AI Developed to Detect Only Changes by Comparing Videos Taken at Different Times and Locations [Reading Science]

To Be Presented at ICML 2026: Performance Validated with Over 1.13 Million Frames
Expected Applications in Patrol Robots and Smart Security Surveillance
Automatic Detection of Door Openings and Object Movements

A domestic research team has developed an artificial intelligence (AI) technology capable of identifying only the objects that have actually changed within the same space, even if the time and location of the footage differ. This innovation is expected to enable patrol robots to autonomously recognize environmental changes between the past and present, and allow security systems to automatically monitor the movement or loss of objects.

On June 2, the Gwangju Institute of Science and Technology (GIST) announced that the research team led by Professor Uiwhan Kim of the Department of AI Convergence has developed an AI model called "VSCDNet (Video-based Scene Change Detection Network)" that detects real-world object changes by comparing videos taken at different times and along different routes.

Example of using VSCD in a real robot environment. A mobile robot repeatedly visits the same space and compares captured videos to detect changes such as doors opening or objects newly appearing or disappearing. The detected change areas can be used for visual surveillance and incremental object learning. Provided by the research team

Existing change detection technologies typically compare photos taken from similar locations and at similar times, resulting in a significant drop in accuracy when camera positions or travel routes differ. This has been a major limitation for autonomous or indoor patrol robots in tracking environmental changes over long periods.

The research team addressed this issue by analyzing the overall flow of the videos rather than comparing individual images. VSCDNet compares a reference video from the past with a current video to identify corresponding scenes and precisely extract only the areas where actual object changes have occurred.

Through this approach, the system can automatically identify situations such as the disappearance of a laptop, changes in object locations, or the appearance of new objects. The areas where changes are detected are visualized as "change masks" and provided to the user.

To validate the technology, the research team also built a large-scale dataset that includes both virtual spaces and real indoor environments. Experiments using the dataset, which consists of a total of 1,090 videos and more than 1.13 million frames, demonstrated that VSCDNet outperformed existing change detection methods.

In particular, the system maintained stable detection accuracy under various conditions, such as differences in video length, quality, and the number of changed objects. In real-world mobile robot experiments, the system automatically detected situations such as doors opening or objects disappearing in videos shot along different routes, and also demonstrated its ability to remember and learn newly appearing objects.

Research team photo. (From left) Professor Euihwan Kim, Department of AI Convergence, Jiae Yoon, Integrated Master-Doctoral Student. Provided by GIST

The research team emphasized that this technology is significant because it goes beyond simply recognizing the current scene and can determine what has changed compared to the past. Since it can detect environmental changes without the need for separate location information or spatial maps, it is highly applicable to indoor patrol robots, smart security surveillance, facility management, and IoT-based smart indoor systems.

Professor Uiwhan Kim of the GIST Department of AI Convergence said, "VSCDNet is an AI model that not only recognizes the current scene, but also autonomously determines what has changed compared to the past. Because it can compare videos taken from different routes without additional location information or spatial maps, we expect it to be widely applied in various indoor environment management fields."

This research was carried out by integrated master's and doctoral student Jiae Yoon of the Department of AI Convergence, under the supervision of Professor Uiwhan Kim. It was supported by the Excellent Young Researcher Program of the Ministry of Science and ICT and the National Research Foundation of Korea, as well as the Self-directed Visual Intelligence Technology Development Project of the Institute of Information & Communications Technology Planning & Evaluation (IITP).

The research results are scheduled to be presented at ICML 2026, one of the world's top AI and machine learning conferences, which will be held at COEX in Seoul next month.

This content was produced with the assistance of AI translation services.