Real-time Exercise Posture Correction Using Human Pose Detection Technique

Abstract

Human pose detection is one of the fascinating research areas in the field of computer vision that has many unsolved challenges. Detecting and capturing human activity is advantageous in many fields like sport analysis, human coordination tracking, public surveillance etc. Due to the COVID-19 driven pandemic, it became hard for society to access exercising hubs and newbies into the fitness industry were left blank with almost no personal guidance that they usually get in gym in terms of exercising in the right way through one-on-one interactions. As these resources are not always available, human pose detection can be a medium to replace a human personal trainer by developing a real-time exercise posture correction system on recorded videos or realtime image stream that allows people to safely exercise at home avoiding injuries. This project uses a pre-trained OpenPose Caffe model with two datasets i.e., COCO and MPII to correct one of the exercises in the fitness industry. This project also discusses various pose estimation and key point detection techniques in detail and different deep learning models used for pose classification.