Researchers at the Gwangju Institute of Science and Technology (GIST) have developed an artificial i..

Researchers at the Gwangju Institute of Science and Technology (GIST) have developed an artificial intelligence (AI) expertise that mechanically detects object modifications by evaluating photographs taken at totally different instances and paths in the similar area.

GIST introduced on the 2nd {that a} analysis workforce led by Professor Kim Eui-hwan of AI Convergence Department has developed “VSCDNet (Video-based Scene Change Detection Network), an artificial intelligence model for scene change detection based on images.

Existing change detection technology mainly compares photos taken from similar locations one by one. For this reason, there was a limitation in that the accuracy was lowered when the shooting location or angle was changed.

Instead of comparing individual images, the research team applied a method of analyzing the flow of the entire image. VSCDNet compares images taken in the past with current images to find corresponding scenes, and accurately detects only areas where actual changes have occurred.

This automatically detects changes in the actual environment, such as the disappearance of laptops or changes in the location of items. The part where the change has occurred is marked with a separate “change masks,” so that the user can check it at a glance.

The research team built a large-scale dataset consisting of 1090 images and more than 1.13 million frames, including virtual spaces and real indoor environments, for technology verification.

Experimental results show that VSCDNet recorded higher accuracy than conventional change detection techniques and maintained stable performance even under various conditions where image length, image quality, and number of changed objects vary.

In an experiment using a real mobile robot, the robot moved to a different path and automatically detected a situation in which a door opened or an object disappeared in the captured image. The function of remembering and learning newly emerged objects was also confirmed.

In particular, this technology is evaluated as a core technology of next-generation autonomous robots and smart surveillance systems in that it can determine “what has modified” by comparing the past and the present beyond simply recognizing the current scene.

Professor Kim Eui-hwan said, “VSCDNet can examine photographs taken from totally different paths with out separate location data or spatial maps,” adding, “We anticipate it for use in varied fields similar to indoor patrol robots, good safety surveillance, facility administration, and IoT-based good indoor methods.”

The research shall be offered at ICML 2026, the world’s most prestigious convention in the subject of AI and machine studying, to be held at COEX in Seoul in July.