A comprehensive review of anomaly detection: from images to videos and compressed video streams

Anomaly detection (AD) is fundamental element in modern computer vision, allowing intelligent systems to identify abnormalities from typical behavior in images, videos, and compressed video streams. Conventional approaches are limited by statistical or reconstruction-based learning, whereas current advancements in deep and hybrid architectures have revolutionized anomaly detection into a data-driven field capable of modeling intricate spatial–temporal dynamics. This review offers a thorough, interdisciplinary synthesis of 196 formally cited works (161 peer-reviewed journal/conference publications, plus arXiv preprints, doctoral dissertations, and a technical report, included primarily for post-2024 coverage) drawn from a broader corpus of 252 systematically screened papers (2015–2025), presenting a three-stage taxonomy that integrates classical, deep learning, and hybrid or multimodal methodologies. The distinctive contribution of this work is its innovative focus on compressed-domain anomaly detection, a novel approach that utilizes motion vectors, discrete cosine transform coefficients, and residuals directly from encoded streams to enable real-time, privacy-preserving analytics without complete decoding. This study incorporates post-2024 developments such as diffusion-transformer hybrids, federated compressed-learning frameworks, and multimodal vision-language reasoning. It critically assesses the trade-offs between accuracy, latency, scalability, and privacy, highlighting ongoing difficulties in codec generalization and distributed robustness. This review also integrates diverse research across three visual domains, presenting a fresh, codec-agnostic perspective and outlining a roadmap for advanced, efficient, and ethically aligned anomaly detection systems tailored for real-world intelligent multimedia environments.