- Статьи
- Internet and technology
- AI flew: the Russian software package will accelerate the training of neural networks for UAVs by 10 times
AI flew: the Russian software package will accelerate the training of neural networks for UAVs by 10 times
Russian engineers have developed their own AI training program designed to be installed on domestic unmanned aircraft systems. This is the first domestic system that, according to the developers, is able to replace popular foreign services and at the same time does not depend on sanctions restrictions. The project has already been tested during the assessment of forest areas, which is especially important during the period of increased fire danger. According to the authors, the system performs AI tasks about ten times faster than its analogues. However, experts express doubts that developers will be able to build a full-fledged ecosystem around the solution.
A complex for accelerating the training of AI systems for drones
NTI engineers have developed a new neural network complex designed to accelerate the training of artificial intelligence systems for unmanned aerial vehicles. The solution is not a separate tool, but a full—fledged platform that covers the entire lifecycle of creating machine learning models for UAVs - from intelligent marking of aerial photographs to training, testing and subsequent deployment of algorithms on onboard computing systems, including domestic Elbrus processors.
The key difference of the project is that all processes take place inside the infrastructure without transferring data to external clouds. This eliminates the risk of leaks and dependence on foreign services. In addition, the system implements mechanisms for intelligent image sorting by angle change, ranking by information content, and automatically transferring markup between intersecting frames. This allows you to speed up the formation of training samples by an average of 3-10 times, and in some scenarios — up to 50 times.
— The advantage of this solution is a trusted environment where the risks of unauthorized access to information are eliminated. We have created specialized marking automation tools for aerial photography from drones, which significantly increase the efficiency of training samples," said Sergey Ivanov, technology developer.
The platform is adapted to the key tasks of computer vision for UAVs: detection, simultaneous tracking of several dynamic objects (multi-object tracking) and tracking targets. An additional element is a coordinate—based tracking algorithm that captures objects by their exact position on the map, reducing the likelihood of false positives.
— The assessment of the forest area has become a separate area, including the determination of the volume of wood, age, breed and number of trees for the purposes of harvesting or forestry management. The neural network automatically calculates the parameters of trees, that is, height, age, diameter of the trunk and crown, doing this based on crown segmentation and three-dimensional clouds of points. This is especially important in the context of the growth of wildfires and the need for operational monitoring," the developer explained to Izvestia.
The project has already been tested on real aerial photography data in the Leningrad and Pskov regions. According to the level of technological readiness, it has reached the stage that allows it to proceed to trial operation. In the future, it is planned to integrate multimodal models and create a centralized library of ready-made neural network solutions, while, as the authors note, the platform is potentially capable of becoming a domestic alternative to international ones such as Roboflow.
Prospects for the implementation of the platform
Experts noted the relevance of the technology for the rapidly developing market of unmanned devices.
— The prospects for development look very broad against the background of the active introduction of drones into forestry. When widely implemented, it can significantly increase the operational efficiency of unmanned systems, reduce costs and ensure the highest quality result," said Nikolai Ivashov, an expert at Aeronet Research Institute, an official representative of Fly Drone.
The implementation should proceed in stages: first in the format of pilot projects, for example, based on national parks, and then with scaling to regions. In the future, the technology may go beyond the forest industry and be used in a wider range of tasks, including various monitoring missions, the expert said.
— The development fits into the global trend of infrastructure development around machine learning models, where data preparation and exploitation tools play a key role. Now there is a lot of progress in the binding and instrumentation of AI agents, rather than in their creation from scratch. Such platforms take over data preparation, preprocessing, and results processing," said Andrey Korigodsky, CEO of Over, and market expert at Aeronet Research Institute.
According to him, the emergence of such solutions is critically important for the large-scale implementation of artificial intelligence in various sectors of the economy.
— The project is already proving its efficiency in practice. Its advantage is its focus on import independence and local deployment, which is becoming a key factor for government and industrial customers," said Yaroslav Seliverstov, a leading AI expert at University 2035.
He added that the model stack used provides a balance between accuracy and performance, but there is still potential for further development through the integration of more modern architectures.
In his opinion, the competitiveness of this AI solution will be determined not only by the quality of the algorithms, but also by the ability of the authors to build a full-fledged ecosystem, including a library of ready—made models, scaling tools and integration with various hardware, which is extremely difficult to achieve in the current market conditions.
Переведено сервисом «Яндекс Переводчик»
