Human Pose Estimation with CNNs and LSTMs

Human pose estimation from images and videos has been a very important research field in computer vision. In this thesis, we present an end-to-end approach to human pose estimation task that based on a deep hybrid architecture that combines convolutional neural network (CNNs) and recurrent neural networks (RNNs). CNNs used to map the input image to feature space (fixed dimensionality), and then deep RNNs to decode the target sequence pose from the feature space. We experimented different RNNs architectures and we found out that deep bidirectional LSTM outperformed other architectures. Additionally, we tested different training strategies and influence of temporal information. Our final model is trained to minimize the average Euclidean distance between the ground-truth 3D joint coordinates and those predicted by our method. To validate our approaches, we experimented on several datasets. Our experiments on Patient MoCap dataset outperformed algorithm, which is deemed stateof-art. We also evaluated on Human3.6M dataset, we achieved 19 cm. These results show the accuracy of the model and the integrity of the estimated pose that learned from the training data. Our quantitative and qualitative observations verify that our method makes significantly accurate. To the best of our knowledge, we are the first to show that CNNs+RNNs models can make accurate 3D joint coordinates estimation from depth images.