Keen AI Labs explore new use cases, technologies and approaches. Many work, others don’t, all are fascinating. Do you have a process that can be improved with AI? Learn more about our Labs process here.

active

Monitoring sand dunes using satellite and train mounted camera data

Synopsis:

Shifting sands pose a threat to rail safety and overhead line clearances. We’re using satellite and ground based imagery to detect dunes and encroaching sand and then trying to model how it will move.

Researchers:
Amjad KArim
Amjad Karim
Petar Gyurov
Petar Gyurov
Complete

Robots looking for bubbles to detect leaks and spot manufacturing imperfections

Synopsis:

Keen AI developed simple computer vision based tools to highlight areas in a video feed where leaks are present and present to a human who can repair the leak.

Outcomes:
Algorithms were reliably able to detect bubbles caused by air leaking from imperfections in the membrane. However we weren’t able to move the implementation since the cameras used to collect live video would physically get in the way of staff.
Further Reading:
Researchers:
Hamzah Reta
Amjad KArim
Amjad Karim
Complete

Neural networks and camera systems for detecting invasive plant species

Synopsis:

We developed a camera system and AI platform for detecting the presence of invasive plant species along linear infrastructure such as rail tracks and roads.

Outcomes:

We trained models able to detect Himalayan Balsam and Ragwort from photographs taken from the car mounted system. Vehicle speed had to be kept under 30 mph for safety and for contiguous footage. Future work to focus on increasing speed to at least 50 mph.

Further Reading:
Researchers:
Hamzah Reta
Amjad KArim
Amjad Karim
Petar Gyurov
Petar Gyurov
Dr. Tom August
ACTIVE

Bootstrapping computer vision model training using synthetic data

Synopsis:

Models require data to train on. Labelling this data manually is time consuming or the data may not exist. We’ve been working with a variety of techniques for generating synthetic data.

Outcomes:
Explored a number of techniques including using Blender to create 3D models as well as procedural models to generate basic examples of GPS text. The GPS synthetic experiments worked really well. 3D models were helpful for training models to identify steelwork towers but require significant effort creating photo realistic scenes.
Further Reading:
Researchers:
Thomas Emment
Hamzah Reta
Petar Gyurov
Petar Gyurov
Amjad KArim
Amjad Karim
ACTIVE

Using model uncertainty to reduce manual labelling effort

Synopsis:
We’re trying to figure out how we can use prediction uncertainties in deep learning network outputs to minimise false positives and prepare more targeted training data.
Researchers:
Qifan Fu
Prof. Greg Sablaugh
Petar Gyurov
Petar Gyurov
Amjad KArim
Amjad Karim
Complete

Identifying ash trees at risk of ash dieback

Synopsis:

Following the learnings from our previous study on identifying invasive species from roadside images, and continuing our partnership with UKCEH, we set out to devise an effective operational pipeline for surveying the ash tree population and diagnosing ash dieback.

Outcomes:

Key insight was determining when in the year to collect footage. Ash tree seed pods are very distinctive and easier to detect in photos collected in late autumn.

Further Reading:
Researchers:
Petar Gyurov
Petar Gyurov
Dr. Tom August