A Semi-automatic Object Identification Technique Combining Computer Vision and Deep Learning for the Crosswalk Detection Problem

• Rúbio, Thiago R. P. M. ^{[1] [2]} ; Cruz, José Aleixo ^{[1] [2]} ; João Jacob ^{[1] [2]} ; Daniel Garrido ^{[1] [2]} ; Cardoso, Henrique Lopes ^{[1] [2]} ; Daniel Silva ^{[1] [2]} ; Rui Rodrigues ^{[1] [2]}
  1. [1] Universidade Do Porto

     Universidade Do Porto

     Santo Ildefonso, Portugal

     
  2. [2] Artificial Intelligence and Computer Science Laboratory (LIACC; Porto, Portugal)
• Localización: Intelligent Data Engineering and Automated Learning – IDEAL 2020. 21st International Conference: Guimarães, Portugal; November 4–6, 2020. Proceedings / Cesar Analide (ed. lit.), Paulo Novais (ed. lit.), David Camacho Fernández (ed. lit.), Hujun Yin (ed. lit.), Vol. 2, 2020 (Part II), ISBN 978-3-030-62365-4, págs. 602-609
• Idioma: inglés
• Enlaces
  • Texto completo
• Resumen

  • Object detection in the traffic domain has faced growing relevance through the years in developing autonomous driving mechanisms. As with vehicles, pedestrians face a very dynamic context, and identifying relevant objects from a pedestrian perspective presents many challenges. Improving the detection of some objects, such as crosswalks, is very relevant in this regard. This paper presents a technique that applies a computer vision approach to automatically generate datasets for training YOLO-based deep learning algorithms. An initial precision of 0.82 achieved with the generated dataset, which is increased to 0.84 after manually removing incorrect annotations. Results show that our approach leverages the dataset building process by reducing the manual workload needed. The approach could be used for training other object detection models used in traffic scenarios.