Scholarlink

<p>Reliable detection of openly carried and concealed knives remains a challenging problem in artificial intelligence (AI) due to the small size, thin geometry, frequent occlusion, and material variability of blade objects. Although advances in deep learning have improved weapon detection performance, the literature remains fragmented across sensing modalities, datasets, and evaluation protocols, limiting reproducibility and systematic comparison. This paper presents a systematic and modality-aware review of AI methods for multisensor knife detection, covering visible-spectrum imaging, thermal infrared, X-ray imaging, millimetre-wave and terahertz sensing, and frequency-modulated continuous-wave radar. A PRISMA-informed literature selection process identified 39 primary studies published between 2018 and 2025. The reviewed works are analysed with respect to sensing modality, dataset characteristics, preprocessing strategies, detection architectures, multimodal fusion methods, and evaluation protocols. The review identifies recurring architectural patterns that improve blade detection, including multi-scale feature representations, small-object-aware training strategies, attention mechanisms, and modality-specific enhancement pipelines. It further provides a cross-modality synthesis and semi-quantitative summary showing that reported performance is strongly shaped by sensing modality, task formulation, dataset difficulty, and reporting practice, including inconsistencies in evaluation metrics, hardware platforms, and inference speed reporting, rather than detector architecture alone. True multimodal fusion remains underexplored, but available evidence suggests that it has potential to improve robustness under occlusion, concealment, and challenging environmental conditions. The review also highlights limitations in dataset diversity, evaluation consistency, and ethical considerations associated with deployment in real-world surveillance contexts. This survey makes three primary contributions. First, it introduces a modality-aware taxonomy that unifies open-carry and concealed-knife detection within a common AI framework. Second, it synthesises datasets, detector architectures, fusion strategies, and evaluation practices with emphasis on small-object performance and operational reliability. Third, it identifies key research gaps and outlines a roadmap toward standardised benchmarks, multimodal learning frameworks, physics-informed simulation pipelines, privacy-preserving deployment architectures, and robust real-world evaluation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A systematic review of multisensor methods for open-carry and concealed knife detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Ogonna Okafor,
  • David Ada Adama,
  • Doratha Vinkemeier

摘要

Reliable detection of openly carried and concealed knives remains a challenging problem in artificial intelligence (AI) due to the small size, thin geometry, frequent occlusion, and material variability of blade objects. Although advances in deep learning have improved weapon detection performance, the literature remains fragmented across sensing modalities, datasets, and evaluation protocols, limiting reproducibility and systematic comparison. This paper presents a systematic and modality-aware review of AI methods for multisensor knife detection, covering visible-spectrum imaging, thermal infrared, X-ray imaging, millimetre-wave and terahertz sensing, and frequency-modulated continuous-wave radar. A PRISMA-informed literature selection process identified 39 primary studies published between 2018 and 2025. The reviewed works are analysed with respect to sensing modality, dataset characteristics, preprocessing strategies, detection architectures, multimodal fusion methods, and evaluation protocols. The review identifies recurring architectural patterns that improve blade detection, including multi-scale feature representations, small-object-aware training strategies, attention mechanisms, and modality-specific enhancement pipelines. It further provides a cross-modality synthesis and semi-quantitative summary showing that reported performance is strongly shaped by sensing modality, task formulation, dataset difficulty, and reporting practice, including inconsistencies in evaluation metrics, hardware platforms, and inference speed reporting, rather than detector architecture alone. True multimodal fusion remains underexplored, but available evidence suggests that it has potential to improve robustness under occlusion, concealment, and challenging environmental conditions. The review also highlights limitations in dataset diversity, evaluation consistency, and ethical considerations associated with deployment in real-world surveillance contexts. This survey makes three primary contributions. First, it introduces a modality-aware taxonomy that unifies open-carry and concealed-knife detection within a common AI framework. Second, it synthesises datasets, detector architectures, fusion strategies, and evaluation practices with emphasis on small-object performance and operational reliability. Third, it identifies key research gaps and outlines a roadmap toward standardised benchmarks, multimodal learning frameworks, physics-informed simulation pipelines, privacy-preserving deployment architectures, and robust real-world evaluation.