Human Pose Recognition using ResNet with RJTM

Abstract

The goal of this study is to develop a system that can automatically detect human poses in
video. Human posture recognition (HPR) is a term used to describe the ability of a computer
to automatically identify the human stances that appear in a video. A wide range of obstacles,
including but not limited to human shape and motion, occlusion, a crowded background and
moving cameras and lighting conditions, make this a challenging task to solve.
As a starting point, the most popular and prominent current methods of investigation are
examined, analysed and presented. Deep learning-based and handmade feature-based
approaches are grouped together in the literature review. Once this handmade and deep
learning-based pose recognition framework is developed, it will be used throughout the
research by integrating new algorithms in both the handcrafted and deep learning domains for
pose recognition. This approach was unable to distinguish between positions with finer
texture and stances with greater similarity in several scenarios. Relative Joint Trajectory
Maps (RJTM) are proposed as a color-coded map to represent the trajectory information of
skeletal joints and trained on a strong Dense ResNet. Publicly available human pose datasets
are used to evaluate the approaches in this study. Discussions and future study ideas in the
field of human pose identification are also included in the conclusion of this research project.