<p>Video-based violence detection has emerged as a critical research area in intelligent surveillance systems, aiming to automatically identify aggressive and harmful human behaviors in complex real-world environments. As a distinct and challenging subclass of <i>video anomaly detection (VAD)</i>, violence detection presents unique difficulties, including motion ambiguity, visual similarity between violent and non-violent actions, and the limited availability of annotated training data. This paper presents a comprehensive review of deep learning techniques for video-based violence detection, with particular emphasis on Vision–Language Models (VLMs) and their role in semantic reasoning, weakly supervised learning, and multimodal violence understanding. The review systematically analyzes the key components of violence detection pipelines, including feature extraction strategies, temporal modeling approaches, learning paradigms, loss function design, regularization techniques, classifier selection, anomaly and violence scoring mechanisms, and benchmark datasets. Furthermore, this study highlights emerging research trends, such as transformer-based architectures and the increasing adoption of vision–language models as complementary feature extractors for enhancing contextual and semantic understanding in complex violent scenarios. By consolidating recent advances, identifying open research challenges, and outlining future directions, this review provides a structured and up-to-date reference for researchers and practitioners working on robust, scalable, and ethically responsible video-based violence detection systems.</p>