Image-Based Scene Recognition for Robot Navigation Considering Traversable Plants and Its Manual Annotation-Free Training

This paper describes a method of estimating the traversability plant parts covering path and navigating through them for mobile robots operating in plant-rich environments. Conventional rely on scene recognition methods that consider only geometric information environment. Those methods, therefore, cannot recognize paths as traversable when they are covered by flexible plants. In this paper, we present novel framework image-based to realize navigation such Our model exploits semantic segmentation branch general object classification estimation pixel-wise traversability. The is trained using an unsupervised domain adaptation with label images generated from robot's traversal experience during data acquisition phase, coined masks. training procedure entire is, free manual annotation. our experiment, show proposed capable distinguishing plants more accurately than conventional non-traversable classes, existing method. We also conducted real-world experiment confirmed robot successfully navigated