Scholarlink

This paper presents a comprehensive approach to adapt and optimize YOLO models for urban environment detection specifically in Tlaxcala, aimed at assisting visually impaired people in navigation on urban environments. We propose and implement a methodology that combines transfer learning, continual learning, and model compression techniques. The study focuses on detecting four critical urban elements: cars, bicycles, traffic lights, and crosswalks, using datasets from Berkeley Deep Drive, KITTI, and a custom dataset of Tlaxcala. The validation results showed a mAP50 of 94.5% and mAP50-95 of 87.3%. However, when tested on simulation experiments using Tlaxcala images, performance dropped to 61.4% mAP50 and 29.8% mAP50-95.

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

YOLO Meets Tlaxcala: Toward Real-Time Urban Object Detection for Assistive Technology

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Emilio Márquez Martínez,
  • Ernesto Mendieta Cuecuecha,
  • Jesús García Ramírez,
  • Eric Ramos-Aguilar

摘要

This paper presents a comprehensive approach to adapt and optimize YOLO models for urban environment detection specifically in Tlaxcala, aimed at assisting visually impaired people in navigation on urban environments. We propose and implement a methodology that combines transfer learning, continual learning, and model compression techniques. The study focuses on detecting four critical urban elements: cars, bicycles, traffic lights, and crosswalks, using datasets from Berkeley Deep Drive, KITTI, and a custom dataset of Tlaxcala. The validation results showed a mAP50 of 94.5% and mAP50-95 of 87.3%. However, when tested on simulation experiments using Tlaxcala images, performance dropped to 61.4% mAP50 and 29.8% mAP50-95.