy Yakovenko

Division of Exercise Physiology, West Virginia University, Morgantown, WV 26506

Presentation Category: Physical Sciences & Engineering (Poster Presentation #162)

Student’s Major: Exercise Physiology

OpenPose is a tool built with an artificial neural network (ANN) approach to extract human body posture from images. ANN approach has many applications ranging from software for self-driving cars to measuring joint angles in orthopedic surgery. The challenge is that the computation performed with ANN has to be precise, efficient, and accurate. To assess whether OpenPose is ready for widespread use in body posture estimation, the focus of this project is to determine this network’s accuracy. Twenty pictures from the internet were gathered ranging from a single person fully in frame to a crowd of people spread all across the image. These pictures were entered into OpenPose following the steps demonstrated by the MathWorks workshop Estimate Body Pose Using Deep Learning [1]. The images were analyzed to extract body posture. To be eligible for scoring, pictured individuals had to be 75% visible and reasonably discernable by the human eye. This subjective process resulted in a 61.1% accuracy rate for those unobstructed and/or directly facing the camera and a 39.7% accuracy rate for those partially obstructed and not directly facing the camera. These results indicate that OpenPose works well in ideal situations, but may not be accurate in most real-world situations. Partial obstruction and posture variations must be overcome if this ANN approach is to be utilized in widespread applications. [1] Estimate Body Pose Using Deep Learning. MathWorks. https://www.mathworks.com/help/deeplearning/ug/estimate-body-pose-using-deep-learning.html#EstimateBodyPoseUsingDeepLearningExample-3

Funding:

Program/mechanism supporting research/creative efforts: WVU's Research Apprenticeship Program (RAP) & accompanying HONR 297-level course