home
ManchesterHi, I'm Archie Mearns. I'm a Robotics and AI DPhil/PhD student at the University of Oxford within the Robotics and AI for Net Zero Centre for Doctoral Training(RAINZ CDT). I am currently at the University of Manchester as a visiting student for the first year of the CDT programme.
Interests
My research interests are:
I am also interested in low level/embedded development and technology ethics, particularly regarding AI, privacy and information ethics.
Education
DPhil Robotics and AI (RAINZ CDT)
BSc Computing Science - First Class Honours
Experience
Research Assistant (University of Aberdeen)
Student Demonstrator (University of Aberdeen)
Projects
Undergraduate Dissertation
Accurate depth estimation is critical for enabling autonomous mobile robots to create 3D
maps
...
Read Abstract...
for use in navigation and other tasks. Deep learning-based stereo matching models
dominate classical methods for depth estimation on standard benchmarks, although have
not yet had widespread adoption in real-time embedded Visual-SLAM. This is in part
due to computational constraints and uncertainty of performance on real world data. This
project investigates optimised deep stereo matching models for improving mapping quality
in an embedded Visual-SLAM pipeline while maintaining real-time operation. Three
deep stereo matching models were integrated into an RTAB-Map pipeline on a TurtleBot3
robot platform equipped with a ZED stereo camera and an NVIDIA Jetson Orin Nano
edge GPU. Quantitative and qualitative evaluation of the resulting maps revealed that
deep learning-based models create higher quality and more complete 3D point cloud maps
than the classical baseline. Selective-IGEV produced the best map quality within the
real-time constraint using an optimised configuration that decreased latency by 23x at the
expense of only 5-12% accuracy loss. Benchmark results alone are insufficient for model
selection as they are unreliable predictors of real world transfer performance, deployment
suitability or downstream VSLAM performance. The results demonstrate that optimised
deep stereo matching models can improve real world mapping quality in real time fulfilling
the research aim and enabling better mapping quality for any
VSLAM system with an embedded GPU.
Show Less
Technologies
help
Navigate with either mouse or keyboard.