Research Proposal: Does RG-chromaticity and depth differ from RGB in distinguishing anomalous objects in stereo image datasets for self-driving vehicles?
Self-driving vehicles are being tested live, on public streets, but when we attempt to validate the underlying image models’ performance, there remains some unexplained underperformance, which prevents reliable validation processes.
Vehicles are dangerous, massive objects that operate at speed near exposed members of the public. Self-driving technologies seek to reduce the probability of collisions, by both removing human inattention and unreliable analysis. However, partially automated systems consistently result in greatly increased human inattention, and the computational analysis it is replaced with is challenging to validate as reliable.
In 2018, the first pedestrian fatality by a self-driving vehicle was observed.
Further dependability research has been advocated. This project is part of asking: What is causing these apparently capable models to be assessed as operating incorrectly?
I recently withdrew from studying (general) Engineering at Durham University, and am now exploring defence, development, and diplomacy: writing for Durham Diplomatic Society, and volunteering remotely as visiting coach for Durham University Negotiation Society.
The Laidlaw Leadership and Research Scholarship's summer research internship, with my supervisor Prof. Toby Breckon, has provided me great insight into computer vision and pattern recognition research, which not only happily furthered my curiosity, but also better positioned me to report on the Royal Aeronautical Society's recent conference on automation, which I expect to mark a turning point in my career.
My previous work in STEM includes co-authoring "Preliminary results from the ROJER-X...", a conference paper which you can find the abstract of through my LinkedIn profile's "Publications" section. The full paper on the ROJER-X remains in preparation.
Please sign in
If you are a registered user on Laidlaw Scholars Network, please sign in