Humanoid skeletal pose estimation using depth sensors

Abstract

This dissertation presents the development of a modern pose estimation system using depth sensors and recent advances in machine learning, designed to improve upon legacy systems such as Microsoft Kinect for use in Virtual Reality (VR) applications. The proposed system integrates high-precision data collection with synthetic data generation to train a robust and generalizable model capable of real-time human pose tracking. Data was collected using SteamVR trackers, which offer sub-millimetre precision and low-latency tracking at a fraction of the cost of traditional motion capture setups like OptiTrack. These trackers enabled the creation of a high-quality dataset that approximates motion capture standards without the associated complexity or expense. To assess system robustness, the trained model was evaluated using an ASUS Xtion PRO depth sensor. Quantitative comparisons against the SteamVR tracker data (used as ground truth) were conducted, focusing on drift, positional accuracy, and performance under challenging conditions, including multi-person tracking and low-light environments. The results demonstrate that the system achieves accurate and reliable pose estimation across a range of scenarios. This work contributes a cost-effective, scalable, and adaptable approach to pose estimation in VR and other interactive computing environments.