Video surveillance systems rely on human observers with limited attention capacity, which prevents visual coverage from becoming a timely alert when a firearm is involved. The aim of this work was to review and organize the literature on firearm detection in video surveillance using deep learning, focusing on datasets, models, evaluation and deployment. Articles published between 2016 and 2026 reporting experimental results on surveillance scenarios with documented datasets were collected. The findings show that the problem is constrained above all by data: the small size a weapon occupies within a frame is the main bottleneck, together with occlusion, lighting and objects visually similar to a weapon. Two-stage detectors, one-stage detectors and cascaded schemes offer different trade-offs between false positives and speed, although published figures do not warrant a ranking among architectures, since they come from custom datasets and heterogeneous metrics. It is concluded that these systems identify the appearance of a weapon rather than determine a threat, so the human verification protocol following the alert is as relevant as the detector itself