Novosibirsk State Technical University has developed a software system for detecting and tracking fast-moving objects in a video stream. The project is based on computer vision technologies, neural network models, and multi-object tracking algorithms.
The development is designed for tasks where stable analytics is important in difficult shooting conditions, such as high-speed movement of objects, blurred images on frames and partial overlaps of the picture. According to Egor Antonyanets, an assistant at the Department of Automated Control Systems at the Faculty of Automation and Computer Engineering, the system processes video in real time and generates object trajectories while preserving their identifiers between frames.
"The key feature of the system is the ability to work in scenarios where classical detectors often miss due to motion blur. We have further trained the model on specialized data and supplemented it with a tracking module to compensate for short—term detection losses," said Egor Antonyants.
Technically, the processing is built in several stages: the neural network detects objects on each frame, the tracker compares detections between frames and maintains continuous trajectories, and the visualization system displays the result in a form convenient for analysis. This approach allows us to obtain not only the fact of detection, but also the dynamics of the movement of objects over time.
"Unlike the basic solutions focused on relatively static scenes, the new tool is adapted to high-speed videos and shows more stable performance on complex fragments. This makes the system useful for video surveillance applications, transport analytics, industrial monitoring, and research projects," says Denis Savochka, the main developer of the project, a 4th—year undergraduate student.
In the future, the system can be expanded, support for additional classes of objects can be added, generalization capability can be improved on new video domains, and the solution can be integrated into existing video surveillance and analytics circuits.
Earlier, students of the Faculty of Automation and Computer Engineering at NSTU-NETI created a visual product search system that allows you to find the right product in the marketplace catalog based on photos. The development is based on artificial intelligence, which understands the essence of search, and not just compares pixels.