3D Human Pose Estimation from 2D Keypoints Detection using Optimization of Kinematic Models

In this thesis, we address the challenge of 3D Human Pose Estimation (HPE)and Hand Pose Estimation from 2D monocular images and video sequences.The ability to accurately determine human pose in 3D space is fundamental for advancing fields such as animating and controlling virtual characters, enabling intuitive human-robot interaction, and supporting various medical applications.Despite its importance, accurately estimating 3D human pose from 2D input remains a significant challenge due to the lack of depth information, which leads to visual ambiguities. This makes it difficult to distinguish between similar poses and to resolve occluded joints. We propose a novel approach that uses OpenPose for initial 2D keypoint detection, followed by a model based optimization technique. By utilizing Denavit-Hartenberg parameters,we develop a kinematic model of the human body and hand and define a cost function. Minimizing this function with the aid of the Jacobian matrix andnatural joint limits, we are able to find a 3D pose that most likely represents thepose from the given image. The results demonstrate that our approach achieves competitive accuracy and computational efficiency compared to state-of-the-art solutions. With further enhancements in error correction and optimization, our method has the potential to outperform existing approaches.