Multi-Scale Human Pose Estimation Using Morphological Segmentation and Deep Learning

Abstract

The intersection of computer vision, computer graphics, and machine learning leads to human modeling and pose estimation. Human pose estimation has been and continues to be a challenging issue in computer vision because of occlusions, differences in sizes of bodies, and intricate joint movements. Even with recent breakthroughs in deep learning, correctly identifying salient events in real-world settings remains a major challenge. To solve these problems, we introduce a new approach to accurate human pose estimation that combines morphological segmentation with deep learning. Morphological operations help segment the input images, and Convolutional Neural Network (CNN) architecture like VGG16 is utilized to extract significant features from the segmented images, which are then classified using classifiers. The model, which is proposed, is trained on two publicly shared datasets, MPII and LSP, to capture diverse human poses with varying conditions and scales. We emphasize the success of our approach in attaining sophisticated results in human pose estimation tasks by engaging in extensive testing and evaluation. Our method effectively deals with occlusions and intricate poses along with accurately detecting key points. We also highlight the model's interpretability and generalizability, presenting its strength in numerous real-life scenarios.