BushNet: Effective semantic segmentation of bush in large-scale point clouds

Abstract

Effective and robust semantic segmentation of bush is the fundamental problem of agroforestry environment understanding. However, the point cloud data of most large-scale agroforestry scenes is extremely large, and it is difficult to perform semantic segmentation on them. In order to realize the effective semantic segmentation of bush point cloud in large-scale agroforestry environment, this paper proposes BushNet, a novel point cloud segmentation network consists of three key components. Firstly, we propose the minimum probability random sampling module which can quickly and randomly sample a huge point cloud while avoiding the problem of random sampling easily causing re-sampling, reducing the consumption of computing resources and improving the convergence speed. Secondly, we propose the local multi-dimensional feature fusion module which makes the network more sensitive to bush point cloud features, thereby showing better bush segmentation performance. Thirdly, we propose the multi-channel attention module to achieve more accurate attention distribution and improved training efficiency. Experiments demonstrate that our approach significantly improves segmentation performance on multiple large-scale agroforestry point cloud data sets.