HeLoDL: Hedgerow Localization Based on Deep Learning

Abstract

Accurate localization of hedges in 3D space is a key step in automatic pruning. However, due to the irregularity of the hedge shape, the localization accuracy based on traditional algorithms is poor. In this paper, we propose a deep learning approach based on a bird’s-eye view to overcoming this problem, which we call HeLoDL. Specifically, we first project the hedge point cloud top-down as a single image and, then, augment the image with morphological operations and rotation. Finally, we trained a convolutional neural network, HeLoDL, based on transfer learning, to regress the center axis and radius of the hedge. In addition, we propose an evaluation metric OIoU that can respond to the radius error, as well as the circle center error in an integrated way. In our test set, HeLoDL achieved an accuracy of 90.44% within the error tolerance, which greatly exceeds the 61.74% of the state-of-the-art algorithm. The average OIoU of HeLoDL is 92.65%; however, the average OIoU of the best conventional algorithm is 83.69%. Extensive experiments demonstrated that HeLoDL shows considerable accuracy in the 3D spatial localization of irregular models.