FONT SIZE
S
M
L
COLOR
Normal
1
2
3

FRMO Top

Home > Research > Research Results > Research Results 2020 > Strip roads can be identified using Artificial Intelligence (AI)

Update:May 1, 2020

Main content starts here.

Strip roads can be identified using Artificial Intelligence (AI)

 

Article title

Strip road detection with semantic segmentation based on deep convolutional neural networks. (in Japanese with English summary)

Author (affiliation)

Kengo Usui (a), Masahiro Mozuna (a)

(a) Department of Forest Engineering, FFPRI, Tsukuba, Ibaraki, Japan.

Publication Journal

ournal of The Japan Forest Engineering Society, 35(1), 7-13, The Japan Forest Engineering Society, Jan. 2020 DOI:10.18945/jjfes.35.7( External link )

Content introduction

There is demand to automate logging equipment in order to improve productivity and resolve labor shortages in the forestry industry, In order for this equipment to safely travel through strip roads to do forestry work, it is necessary to detect the area that can be traveled (road surface). However, compared with regular roads that have clearly defined boundaries, strip roads in forests merge with natural elements, making it difficult to detect road surfaces. In the present study, a method for detecting road surfaces was developed with AI using still images and a deep learning system.

First, still images were created for specific time intervals from video images that had been taken from the front of a forestry machine while it was traveling. Next, the road surface in the still images was identified visually, and ground truth label images were created. The images that were extracted here were subjected to image processing such as rotation, color changing, and so on, and then they were entered into a computer, where the road surface patterns were subjected to AI learning. Photos that were not used in this processing were used as control images to make evaluations and verify the effectiveness of the AI learning.

The control images were read into the learning computer, and each pixel (image element) was classified as either road surface or background. With this method, the road surface could be detected with an accuracy of 96.7 to 97.5%. In this study, the method that was developed for recognizing still images that is needed for automatic traveling of logging equipment will become technology that will contribute greatly to the future automation of forestry machinery.

 

Photos: (a) View of an actual strip road 

Photos: (a) View of an actual strip road, (b) ground truth label image of the road surface, and (c) strip road detected with deep learning. The greater the resemblance between (b) and (c), the greater the accuracy of strip road detection.