This paper has described the adaptability of semantic segmentation using deep convolutional neural networks for detecting forest strip roads. We used fully convolutional networks, which are one of the most commonly used methods in semantic segmentation. We established our datasets by capturing videos of strip roads. We performed a 3-class classification of strip roads, their backgrounds, and the buffer between the roads and the backgrounds and achieved good overall accuracy (96.7–97.2%). Our findings showed that our approach can detect strip roads to the same level as in the previous research. Our precision for the 3classes of strip road, buffer, and background was 96.7–97.5%, 67.4–71.6%, and 98.5–98.6%, respectively. There was some difficulty in detection of the edge part of strip roads as a buffer class using the model in this study. However, our trained model showed good precision in detecting the strip road class. These findings suggest that our models might predict strip roads for local path planning when strip roads are sufficiently wide for machines to pass.
View full abstract