Automatic tree classification method based on laser scanning 3D point cloud based on deep learning

Abstract

The invention discloses an automatic tree classification method based on deep learning based on laser scanning three-dimensional point cloud, which can automatically classify extracted single trees, and the classified single trees include trunk and crown structures. The present invention adopts the method of rotating side projection, and model training can still be carried out effectively in the case of few tree data, and due to the adoption of a normalized preprocessing method, the disadvantage of uneven distribution of point cloud data density (distance from the scanner) is overcome , making the result less affected by the acquisition equipment and more stable. In addition, the accuracy of automatic classification of a variety of trees is improved due to the use of deep learning for model training. The present invention adopts eigenvectors as units for calculation, has fast calculation speed, is more suitable for large-scale point cloud scenes, and has practical significance and application value.

Description

technical field [0001] The invention relates to the field of spatial information processing, in particular to an automatic tree classification method based on deep learning for laser scanning three-dimensional point clouds. Background technique [0002] Traditional forest surveys are mainly manual surveys, which have the following disadvantages: 1. Due to the lack of clear signs in the forest, it is difficult to determine the sample plots to be surveyed, and it is often necessary to rely on old farmers to lead the way, which is time-consuming and laborious. 2. During the investigation, the investigators have to go deep into the dense forest and face natural dangers such as landslides, poisonous snakes and insects. 3. The terrain in the forest is complex, and some trees even grow on very steep slopes. Manual measurement will cause errors or even fail to measure. [0003] The use of laser scanning technology for forest surveys can effectively solve many defects in artificial ...

AI Technical Summary

Problems solved by technology

[0002] Traditional forest surveys are mainly manual surveys, which have the following disadvantages: 1. Since there are no clear signs in the forest, it is difficult to determine the sample plots to be surveyed, and often need to rely on old farmers to lead the way, which is time-consuming and laborious

2. During the investigation, the investigators have to go deep into the dense forest and face natural dangers such as landslides, poisonous snakes and insects

3. The terrain in the forest is complex, and some trees even grow on very steep slopes. Manual measurement will cause errors or even encounter situations that cannot be measured

[0007] (1) Directly classify the extracted features of 3D objects. Although the information of the original data can be more fully utilized, the 3D objects are multiplied compared to the features contained in the image.

Direct extraction of three-dimensional object features will cause the unfavorable situation of feature dimension explosion, resulting in the consequence that the amount of data is too large to train or the training time is too long;

[0008] (2) The 3D grid data has the advantage of high integrity. Most methods can achieve good classification results on this data set, but they are not necessarily suitable for point cloud data;

[0009] (3) When using the method of projecting three-dimensional data to an image, the projection surface is too small, and the projection angle is relatively single, so the projected information cannot be fully utilized

Images