Are Current Computer Vision Models for Handgun Detection Ready to Be Deployed in Camera Surveillance Systems?

• Jorge Batista UFBA
• Tatiane Rios UFBA
• Cassio Prazeres UFBA
• Rubisley Lemes UFBA
• Maycon Peixoto UFBA
• Gustavo Figueiredo UFBA
• Frederico Durão UFBA
• Eduardo Santana de Almeida UFBA
• Ivan Machado UFBA
• Hérsio Iwamoto Positivo Tecnologia
• Ricardo Rios UFBA

Resumo

Computer Vision (CV) has become a key component of modern camera surveillance systems, enhancing public and private security. One persistent challenge in this domain is the reliable detection of handguns in low-quality images, in which weapons are often small, partially occluded, and carried discreetly to avoid attention. To address this, recent advances have focused on specialized models trained on specific weapon types and human poses. However, two crucial aspects remain underexplored. First, while state-of-the-art models have improved handgun detection rates, they also exhibit a noticeable increase in false positives, often mistaking unrelated objects for firearms. Second, although combining gun detection with pose estimation is a common approach, it may offer limited value in practice, as criminals tend to maintain natural, non-suspicious postures to avoid drawing attention. In this work, we analyze the effectiveness of such models in scenarios that reflect real-world urban settings: low-resolution surveillance footage, small concealed weapons, and non-suspicious human behavior. To address these limitations, we propose two key contributions: an enhanced approach that extends a gun-and-pose model with a different method for computing the intersection between predictions and ground truth; and a new set of images designed to better train models for detecting handguns in the context of non-suspicious human behavior. In addition to improving performance in our experiments, our findings also highlight the need for more robust and context-aware handgun detection systems in CV-based surveillance.