Kidnapping for ransom continues to pose a significant security threat in Nigeria, and the rapid identification of hand-carried weapons in surveillance footage could offer early warnings to law enforcement agencies. This study presents a computationally efficient two-stage vision pipeline that integrates the speed of You Only Look Once (YOLOv5s) with the localization capability of a Faster RCNN (ResNet50FPN) to detect knives and related weapons in real time. The system is evaluated in a zero-shot manner, utilizing off-the-shelf Common Objects in Context (COCO) weights without any domain-specific fine-tuning on the 928-image Sohas weapon dataset. Experimental results indicate that the hybrid cascade achieves image-level coverage of 99.6% and processes a frame in 0.19 seconds on a single Tesla T4 GPU (approximately 5 fps), meeting the latency requirements of typical Nigerian Closed-Circuit Television (CCTV) deployments. However, the detection accuracy was modest: the mean Average Precision was 0.0019 at IoU 0.50 and 0.0168 at IoU 0.30, indicating that localization error is the predominant failure mode. When compared with recent fine-tuned models that report mAP ≈ 0.65–0.75 on weapon-specific datasets, the zero-shot baseline quantifies the performance gap attributable to the domain shift. Qualitative analysis further identified the small-object scale, metallic false positives, and class imbalance as major sources of error. The presented code, pretrained weights, and evaluation logs were released to provide an open, reproducible benchmark for subsequent research. By establishing both the feasibility of real-time inference on commodity hardware and the limitations of generic weights, this work lays the foundation for future, domain-adapted systems aimed at mitigating kidnapping incidents in Nigeria through automated weapon detection.

Hand Carried Weapons Detection Kidnapping Mitigation Nigeria real time surveillance YOLOv5 Faster R CNN Hybrid Detector Zero Shot Baseline