Skip to main navigation Skip to search Skip to main content

A multi-layer CRF based methodology for improving crowdsourced street semantics

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

This paper presents an intuitive and novel method for improving the semantic quality of streets in crowdsourced maps. Two factors negatively affecting the quality are incorrect and ambiguous semantics. Toward overcoming these, a multi-layer CRF based model is proposed that performs a simultaneous hierarchical classification of streets into fine-grained (crowdsourced; therefore, rich but ambiguous) and coarse-grained (familiar and standard) semantics. Inference is performed using Lazy Flipper algorithm which is fast for street network consisting of several hundred thousand streets. The model achieves a classification accuracy of 61% for fine-grained classification and 77% for coarse-grained classification respectively.

Original languageEnglish
Title of host publicationIWCTS 2018 - Proceedings of the 11th ACM SIGSPATIAL International Workshop on Computational Transportation Science
PublisherAssociation for Computing Machinery (ACM)
Pages29-38
Number of pages10
ISBN (Electronic)9781450360371
DOIs
Publication statusPublished - 6 Nov 2018
Event11th ACM SIGSPATIAL International Workshop on Computational Transportation Science, IWCTS 2018 - Seattle, United States
Duration: 6 Nov 2018 → …

Publication series

NameIWCTS 2018 - Proceedings of the 11th ACM SIGSPATIAL International Workshop on Computational Transportation Science

Conference

Conference11th ACM SIGSPATIAL International Workshop on Computational Transportation Science, IWCTS 2018
Country/TerritoryUnited States
CitySeattle
Period6/11/18 → …

Keywords

  • Conditional Random Fields
  • Hierarchical Classification
  • Open-StreetMap
  • Semantics
  • Street Networks

Fingerprint

Dive into the research topics of 'A multi-layer CRF based methodology for improving crowdsourced street semantics'. Together they form a unique fingerprint.

Cite this