Researchers at the Competence Center of the National Technology Initiative (NTI), created based on the Moscow Institute of Physics and Technology, are working on creating an autonomous system for intelligent analysis of photo and video information for drones and situational use analysis systems. The prototype will be tested next year, the press service of Russian Venture Company told TASS on Wednesday.
The system's peculiarity is that the neural network will be able to independently search and analyze the necessary templates for training, which will allow processing much larger volumes of data than using other machine learning methods that require manual marking of training images. The new approach will reduce the period of adaptation of such systems to the assigned tasks and increase their efficiency, the authors say.
The development is based on a unique technology of variational autoencoders, making it possible to search for objects atypical for the area in one image. Unlike existing systems for detecting objects using security cameras, the invention combines classic tools for analyzing photo and video streams with neural network algorithms. The authors expect that the new technology will be useful in real-time recognition of objects moving at high speed against a cloudy sky or forest background — existing systems cannot yet overcome this barrier due to the inability to analyze the distorted image. Simultaneously, the flexible architecture will allow the system to be embedded in small devices, including drones and mobile structures for work on the ground.
Functional tests of the prototype system are scheduled for the second half of 2021. The developers plan to test the mobile complex on the territory of the institute.