Mapping traffic signboard from mobile mapping systems using deep learning approach

Pei Cheng Chen, Tee Ann Teo

Research output: Contribution to conferencePaper

Abstract

The high definition map (HD map) requires precise traffic sign information. For example, the OpenDRIVE standard requires the attribute of traffic sign such as type, subtype, location, height, and orientation. The Mobile Mapping System (MMS) collects color images, lidar 3D point cloud, and trajectory from positioning and orientation system (POS). To fulfill the needs of traffic signboard information for HD map, the lidar point clouds may provide geometrical information such as location, height, orientation while the color images may provide semantic meaning about traffic sign. The objective of this study is to develop an automatic framework to generate traffic signboard information for HD map. Most studies utilized either color images or lidar point clouds for traffic signboard detection. As the fusion of images and 3D point clouds may obtain both geometric and semantic information, this study adopts a data fusion approach to fuse the information from color image and lidar point clouds for traffic signboard recognition. The proposed method includes three steps. The first step detects the traffic signboard using lidar intensity. The traffic signboard is covered by a highly reflective surface. Therefore, the lidar intensity for traffic signboard is generally higher than other objects. The geometrical relationship between lidar and color image is based on the exterior orientations from the trajectory. The second step utilizes the exterior orientations of the color image to estimate the location of traffic signboard in image space. The traffic signboard in image space is classified by a modified VGG network to obtain the name of signboard among the 78 classes in Taiwan's traffic system. The last step extracts the geometrical parameters such as location, height, the orientation of traffic signboard from lidar point. The experimental results indicate that the attributes of traffic signboards can be extracted automatically.

Original languageEnglish
StatePublished - 1 Jan 2020
Event40th Asian Conference on Remote Sensing: Progress of Remote Sensing Technology for Smart Future, ACRS 2019 - Daejeon, Korea, Republic of
Duration: 14 Oct 201918 Oct 2019

Conference

Conference40th Asian Conference on Remote Sensing: Progress of Remote Sensing Technology for Smart Future, ACRS 2019
CountryKorea, Republic of
CityDaejeon
Period14/10/1918/10/19

Keywords

  • Deep learning
  • High definition map
  • Mobile mapping system
  • Traffic sign

Fingerprint Dive into the research topics of 'Mapping traffic signboard from mobile mapping systems using deep learning approach'. Together they form a unique fingerprint.

  • Cite this

    Chen, P. C., & Teo, T. A. (2020). Mapping traffic signboard from mobile mapping systems using deep learning approach. Paper presented at 40th Asian Conference on Remote Sensing: Progress of Remote Sensing Technology for Smart Future, ACRS 2019, Daejeon, Korea, Republic of.