³ÉÈË´óƬ

Abstract: The application of AI for real-time video analytics has the potential to revolutionize video surveillance, enhancing safety, compliance, and convenience for all. However, this promising advancement also presents formidable challenges, primarily centered around cost and accuracy. One pressing concern is the substantial computing resources required to run AI models, particularly in large-scale scenarios with thousands of cameras. The scale of computation can become a significant barrier, necessitating the development of AI inferencing pipelines for high throughput processing. Moreover, deploying a single fixed AI model for all surveillance scenes introduces the risk of numerous errors and inaccuracies. Consequently, post-deployment model improvement becomes a crucial requirement to meet the diverse demands of varying surveillance environments.