| With the continuous deepening of human society’s understanding of the natural environment,contemporary science has paid more and more attention to the sustainable development of the natural environment while maintaining social development,so as to maintain the balanced relationship between mankind and nature.Among them,the detection of trees and vegetation in the wild is very important.Closely analyzing the dynamic changes of trees and vegetation can help humans better deal with issues of environment.With the vigorous development of lidar technology and deep learning algorithms,humans have begun to try to use these technologies to complete tree detection tasks.However,the large amount of data inherent in the 3D point cloud of outdoor large scenes and the characteristics of trees with very complex and sharp features have brought huge obstacles to the innovation and development of this task.Therefore,this article focuses on the segmentation of 3D point clouds in large scenes,and the following work is done:(1)In order to improve the efficiency of the network in processing large-scale 3D point clouds,a random sampling algorithm with a time complexity of O(1)is used when extracting key points.In order to compensate for the loss of key information caused by random sampling algorithms,this paper adopts a multi-level feature extraction method to continuously increase the receptive field of local neighborhoods,so hold as much key feature information as possible.(2)This article focuses on tree detection,that is,it only cares about the segmentation accuracy of the tree object in the prediction result.Aiming at the characteristics of tree objects with a wealth of sharp information on corners,this paper extracts the normal vector information of point cloud objects when training the deep learning network to enrich the data features,so as to provide support for the final prediction results.(3)In order to facilitate the work of tree detection personnel,this paper encapsulates the proposed algorithm into a tree detection system.While the system provides preset algorithm models,it also supports user-defined hyperparameters to train models that are more in line with their needs.The tree detection system is mainly implemented in java language,which provides important support for the efficiency of the entire processing flow. |
