Edge AI Enabled Real-Time Video Analytics: A Full-Lifecycle Perspective

The unprecedented surge of video data from diverse domains such as intelligent transportation, security surveillance, and embodied perception (EP) has created an urgent need for efficient extraction of useful information to enable smart decision-making. As a result, real-time video analytics (VA) has become a key enabling technique for these applications. Recently, real-time VA has benefited from the rapid advancements of artificial intelligence (AI). However, cloud-based VA suffers from considerable latency, substantial bandwidth consumption, and potential privacy leakage. By moving service and intelligence from the cloud to the network edge, edge AI enabled VA has emerged as a transformative paradigm that enhances the efficiency and security of VA systems. Unfortunately, existing surveys on edge AI enabled VA rarely consider the lifecycle of AI models, ranging from training, deployment to inference. Motivated by the above gap, this paper presents the first comprehensive survey of edge AI enabled real-time VA from a fulllifecycle perspective, covering edge-based training paradigms, deployment strategies, post-deployment adaptation mechanisms, and inference optimizations. Furthermore, considering that traditional AI models for VA still face significant challenges in multimodal reasoning, contextual understanding, and generalization to unseen scenarios, we review the recent advancements of edge AI enabled real-time VA powered by large language models (LLMs). Specifically, we present several representative vision tasks, outline the deployment challenges of LLMs for edgebased VA, and discuss the corresponding inference solutions. Finally, we highlight the key research challenges and discuss future opportunities to guide continuous efforts in this field.