Design of a Multisensory System for Detecting Guns in Video Surveillance Environments in Metropolitan Lima

Berrospi L. · IEEE Access (2026) · DOI: 10.1109/access.2026.3678942

Resumen

The automated detection of handguns in video surveillance systems faces the critical challenge of dynamic occlusions and perspective variations, where conventional unimodal approaches suffer from high omission rates (false negatives). To overcome this vulnerability, this research proposes and develops a homogeneous multisensory fusion architecture. The main objective of this research is to design a highly robust system that confirms the threat of a handgun presence in real time through geometric and probabilistic consensus from multiple cameras, integrating automated alerts through the Internet of Things (IoT). The methodology consisted of concurrent unimodal evaluation using YOLOv11n, followed by a cross-spatial registration stage using a homography matrix calibrated at chest height. To resolve the three-dimensional variation error that invalidates classic metrics such as Intersection over Union (IoU), a Euclidean Distance filter was implemented on dehomogenized and denormalized centroids. Upon exceeding a probabilistic threshold of joint certainty, the system transmits asynchronous notifications via the Telegram API. The results demonstrated outstanding effectiveness, achieving a Recall of 96.30% and an F1-Score of 94.21%, while maintaining an operational fluidity of 53 frames per second (FPS) (18.8 ms latency per cycle). In conclusion, the integration of Euclidean analysis in projective spaces optimizes the reliability of the multisensory system, eradicating false negatives due to depth disparity without compromising the low computational cost required for public safety.

Instituciones: Universidad Tecnológica del Perú