Human Joint Profile Extraction using Deep Learning Approaches

Abstract

Digital human modeling and gait analysis are essential for improving hip replacement surgery (HRS). In this study, Convolution Neural Networks (CNN) are used as a machine learning method to extract the most accurate stick-model from videos captured on a simple camera to represent gait and body components. We developed and tested multiple approaches to create an equitable skeleton model from an image. This process consists of two main parts: defining the joint locations using a CNN network in different architectures, and defining the connections into the final skeletons. A CNN has been trained, validated, and tested using the OpenPose software, which combines two different networks that have been tested on three data-sets for learning and evaluation. The results were satisfactory, but MobileNetV1 was evaluated for optimization of OpenPose computations and definitions. Several hyper-parameters were investigated to provide better representations. As a result of utilizing OpenPose methodology in conjunction with heavily optimized network design and post-processing code, and implementing MobileNet, the proposed solution has provided improved accuracy ratios.