The future of intelligent traffic management: KTU student develops a model for autonomous systems

According to the US National Highway Traffic Safety Administration (NHTSA), as many as 94 per cent of traffic accidents are caused by human error or decision-making. This figure is often cited as a key argument for the development of autonomous vehicles. Can technology do what we frequently fail to – assess the environment objectively, quickly, and consistently?

Bartas Lisauskas, a student at Kaunas University of Technology (KTU), believes it can. He is developing a system that enables computer vision to function efficiently, even with extremely limited computing resources.

Bartas, a student of Software Engineering, says he has always looked for solutions with practical value. From the very beginning of his studies, he was more interested in how technology works in real-life situations than in theoretical models. Today, this approach is becoming increasingly relevant, as autonomous systems are being applied not only in transportation but also in industry, logistics, and service sectors.

Smart traffic management in cities

“Developing efficient and safe autonomous technologies that are useful in real life is a task, requiring a lot of time and effort,” says Lisauskas.

The situation is improving, but for these technologies to become widespread, their practical implementation needs to be considered, both in infrastructure and everyday services. One area where Lisauskas sees significant untapped potential is smart traffic management in urban environments.

Autonomous systems require more than just the ability to move – they need to perceive their surroundings, make context-based decisions, and do so quickly. Computer vision is one of their key components. It enables systems to observe, identify, and classify objects in their environment, such as vehicles, pedestrians, road signs, or traffic lanes.

“Human decisions are often driven by experience or emotion, whereas computer vision, if properly implemented, can ensure consistent, objective assessment,” he adds.

Perceiving the environment with limited resources

One of the core challenges is ensuring that environmental awareness is not only accurate but also achievable under limited-resource conditions. To address this, Bartas developed a semantic segmentation model optimised for efficient performance without requiring powerful computing hardware.

The model is designed to classify an image into distinct categories, such as road, car, sky, or pavement. This process typically demands high computational capacity, especially when high precision is desired.

“This system is being developed specifically for resource-constrained environments, where powerful equipment is unavailable and all computations must be performed quickly and efficiently,” he says.

“The transformer architecture allows for efficient data processing, while attention management helps the system concentrate on the most relevant image regions. This combination delivers better results under constrained conditions,” Bartas explains.

According to him, the main challenge is finding a balance between model performance and efficiency: “Many existing models perform well but are resource heavy. My goal is to create a system that’s lightweight enough for real-device use, yet accurate enough to be practical”.

Maskeliūnas highlights that this work on resource-efficient semantic segmentation using transformer-based architectures demonstrates not only technical depth but also a strong awareness of practical deployment constraints – an essential consideration in intelligent systems engineering.

“Bartas’s solution does not require a powerful GPU cluster or expensive GPU boards – it works perfectly on an old computer. It aligns with the broader direction of AI-driven innovations developed at KTU – integrating AI with signal processing to create systems that are both intelligent and efficient in dynamic, real-world environments,” he adds.

Article Efficient Transformer-Based Road Scene Segmentation Approach with Attention-Guided Decoding for Memory-Constrained Systems can be accessed here.