Research On Point Cloud Segmentation Of Street Tree Target Based On Deep Learning

As an important part of urban landscape and ecosystem,street trees are of great significance for urban air purification,noise control and wind and sand prevention.At present,the management and maintenance of street trees are mostly carried out manually,including pruning of branches and leaves,watering and spraying of pesticides.Among them,pesticide spraying is particularly important for the pest control of street trees.The current variable-target spraying technology has been successfully applied in the process of agricultural pest control.How to extend the variabletarget spraying technology with high efficiency and low pollution to the pesticide application of street trees become the focus of research.Aiming at the difficulty of detecting the target point cloud of road trees in complex urbanization environment,which makes it difficult to promote the target variable spray technology,this paper proposes a segmentation method of road tree target point cloud based on deep learning,which provides basic data for road tree to target pesticide application.The specific research contents are as follows:(1)The color street point cloud image is constructed by using the depth,primary echo intensity and echo frequency information in the street point cloud data,and a street tree point cloud image dataset is established based on this.From the original street point cloud data to the street point cloud image,a series of transformations are required.First,a two-dimensional lidar mobile scanning device is used to collect the street point cloud data;then analyze the original street point cloud data to obtain the three-dimensional image of the measurement point.Coordinates,the intensity of the first three times and the number of echoes;then select the depth of the measurement point,the intensity of the first echo and the number of echoes as the R,G,B channels of the point cloud image to construct a color street point cloud image;finally,based on the color street point cloud image Build a street tree point cloud image dataset.In the color street point cloud image constructed by this method,the measurement points correspond strictly to the pixels.(2)A street tree target point cloud segmentation algorithm based on Mask R-CNN is proposed,which uses deep learning to automatically detect and segment street trees.Based on the established street tree image data set,the Mask R-CNN-based street tree target point cloud segmentation algorithm is trained and the experimental results are analyzed.The results show that the method can accurately segment the street tree from the target objects such as buildings,pedestrians,street lights,etc.,the average precision rate in the test set is 99.8%,the average recall rate is 99.4%,the average F1 score is 99.1%,and the average processing time per radar scan line is 36.364 ms.In order to improve the performance of this method,the pre-trained network parameters on the largescale image dataset COCO are transferred to the street tree target point cloud segmentation model based on Mask R-CNN using the transfer learning method for training,and the experimental results are compared with the non-transfer model.Comparative analysis.The results show that the transfer learning results are basically consistent with the performance of this method,and the improvement effect is limited.(3)In view of the importance of real-time performance for the segmentation of the street tree target point cloud,consider replacing the two-stage algorithm Mask R-CNN in the model with the one-stage algorithm YOLACT,and propose a street tree target point cloud segmentation model based on YOLACT.In order to effectively improve the performance of the model,considering the correlation between street tree images and point cloud images,a YOLACT-based street tree image instance segmentation model is trained.Migrate the parameters of the instance segmentation model,train the YOLACT street tree target point cloud segmentation model for secondary migration,and obtain an average precision rate of 98.1%,an average recall rate of 99.0%,an average F1 score of98.6% and scanline processing speed.It is the experimental result of 12.1ms.Compared with the one-time migration YOLACT’s street tree target point cloud segmentation model,the performance is close to the one-time migration of the Mask R-CNN street tree target point cloud segmentation,and the detection speed is greatly improved.