Machine Vision and Image Processing

Training a Deep Learning (DL) model requires large quantities of labeled training data to be used as ground truth during the training process. This can be a bottleneck for rapid deployment of new sensors to a platform as data labeling is typically a highly repetitive and person-hour intensive process.

It is commonly assumed that image data can be classified using ready-made DL models. Plextek explored using one such model to provide the ground truth for automatic labeling of radar data, as well as a base model for transfer learning to thermal imagery.

Maximising the efficiency of solar farms is critical for operators, with drones often used to capture precise thermographic and video diagnostic data to warn of any defects.

When Above Surveying approached Plextek for help with automating its labour-intensive post-survey work, the challenge was to translate its data into a format that machine learning tools could process – removing solar glint along the way.

We recommended five areas of investigation along with example processing chains which could improve the client’s current processing workflow to increase overall efficiency and add extra capabilities.

Read the full case study here. 

Non-GPS reliant Alternative Navigation (AltNav) methods have been gaining traction recently, one of which being Visual Scene Matching (VSM). VSM algorithms are not well suited to address changing environments, often failing due to vegetation changing with the season, or due to temporary objects being included in the reference points, such as parked cars.

This is where Semantic Segmentation comes in handy, by adding context to the reference images, what objects are temporary or permeant, traditional VSM algorithms can be augmented to greatly improve performance.

Read the full case study here.

Unauthorised drones flying near airports cause significant safety and security issues and extensive disruption to travellers and airlines. But differentiating between drones and birds can be a challenge for traditional radar, leading to a high rate of false alarm.

In response to this challenge, the Plextek team developed an automatic target recognition framework that enabled a wide range of different ML algorithms to be explored and optimised based on data from our PLX-U16 radar. The ML algorithms exploit the micro-Doppler effect caused by small, internal movements within an object – the rotation of rotor blades on a drone, for example, or wing movements relative to the body of a bird.

The result was the development of an ML classification technique that can accurately distinguish between birds and drones and even identify the type of drone.