Volume XLII-2
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2, 731-738, 2018
https://doi.org/10.5194/isprs-archives-XLII-2-731-2018
© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2, 731-738, 2018
https://doi.org/10.5194/isprs-archives-XLII-2-731-2018
© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 4.0 License.

  30 May 2018

30 May 2018

CLASSIFICATION OF POLE-LIKE OBJECTS USING POINT CLOUDS AND IMAGES CAPTURED BY MOBILE MAPPING SYSTEMS

Y. Mori, K. Kohira, and H. Masuda Y. Mori et al.
  • The University of Electro-Communication, 1-5-1 Chofugaoka, Chofu, Tokyo 182-8585, Japan

Keywords: Mobile Mapping System, Point-Cloud, Pole-Like Object, Machine Learning, Convolutional Neural Network

Abstract. The vehicle-based mobile mapping system (MMS) is effective for capturing 3D shapes and images of roadside objects. The laser scanner and cameras on the MMS capture point-clouds and sequential digital images synchronously during driving. In this paper, we propose a method for detecting and classifying pole-like objects using both point-clouds and images captured using the MMS. In our method, pole-like objects are detected from point-clouds, and then target objects, which are objects attached to poles, are extracted for identifying the types of pole-like objects. For associating each target object with images, the points of the target object are projected onto images, and the image of the target object is cropped. Each pole-like object is represented as a feature vector, which are calculated from point-clouds and images. The feature values of a point-cloud are calculated by point processing, and the ones of the cropped image are calculated using a convolutional neural network. The feature values of point-clouds and images are unified, and they are used as the input to machine learning. In experiments, we classified pole-like objects using three methods. The first method used only point-clouds, the second used only images, and the third used both point-clouds and images. The experimental results showed that the third method could most accurately classify pole-like objects.