CNN-based Tree Model Extraction

We propose a method for the segmentation and structural reconstruction of tree stems and branches in cluttered environments. We use single images of monocular cameras and convolutional neural networks for the segmentation, centerline, and contour detection of trees (trunks and branches), and a deterministic approach to build tree model graphs, defining the positions of vertices and edges, on the points of segments' centerlines. Compared to previous stochastic methods, this approach, tested on a synthetic dataset, gives better segmentation accuracy with a significantly smaller computational complexity.

I. Introduction

The detection, segmentation, and structural analysis of piecewise linear or curvilinear objects such as road networks, blood vessels, and trees are important tasks in the field of computer vision. The detection of tree parts is a fundamental question in many agricultural applications such as pruning by robots or trees state monitoring in forests and city parks.