Weapon detection in video surveillance systems is critical for enhancing public safety by enabling early identification of potentially dangerous situations. Traditional systems often rely on manual oversight leading to delays and inaccuracies. With recent advancements in machine learning, particularly deep learning, automated weapon detection models have gained prominence for real-time monitoring. This paper surveys various models used in weapon detection including YOLOv5, YOLOv8 with quantization, Faster R-CNN, and SSD MobileNet. These models have been enhanced using techniques like background subtraction combined with SVM and Canny edge detection, and further refined with ensemble methods such as MC-CNN, EfficientDet with EfficientNet-B0, and NMS (Non-Maximum Suppression). Several advanced models, including YOLOv4, Faster R-CNN, Inception-V2, and ResNet variants (ResNet18, ResNet34, ResNet50), have been applied for weapon classification tasks. Performance evaluations highlight the trade-offs between accuracy and computational efficiency, with YOLOv5 and YOLOv8 standing out for real-time applications. This paper also discusses the impact of fine-tuning pre-trained models like YOLOv4 for improved weapon detection in dynamic environments, thereby demonstrating the potential for these models to improve public safety and security in surveillance scenarios.