Real-Time Permanent Change Proposals for Abandoned Object Detection

In this work, we address the significant challenge of detecting permanent changes in video scenes, a foundational element for the advancement of efficient surveillance systems. We introduce a unique two-step algorithm. The first step involves temporal smoothing by utilizing methods such as median operation and Mixture of Gaussians (MOG) to effectively model the background frame, thereby eliminating temporary objects within a dynamic window. The subsequent phase leverages a MOG-based strategy to identify alterations over the derived short-term background frames. Specifically tailored for real-time operations, our approach is adaptable to edge surveillance devices, even those equipped with minimal GPU capabilities. The algorithm adeptly identifies static foregrounds or lasting scene modifications, setting the stage for possible detection of abandoned objects in later processes. Tested on widely recognized datasets, our method demonstrates almost impeccable recall rates. Despite a lesser emphasis on precision, the produced false positives can serve as valuable data in subsequent surveillance stages. Optimized to operate near real-time, our system is compatible with compact embedded GPUs, such as the Nvidia Jetson. This approach offers a pragmatic and robust solution to a growing need in the automated video surveillance sector, aiming to bolster public safety. Notably, our method approaches state-of-the-art outcomes on standard benchmarks while significantly reducing computation time, especially when deployed on suitable modern edge computing devices.