AI Feedback Architecture of Video Surveillance System

Abstract

The learning capacity of general deep learning models for object detection would not be large enough to represent real-world scene dynamics, and thus such models would be weak to ‘unseen’ data due to environmental changes. Therefore, in this study, we propose a new method to continuously improve the object detection algorithms by applying negative and positive learning mechanisms, especially for intrusion detector in video surveillance systems. By applying an iterative process where the current model is updating using new incoming data with a state-of-the-art model in a continual process of adaptation. The experimental results in various challenging videos for real video surveillance systems demonstrate that the proposed method offers a significantly improved algorithm accuracy with a low complexity, thus it is adapted for real-world systems.