- Статьи
- Science and technology
- Street algorithms: a neural network for tracking transport has been created in the Russian Federation
Street algorithms: a neural network for tracking transport has been created in the Russian Federation
Russian scientists have developed a neural network for tracking vehicles in low light conditions. Functionally, it is an intelligent vehicle tracking and classification system. With LightHead technology, cars can be distinguished even with unstable frame rates, low visibility, and limited computing power. It is able to dynamically recognize and classify moving objects, which is especially important in urban environments where the viewing angle and lighting can vary. At the same time, full-fledged testing in real conditions has not yet been conducted.
What is a smart tracking system for transport?
MTUCI specialists have taught neural networks to track cars even with unstable frame rates, poor lighting, and limited computing power. They have adapted computer vision technologies to real-world conditions. The developed algorithm allows you to accurately "recognize" an object, even if it temporarily disappears from the frame, is blocked by other vehicles, or is captured on an inexpensive low-resolution camera.
The experts took the DeepSORT object detector as a basis and adapted it to work with lightweight neural network models, such as the integrated YOLO-NAS system, which is the latest AI model for detecting objects.
Then the scientists implemented dynamic frame skipping (1-3 at a time), while maintaining the accuracy of object identification. This approach is very important for autonomous systems, video surveillance and unmanned platforms, where saving resources is critical.
— Our task is to make tracking work not only in a data center with the appropriate capacities, but also on a street camera in a small town, where there may be a weak signal, and recording is carried out with skipped frames, - explained Timur Potapchenko, Candidate of Technical Sciences, Associate professor of the Department of Software Engineering at MTUCI.
As the expert explained, scientists are currently testing a combination with the YOLOv5 and YOLOv8 models, their technology claims to be a new generation of YOLO models. She got the name TrackFusion. The launch of a pilot project is nearing as part of tracking at an intersection with unstable lighting and high traffic — real cameras, real urban traffic, the developers told Izvestia.
The second aspect of the tracking application was the task of "seeing" the car and understanding what it is like. In the basic version, algorithms can confuse a passenger car with a truck, especially when viewed from a poor angle. Experts have added an additional classification add—on to DeepSORT, which "refines" the model as data accumulates from different frames, which allows dynamically increasing the accuracy of determining the type of object. It is called LightHead and is based on a lightweight neural network, also known in classification as CNN.
— This approach works like a "second opinion" — if the first trigger gave an inaccurate grade, the system has a chance to correct its decision, — said Yuri Sadiev, a graduate student at MTUCI.
Vehicle recognition
This development can be used not only as an independent tool, but also as a basis for creating various solutions in the field of logistics and security. The system may be in demand among federal and regional authorities (the Ministry of Internal Affairs, the Ministry of Transport, Rosavtodor, Rosstat, and the Ministry of Finance), logistics companies, and retailers, for example, to select the most optimal routes, taking into account the congestion of various types of vehicles, said Deputy Director for Technology Transfer at the NTI Competence Center for Big Data Storage and Analysis Technologies." Timofey Voronin is based at Moscow State University.
— The development looks very promising with a wide range of applications. Two points can be noted — the ability to work in low light conditions and with low computing power, which makes the vehicle classification model accessible and attractive. Accessibility is also due to the fact that small companies will be able to afford the use of this model without high capacity costs," he added.
Computer vision technologies can currently classify objects according to predefined data. However, the process of determining in urban conditions is complicated by the constantly changing landscape, seasons, weather patterns, lighting due to the alternation of day and night, and other factors. The system can also, for example, identify a nearby flying bird as a car, says Anton Averyanov, CEO of the ST IT group of companies.
— Such systems are now used to predict traffic, intelligently switch light signals at intersections and optimize car movement, so it is especially important that they correctly identify vehicles and are constantly evolving. It is this technology that has been improved and refined by experts from MTUCI. Now, after passing the tests, the technology can be implemented for use. The accuracy of traffic detection and prediction in intelligent systems will improve, due to which there will be more intersections using this technology," he noted.
According to Igor Mishin, an expert at NTI Avtonet, the development will improve the efficiency of vehicle monitoring and improve road safety. Perhaps the neural network will help to find stolen cars, recognize people and track interference on the roadway. In addition to highways, the neural network will find applications in the field of highway maintenance and emergency services.
Road AI remains one of the leading trends in the development of smart cities around the world. The use of neural networks in urban life is developing the infrastructure for highly automated vehicles — self—driving cars, self-driving taxis - that are becoming part of our lives. Many world capitals are actively developing AI algorithms to automate urban processes and improve transport network management. Moscow sets its own trends in this regard, and is one of the digital leaders, the expert noted.
Alexey Ermakov, an IT expert and lecturer at Synergy University, believes that the authors of the project need to maintain a balance between adapting existing developments and developing their own innovations. At the same time, the idea of combining YOLO-NAS—type models with DeepSORT-type tracking algorithms demonstrates a rational approach - instead of creating technologies from scratch, it upgrades and refines already proven methods. This allowed the MTUCI scientists to save resources and, ideally, accelerate the implementation of new technologies in industrial and transport tasks, he concluded.
Переведено сервисом «Яндекс Переводчик»
