Student Findings Could Guide Vision-Based Detection Research
Dr Jasmin Martin is giving engineering students first-hand project experience that could guide future research on vision-detection technologies.
Martin supervises final-year students for their capstone project on vision-based detection through the QUT School of Electrical Engineering and Robotics.
The project is closely related to her work on an algorithm that enables unmanned aerial vehicles (UAV) to detect aircraft at a range of more than 2km.
“In my research, I’ve taken a control systems approach to detecting aircraft in image sequences, but students take a machine and deep learning approach, which is something I’d like to investigate more,” Martin said.
“Students work on detecting a medium-sized fixed-wing aircraft at distances ranging from about 500 metres to 3 kilometres.
“They investigate spatial and temporal aircraft behaviour — to learn what the aircraft looks like to a vision-based system, how it approaches and moves in an image sequence.
“Findings from these student projects could be used to guide research direction.”
Each year students have access to previous students’ findings and, according to Martin, have made significant progress towards aircraft detection.
Detecting more detail in images
This year, mechatronics student Somayeh Hussaini is working on the vision-based aircraft detection problem under the co-supervision of Martin and Professor Jason Ford from the QUT Centre for Robotics.
She is developing a machine learning model that uses the same behaviour a human brain uses to function — deep learning.
“Deep learning is a technique where artificial neural networks are able to learn data features from large amounts of data,” Hussaini said.
“Deep learning models are inspired by the complex dynamic characteristics of our brain. Larger network depths can give us more accurate predictions.
“In terms of identifying images, that means we can detect more detail and identify or accurately predict images of aircraft as small as possible.”
A machine learning model can be trained to identify patterns in images and make predictions using the deep learning technique.
“If I gave a machine learning model lots of pictures of human faces and trained it to identify patterns that divided them into genders, it would then be able to predict the gender of a human if I gave it a new image,” Hussaini said.
“In this case, we are training the machine learning model to identify and track an aircraft when it’s a great distance away — only a few pixels on-screen — and surrounded by clouds.”
Because of the complexity and time needed to create machine learning models, Hussaini is using an existing set of algorithms proven to be useful in detecting small objects in sequential images.
She started by using synthetic data, images in which researchers added aircraft to synthesise the most ideal scenario of visible aircraft.
“By having a bounding box around the aircraft in sequential images, I can train the machine learning model to identify and track its movement so that it can predict the location of the aircraft in unseen video sequences,” Hussaini said.
“In the future, when an aircraft takes off it will start capturing its surroundings and continually detect other aircraft.”
Next, Hussaini will use real data — video sequences of aircraft in clear and cloudy conditions — from researchers who have published their data sets.
“Researchers are always publishing papers on new ways to improve the model so it can predict more accurately,” Hussaini said.
“These could apply to everyday products like vacuum cleaners, not just aircraft, but would be very useful for UAV to operate autonomously while protecting other aircraft through early detection and avoidance.
“My project is only on detection at the moment, but the next step would be trying to implement a system to avoid other aircraft.”
Mechatronics students have exposure to mechanical, electrical and computer engineering fields giving them a range of options for career specialisation.
Hussaini focused her studies on mechanical and software engineering while holding multiple scholarships and working with the university as a student ambassador helping high school students learn about STEM.
From her second year, she has also worked as a software engineering intern for global technology company Deswick but is yet to decide if she will take a graduate position in industry or continue with a career in research.
“If I have more time in research, I would like to explore algorithms more to see how I can improve existing models,” Hussaini said.
Robotic kangaroo leg was a springboard to research
Martin herself followed the path to research after her capstone project to develop a robotic kangaroo leg before completing her mechatronics engineering course.
“I really enjoyed my final-year project and wanted to continue the research. I was also a sessional academic for about three years during my undergraduate course and wanted to keep teaching,” Martin said.
After graduating, Martin started her PhD studies and joined an industry-funded vision-detection project led by Professor Ford and Dr Tim Molloy from the QUT Centre for Robotics.
Since then, she has helped develop a new ‘quickest detection’ mathematical theory and design algorithms that identify aircraft among visual clutter such as clouds.
“Until we have adequate sense-and-avoid systems, there is a barrier to flying UAV in national airspace,” Martin said.
“The final safety layer for unmanned aircraft is a sense-and-avoid system, which is the implied regulatory requirement that a UAV has to be able to sense and avoid a collision threat, either as well as or better than a human pilot.”
The right teacher can define your career pathway
Among her supervisory and research work, Martin also teaches first-year students the foundations of electrical engineering and third-year control engineering through the School of Electrical Engineering and Robotics.
If not for her own teacher explaining maths in just the right way, Martin says she may not have found herself on the path to engineering or research.
“I was really lucky in high school. I had an awesome maths teacher who explained concepts in a way that I understood easily and encouraged me to do engineering,” Martin said.
“I really enjoy problem-solving, and that’s what I liked about my mechatronics course. There were so many projects in which we were just given a goal and had to find the solution on our own.”
While some students can struggle with not having defined outcomes to problems in Martin’s experience, she believes that engineering training teaches them how to approach questions and find answers.
“Learning how to find answers is not just a part of your degree but how you want to live your life. You can’t just rote learn — you have to actually understand problems,” Martin said.
Learn more
Explore the QUT School for Electrical Engineering and Robotics
Visit the QUT Centre for Robotics
Explore QUT research
