L-BFGS Optimization-Based Human Body Posture Rectification—A Smart Interaction for Computer-Guided Workout

Human pose estimation possesses a significant potential in reducing the cases of injuries sustained during strenuous physical activities like gym workout sessions. Pose estimation aims toward locating human body joints accompanied by visual inputs. Major challenges in perceiving inputs from a cluttered background have been addressed in this work. The paper presents an end-to-end methodology, based on two different algorithms, MediaPipe Holistic Pipeline and Modified BlazePose, that perform real-time pose detection and warn users regarding incorrect body movements. The reason for choosing these models over more popular and robust Human Pose Estimation models like OpenPose is the lightness of the models. A comprehensive review of the two models has been illustrated in this research work for 2D as well as 3D pose estimation. The research piece has also incorporated the subsequent methodology applied to identify faulty body joint angles by comparing them with the optimum ones. Considering an uncontrolled environment, the testing of the models have been done. The models have confirmed the identification of human Region Of Interest as foreground from background objects and calculated the pose angle of joints. A systematic study pertaining to the visibility of joints for each model is also presented in the paper for better reference. The deviation of the measured angle of joints from the optimal ones is properly annotated followed by optical analysis.