Human Pose Estimation And Action Recognition Using Deep Neural Networks

Human pose estimation and action recognition in videos are fundamental tasks to automatically analyse and understand the human behaviour for various applications.Both tasks are necessary to get information about the human body in both its motion and stillness cases.With the recent fast progress in machine learning and the advance in deep learning techniques,processing images and videos for pose estimation and action recognition using those methods became a key factor in improving the performance,especially deep neural networks models that offer an end-to-end efficient structure for feature extraction and classification or regression.In this thesis,we propose new techniques for pose estimation and action recognition using deep convolutional neural networks,which is a class of deep neural networks that was specially designed for 2D feature extraction.Convolutional neural network based techniques outperformed previous handcraft-based methods because of the ability of automatically learning low level and high-level features from training data.Since the key success for any image-based recognition process is based on extracting the relevant features for the desired task,in our proposed techniques,the focus was on how to improve the feature extraction by using new convolutional neural network structures.Our work tackles the problem from different aspects in term of data type and problem nature.First,we treated the 3D pose estimation in depth images and 2D pose in colored images as a regression problem that maps directly the input image to the pose locations using deep convolutional neural network models in an end-to-end learning process.Second,we tackled the action recognition problem using both depth images and 3D posture data to construct two descriptors that offer different types of motion features,then three channels of convolutional neural networks were designed for feature extraction and action classification.In the end,as a complementary work,and in the same context of analysing human behaviour,we propose a motion quantification algorithm to represent the body movement using 3D postures based on a calculation metrics without the need to a learning technique.Furthermore,a motion comparison algorithm is proposed to evaluate the similarity between two body movements based on the quantification algorithm.In order to test the performance of our work,the proposed pose estimation and action recognition techniques were evaluated on public benchmark datasets.Both quantitative and qualitative evaluation showed competitive results with the state-of-the-art methods.