Video Surveillance for Indoor Office Environment Based on Object-Level Anomaly Detection

Abstract Traditional methods of Abnormal Behavior Detection (ABD) process the surveillance video on frame-level, which ignores object-level abnormal behavior patterns. To address the problem, this paper presents Object-Level Anomaly Detection model (OLAD), which aims to model various normal behavior patterns of different objects and the normal interaction patterns between them. Specifically, OLAD introduces an encoding-embedding network to transform object-level information into the feature space. By integrating such information, OLAD processes both frame-level and object-level cues in the video for ABD. In addition, we construct our own dataset Northking2022 especially for office scenes because of the lack of public datasets for indoor office environments. Experimental results show that OLAD gains better performance on both the public benchmark and Northking2022.