Estimating Human Poses Using Deep Learning Models

Abstract

Over the past decade, extensive research has focused on the extraction of 3D human poses from images. The existing datasets must effectively address common challenges related to pose estimation. These datasets serve as valuable resources for evaluating, informing, and comparing different models. Deep learning models have gained widespread adoption and have demonstrated impressive performance across various domains of research and engineering. In this study, we employ these models, leveraging the open-source libraries OpenCV and Keras. To enhance the diversity and complexity of the training and testing process, we utilize the MPII Human Pose dataset. Specifically, we train and test the ResNet50 and VGG16 models using this dataset, resulting in significant improvements. The model's performance is evaluated based on the validation rate of the dataset and the accuracy of our model was 88.8 percent for VGG16 and 67 percent for ResNet50.