MIPT has designed an automatic tracking system that coordinates observations from multiple cameras to follow objects over time. Andrey Leus, a leading researcher at the Laboratory of Special Purpose Digital Systems at the Moscow Institute of Physics and Technology, described the project to socialbites.ca, outlining how the technology works in practice.
Today’s computer algorithms can recognize images from photos and video, including faces. Yet most systems still only confirm that a target appears in a single camera frame and struggle to connect distant sightings into a continuous timeline. For reliable long term tracking of a person or a machine, human operators typically intervene to piece together a sequence of frames from different viewpoints.
Leus explained that the team created a method to generate tracks that link together images of observed objects captured by different cameras. In many modern surveillance setups, each object found by image recognition is treated as a new discovery, with no obvious link to earlier or later views. The new approach resolves this gap by matching distinctive features such as vehicle footprints or characteristic marks across cameras, even when a key detail like a license plate is visible in one frame but not in another. Consider a car moving through a city where the plate is legible only on one camera; the system can still associate that car with the same vehicle on other cameras where the plate is not readable.
Through this mechanism, the surveillance network can automatically track a target object as it transitions from one camera to the next. The concept mirrors human perception: if a person appears in a close-up wearing a hat and a long coat, the observer can recognize the same individual in other frames by considering overall silhouette and context, even without every identifying detail visible. In this design, each object is represented not by a single image but by a continuous series of observations gathered from various angles and distances.
Leus stated that when a track is recorded, each camera contributes a sequence of frames rather than a single image. This yields a robust dataset—hundreds and thousands of images that encode the object’s appearance across time and space. After the system memorizes this dataset, recognizing the object in subsequent cameras becomes straightforward, enabling uninterrupted surveillance across the network. The approach emphasizes continuity and consistency rather than isolated snapshots, which improves the reliability of tracking under changing conditions and viewpoints.
One potential application is integration with the physical layout of the camera network to monitor real-world activity more precisely. For instance, the system could be used to assess attendance patterns in educational settings by determining which classroom a person occupies and at what times, all while maintaining a non-intrusive, observational stance. The design supports scalable monitoring across multiple locations, offering a coherent flow of data as individuals move through spaces with different vantage points.