Scene Change Detection for Robotic Patrol System

Scene change detection (SCD) aims to obtain semantic object variance so that it enables patrol robots to infer the degree of risk in the surveillance areas as they repeatedly monitor designated areas. As previous studies mainly handle SCD regarding static environments, they may fail to adapt to the patrol robot system where large scene change occurs due to frequent activity of movable objects and viewpoint changes by varied localization results of robots. As practical solution, we introduce a SCD method employing a graph neural network (GNN) in which entire objects in a scene are designated to graph nodes and tracked for calculating the translation of these nodes (correspondences). We also present a practical scheme for patrol robots to handle sequential query events during the analysis of the object transitions compared to the past scene. From an experiment with a custom real-world dataset, we demonstrate the method achieves around 92% recall rate.