Graph-based human pose estimation using neural networks

This thesis investigates the problem of human pose estimation (HPE) from unconstrained single two-dimensional (2D) images using Convolutional Neural Networks (CNNs). Recent approaches propose to solve the HPE problem using various forms of CNN models. Some of these methods focus on training deeper and more computationally expensive CNN structures to classify images of people without any prior knowledge of their poses. Other approaches incorporate an existing prior knowledge of human anatomy and train the CNNs to construct graph-representations of the human pose. These approaches are generally characterised as having lower computational and data requirements. This thesis investigates HPE methods based on the latter approach. In the search for the most accurate and computationally efficient HPE, it explores and compares three types of graph-based pose representations: tree-based, non-tree based, and a hybrid approach combiningbothrepresentations. Thethesiscontributionsarethree-fold. Firstly,theeffectofdifferent CNN structures on the HPE was analysed. New, more efficient network configurations were proposed and tested against the benchmark methods. The proposed configurations achieved offered computational simplicity while maintaining relatively high-performance. Secondly, new data-driven tree-based models were proposed as a modified form of the Chow-Liu Recursive Grouping (CLRG) algorithm. These models were applied within the CNN-based HPE framework showing higher performance compared to the traditional anatomy-based tree-based models. Experiments with different numbers and configurations of tree nodes allowed the determination of a very efficient tree-based configuration consisting of 50 nodes. This configuration achieved higher HPE accuracy compared to the previously proposed 26-node tree. Apart from tree-based models of human pose, efficient non-tree-based models with iterative (looping) connections between nodes were also investigated. The third contribution of this thesis is a novel hybrid HPE framework that combines both tree-based and non-tree-based human pose representations. Experimental results have shown that the hybrid approach leads to higher accuracy compared to either tree-based,or non-tree-based structures individually