Performance optimization of computer vision systems for human-robot interaction : a containerized approach using specialized hardware.

Abstract

This thesis investigates the implementation of a transparent computer vision system for humanrobot interaction, comparing specialized hardware solutions against traditional webcam-based approaches. With the increasing integration of robots in various settings, there exists a growing need for efficient, reliable, and transparent vision systems that can operate effectively in resource-constrained environments. The research addresses limitations in conventional webcam implementations, particularly high CPU utilization, limited frame rates, and lack of transparency in data processing. Through empirical testing, the study demonstrates that the specialized Oak-D camera implementation with on-device neural processing significantly outperforms traditional webcam approaches, achieving a 7.6x improvement in frame rate (23 fps versus 3 fps) while reducing CPU utilization from 80-100% to 20-40%. This performance enhancement addresses critical limitations in resource utilization and system responsiveness. The research also incorporates containerization through Docker to ensure deployment consistency across diverse systems, enhancing portability and maintaining standardized configurations. This approach facilitates simplified dependency management while introducing negligible performance impact, addressing the challenge of platform dependency in robotics applications. By combining specialized hardware with open-source software and containerization technologies, this study provides a practical framework for developing transparent, efficient computer vision systems suitable for human-robot interaction in resourceconstrained environments. The findings contribute to the field by establishing quantifiable advantages of specialized vision hardware and transparent implementation approaches, with implications for enhancing system performance in robotic applications.