This paper offers a systemic review of the deep learning methods to detect violence on campus, which is a critical issue in intelligent surveillance to improve the student safety and prompt cut off of violent accidents. The review reviews studies published 2018-2025, concentrating on model structure to detect fights, bullying, vandalism, and aggressive behavior on problematic campuses due to occlusion and light variations and complicated human interactions. The research design includes a comparative study of different deep learning networks, such as CNNs, RNNs, 3D CNNs, attention-based networks, transformers, graph neural networks, neuro-fuzzy, and multimodal systems and federated learning methods. The paper also assesses benchmark datasets frequently utilized, performance measures, and even real-time deployment considerations. Findings show that CNN models of light weight can fit well into real-time use but are not capable of time modeling but hybrid CNN-RNN and attention based models may provide better accuracy at increased computing cost. Transformer and multimodal models have shown promising performance, but are computationally expensive to e.g. deploy to edges. The review presents important research gaps, such as inadequate datasets to the specific campus, insufficient multimodal integration, privacy issues, and the necessity of explainable and lightweight implementation. This work can guide further research on viable solutions, effective, and privacy-conscious violence detection systems in a learning setting.
