Semantic and Depth Learning for Autonomous Forest Mapping

• Guilherme V. Nardari USP
• Roseli A. F. Romero USP

Resumo

While overhead data can provide general information about a forest, inside the forest we can identify understory plants, measure the diameter and count the trunks of each tree and check the number of dry leaves on the ground, which is an indication of the risk of fire. To get these measurements, we rely on human expeditions, which can be slow, expensive, and even dangerous. For this reason, robots that can autonomously navigate and extract data from inside the forest could revolutionize the way to monitor forests around the world and the amount of information we have about them. With no reliable GPS signal, uneven terrain covered by plants and leaves, and trees with branches moving with the wind, classic algorithms developed with urban environments in mind may fail in forests as their assumptions may not be valid in this environment. Semantic information such as classes and forms of objects expected in the environment can be a promising way to increase the robustness and performance of autonomous systems. In this paper, we propose a learning-based approach to the stereo depth estimation problem that simultaneously outputs semantic labels and stereo disparity estimates. This method can obtain pseudo-Light Detection and Ranging (LiDAR) readings that can be incorporated into localization and mapping frameworks to create a semantic map of a forest. The proposed method balances between cost map resolution and speed while learning two tasks. Simultaneously, our architecture maintains an inference speed performance acceptable for real-time semantic mapping. Several experiments in simulated forests are performed, demonstrating that our method achieves a good balance between inference speed and accuracy that would be suitable for real-time semantic mapping.

Palavras-chave: Localization, Mapping, Semantic Segmentation, Machine Learning