Pose estimation neural networks in the context of the RoboCup Humanoid League

In this dissertation, we will be presenting research on pose-estimation neural networks applied to systematically learning a defined two-dimensional region. To date, most deep learning pose estimation research has been on random camera trajectories. However, robotic platforms in domestic and commercial environments require the ability to navigate freely within a defined space. RoboCup was used as the case study for this research. Founded in 1997, the international RoboCup competition seeks to advance humanoid robotics through the medium of competitive robotic soccer, with the goal of a team of humanoid robots defeating the human world champion team. Humanoid robots face limitations in computational power preventing them from taking full advantage of many state-of-the-art localisation techniques. Deep Learning has been used for pose estimation since 2015 by modifying image classification neural networks for the task. However, when implementing a ResNet50 pose estimation architecture in a small indoor soccer room, the technique exhibited an unexpected poor accuracy. Six hypotheses were taken from the literature and from the experimental results of this dissertation to attempt to resolve this issue. However, none of the literature based hypotheses were supported by the data. Two approaches were developed that was able to correct the issue: a smaller ResNet50 architecture, and the simplification of the solution space. This thesis proposes a modification to the architecture and such variation reduced the median position error by between 39% to 83%. The simplification of the solution space approach used a 360 degree camera in a rigid 0.5 metre grid, which resulted in further improvements. Additionally, the effects of noise on both the original and proposed architectures was tested, with the proposed architecture achieving a 44% reduction in the same error metric compared to the original architecture.