Video Anomaly Understanding: From Pixel-Level Detection to Semantic Reasoning

Video anomaly understanding is an emerging frontier in artificial intelligence that seeks to automatically detect, interpret, and explain unusual events, behaviours, or objects in video streams. As an emerging field in dynamic image analysis, video anomaly understanding aims to extend traditional image retrieval and analysis paradigms toward temporal reasoning and semantic interpretation in visual data. These capabilities power diverse applications ranging from intelligent surveillance and industrial safety to healthcare and autonomous navigation. Despite rapid progress in deep learning, the current state of research remains limited by challenges such as sparse anomaly occurrence, weak supervision, poor generalization across domains, and the need for explainability and efficient deployment on edge devices. Traditional anomaly detection models are typically confined to frame‑level deviations or handcrafted feature spaces, which fail to capture higher‑order context or semantic reasoning.Recent breakthroughs in self‑supervised representation learning, foundation and large‑scale visual‑language models, and spatio‑temporal transformers have opened new possibilities for video anomaly understanding. Researchers are exploring paradigms that integrate prior knowledge, reasoning modules, and causal inference to discern context‑driven anomalies. Concurrently, the democratization of aerial and egocentric video datasets, coupled with advances in edge AI optimization and energy‑aware inference, has broadened the scope of practical deployment scenarios. Nevertheless, bridging pixel‑level signals with high‑level semantics, explaining the causes of anomalies, and ensuring robust operation under unseen conditions remain open problems requiring coordinated research efforts.This Research Topic aims to stimulate cross‑disciplinary collaboration on next‑generation video anomaly understanding that unifies perception, reasoning, and efficiency. The goal is to advance models capable of robust spatio‑temporal awareness, explanation generation, and real‑world adaptability. Key objectives include exploring data‑efficient learning strategies (one‑/few‑/zero‑shot), building interpretable representations for contextual and behavioural anomalies, and enabling edge‑ready AI systems that preserve accuracy and transparency. Contributors are encouraged to integrate methods from computer vision, multimodal learning, knowledge reasoning, and human‑in‑the‑loop evaluation.To gather further insights in this fast‑evolving field, we welcome contributions addressing, but not limited to, the following themes:—Weakly, semi‑, or self‑supervised approaches for video anomaly detection and understanding—Spatio‑temporal anomaly localization, tracking, and reasoning—Explainable and interpretable video anomaly reasoning models—Large‑model and multimodal foundation model adaptation for anomaly understanding—Edge AI and resource‑efficient deployment of video anomaly detection systems—Aerial, drone‑borne, or multi‑view anomaly analysis scenarios—Benchmark datasets, evaluation protocols, and standardized metrics—Human‑centric, safety‑critical, or ethical aspects of automated anomaly reasoning Topic editors: Jiaxu Leng, Ji Gan, Manish Khare, Boying Wang, Haosheng Chen Source: https://www.frontiersin.org/research-topics/81965/video-anomaly-understanding-from-pixel-level-detection-to-semantic-reasoning