• Zielgruppen
  • Suche
 

Traffic-sign Recognition from Crowdsourced Street-level Photos: a Deep Learning approach

Leitung:Zourlidou
Bild Traffic-sign Recognition from Crowdsourced Street-level Photos: a Deep Learning approach

Objective

The objective of this thesis will focus on the study of the detection and recognition of traffic signs from crowdsourced street-level photos (see photo on the right). 

By the term crowdsourced photos, we refer to photos that individuals capture and upload to platforms for obtaining needed services or ideas. The motivation behind detecting and classifying traffic signs from such (geotagged) photos is the enrichment of maps with the respective information, which can be further used for other services (e.g. driver-assistance tasks).   

For the pattern recognition task, this thesis will explore the applicability of some machine learning methods with focus on deep learning. Beyond the implementation and testing of the algorithms, an additional goal of this research work will be the approximation of the position of a single traffic sign captured from users in different photos. 

Tasks and time schedule

  • Literature review on computer vision for traffic sign recognition
  • Algorithms’ implementation 
  • Experiments and testing
  • Writing thesis documentation

Data and Tools

  • Dataset: Traffic sign photos for training and testing of the algorithm(s). 
  • Scripts for retiriving street-level photo sequences form Mapillary API as well as recognised traffic objects for training/comparison study are provided
  • Tools: PostgreSQL database with PostGIS extension (tutorial of installation and usage is provided), Python and relevant packages (e.g. Tensorflow)
  •  

Prerequisites

  • Programming: knowledge of a programming language (preferably Python) as the tasks are heavily dependent on implementing algorithms
  • Data Analysis/Image recognition : prior experience is desirable but not mandatory
  • Interest and willingness to work on a computer vision related topic
  • Ability to work independently

 

Contact: Stefania Zourlidou

Übersicht