A Semantic Compression Framework for Video Surveillance Applications

We develop a semantic compression framework for video surveillance in low-bandwidth communication. The framework aims to extract abstract representations of input videos, enabling the monitoring task while transmitting minimal amounts of data. We propose different levels of abstraction, including captions describing the scene, detection of objects of interest, and sample frames of relevant actions. The framework combines state-of-the-art object detection and multimodal large language models to generate informative and concise representations of the monitored areas, providing efficient data compression rates. We evaluate the framework by accounting for compression rate, quality of information, and delay.