ابراهيم جعفري مجنده

Human pose estimation is one of the most basic and challenging problems in the field of computer vision, where we want to extract human body pose from images or input video sequence. In estimating the pose of a single person from two-dimensional images, the position of the joints of the human body is calculated and determined in the input image. In other words, body pose estimation is the process of estimating and eva‎luating the body configuration from an image. Recent advances in deep learning techniques have led to huge and significant progress in the field of human gesture estimation, the main reason for which is the increasing number of new applications such as human-robot interaction, game industry, sports activity analysis, etc. Deep learning models for pose estimation need many images to train and learn human pose features, and this issue is considered one of the challenges of this issue. One of the most powerful deep learning algorithms in machine vision is the convolutional neural network algorithm, which requires a huge data set for learning. This type of network has the ability to extract the feature and perform the desired operation of the problem automatically. Transfer learning is used to solve one of the problems of training convolutional neural networks to perform gesture estimation and improve the training process. In this thesis, we intend to use convolutional neural networks and transfer learning to estimate human pose from two-dimensional images. After designing and implementing the proposed model, the model was trained and tested using the Tensorflow library and Python language. The PCKh criterion was used in the eva‎luation of the proposed model. This criterion showed good accuracy and performance by obtaining 83.78% accuracy. As expected, models based on convolutional neural networks and also using transfer learning solved some of the problems in this issue.

تخمين ژست انسان , شبكه هاي عصبي كانولوشني , يادگيري عميق

human pose estimation , convolutional neural network , Deep learning

ebrahim jafari

Seyed Aliasghar Beheshti Shirazi