2021 Volume 23 Issue 2 Pages 369-376
The degradation of wetlands caused by the overgrowth of aquatic plants is a problem in many areas; therefore, low-cost, labor-saving vegetation management using robot boats is under development. Herein, we propose a method to classify aquatic plants and obstacles by real-time image processing to enable the full-autonomous operation of a robot boat. We adopted a semantic segmentation method using deep learning for the image processing and conducted teaching and testing on our own dataset. Regarding classification by segmentation, it was possible to classify the lotus (Nelumbo nucifera), which is the target of harvesting, and a rare species of fringed water lily (Nymphoides peltata), with an accuracy rate of about 90%. The deep neural network could learn the color, shape, and position information in the image. This reduced the influence of external environmental disturbances such as changes in illuminance and harsh sunlight reflections on the water surface, and the inference was robust. Regarding the obstacles, it was possible to classify with an accuracy rate of 80% or more within the depth-obtainable range of an RGB-D camera. It was also confirmed that classification was possible even when the obstacle was located 4.0 m or more away.
