Testing different CNN architectures for semantic segmentation for landscaping with forestry robotics

Abstract

Increasingly mechanized, leading to the emergence of forestry robotics. In this article, we present the results of our evaluation of a set of state-of-the-art convolutional neural network-based solutions for semantic segmentation using the Bonnetal open-source training and deployment framework, together with a custom-made solution based on an adaptation of an alternative decoder and encoder for that framework, the Adapnet++–eASPP architecture, in the context of a robotic perception pipeline designed to perform landscaping in woodlands to reduce the amount of living flammable material (the Fuel class) for wildfire prevention. Results show that, overall, Adapnet++–eASPP was the most robust and comprehensive encoder for our application, demonstrating a consistently high average level of performance in comparison to the other architectures, and displaying the greatest robustness of the group. With this solution, we demonstrated to be able to satisfy our requirements of a low rate of false positives for the Fuel class and operational performance of 10fps.