Scholarlink

Most navigation algorithms are developed for controlled environments, such as warehouses and factories, or for autonomous vehicles, where the challenge lies in fast decision-making over a limited set of obstacles. These conditions differ significantly from pedestrian spaces, which are highly unstructured and low-speed in nature. This work focuses on the adaptation of object detection algorithms for pedestrian navigation via transfer learning. Several models were fine-tuned using the VIDVIP dataset. While all the trained models showed similar performance, YOLOv12x managed to achieve a mAP50-95 score higher than 0.6, which shows a visibly better performance on test environments. This work highlights the potential of transfer learning with urban datasets for enhancing pedestrian navigation systems.

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

Adaptation of Object Detection Models Using Transfer Learning for Navigation in Urban Environments

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Aldo Yael Hernández Flores,
  • Jorge Alonso Vasquez Trujillo,
  • Daniel Sánchez-Ruiz,
  • Jesús García-Ramírez

摘要

Most navigation algorithms are developed for controlled environments, such as warehouses and factories, or for autonomous vehicles, where the challenge lies in fast decision-making over a limited set of obstacles. These conditions differ significantly from pedestrian spaces, which are highly unstructured and low-speed in nature. This work focuses on the adaptation of object detection algorithms for pedestrian navigation via transfer learning. Several models were fine-tuned using the VIDVIP dataset. While all the trained models showed similar performance, YOLOv12x managed to achieve a mAP50-95 score higher than 0.6, which shows a visibly better performance on test environments. This work highlights the potential of transfer learning with urban datasets for enhancing pedestrian navigation systems.