Scholarlink

<p>Security surveillance systems in unconstrained environments deter criminal activities and safeguard public safety by enabling real-time monitoring of threats. Such environments, including protests, riots, and crowded venues, are particularly vulnerable to security breaches. Fast and accurate detection of dangerous objects is crucial for preventing incidents and ensuring public safety, especially in highly challenging environments such as protests and riots, where factors like tear gas, water, fire, moving crowds, and occlusions hinder object detection. This research presents a comprehensive object detection framework tailored for identifying suspicious objects in public disorder scenarios using a one-stage object detection model for optimal speed and accuracy. A key contribution of this study is the development of a benchmark dataset designed to evaluate object detection algorithms in unconstrained environments. Eight You Only Look Once (YOLO) versions (v3–v10) were evaluated using this dataset, comprising five object classes: knives, swords, axes, stones, and sticks. YOLOv10 achieved the highest mean average precision (mAP) of 82.50%, followed by YOLOv4 (80.91%), YOLOv8 (80.60%), and YOLOv5 (71.78%). YOLOv3 recorded the lowest mAP at 65.88%. For detection speed, YOLOv10 achieved 30 FPS, followed by YOLOv9 and YOLOv8 (29 FPS each). Additionally, YOLOv10 outperformed all models in average precision (81.96%), recall (83.82%), and F1-score (82.48%) across all object classes. With its superior speed and accuracy, YOLOv10 is ideal for real-time security surveillance, while YOLOv4 and YOLOv8 remain suitable for accuracy-focused scenarios. Our study represents a major advancement in AI-driven public security surveillance, offering both an optimized detection framework and a benchmark dataset for object detection in unconstrained environments.</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

Advancing suspicious object detection in unconstrained environments using computer vision

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Kanagasabai Thiruthanigesan,
  • Ruwan D. Nawarathna,
  • Roshan G. Ragel

摘要

Security surveillance systems in unconstrained environments deter criminal activities and safeguard public safety by enabling real-time monitoring of threats. Such environments, including protests, riots, and crowded venues, are particularly vulnerable to security breaches. Fast and accurate detection of dangerous objects is crucial for preventing incidents and ensuring public safety, especially in highly challenging environments such as protests and riots, where factors like tear gas, water, fire, moving crowds, and occlusions hinder object detection. This research presents a comprehensive object detection framework tailored for identifying suspicious objects in public disorder scenarios using a one-stage object detection model for optimal speed and accuracy. A key contribution of this study is the development of a benchmark dataset designed to evaluate object detection algorithms in unconstrained environments. Eight You Only Look Once (YOLO) versions (v3–v10) were evaluated using this dataset, comprising five object classes: knives, swords, axes, stones, and sticks. YOLOv10 achieved the highest mean average precision (mAP) of 82.50%, followed by YOLOv4 (80.91%), YOLOv8 (80.60%), and YOLOv5 (71.78%). YOLOv3 recorded the lowest mAP at 65.88%. For detection speed, YOLOv10 achieved 30 FPS, followed by YOLOv9 and YOLOv8 (29 FPS each). Additionally, YOLOv10 outperformed all models in average precision (81.96%), recall (83.82%), and F1-score (82.48%) across all object classes. With its superior speed and accuracy, YOLOv10 is ideal for real-time security surveillance, while YOLOv4 and YOLOv8 remain suitable for accuracy-focused scenarios. Our study represents a major advancement in AI-driven public security surveillance, offering both an optimized detection framework and a benchmark dataset for object detection in unconstrained environments.