Research On Feature Extraction And Recognition Of Single Wood From Laser Point Cloud For Tending And Harvesting In Forest

The existing planted forest area in China is the largest in the world,but the growth quality of planted forests is very poor,requiring reasonable tending and logging.At present,there have been independent research and development of forest combine harvester in China,but the selection of harvesting targets still relies on manual selection,which is inefficient and dangerous.In this paper,the feature extraction and recognition of single wood in tending and harvesting were taken as the research object,and the 3D point cloud data of harvesting area were obtained by using Li DAR.Finally,the picking and breeding targets in the experimental plots were selected through the point cloud intelligent processing method,and the intelligent perception of the forest area of the picking and breeding operations was realized.The main research contents of this paper are as follows:(1)A backpack-type point cloud collection experiment system was built,and the point cloud data used in the experiment was obtained and preprocessed to provide data support for subsequent research.The open source Loam-SLAM algorithm was implemented in the robot operating system environment,and then the point cloud data of different areas(campus,park plantations,experimental forest farms,etc.)were obtained,including long standing trees,multi-branched standing trees,and low shrubs,buildings,flat land,and sloped land.The collected data were processed by blocks,and 50 blocks of experimental plot data were extracted for filtering and downsampling.(2)It improved the semantic segmentation deep learning network suitable for plantation forests,and produced a forest area semantic segmentation data set for harvesting targets.A dedicated point cloud semantic segmentation data set was made according to the 50 sample plot data and the selection requirements of the collection and breeding targets.A total of 200 data samples were obtained through optimization,and then divided into training sets and test sets,and their corresponding sample numbers were respectively 120 and 80.Two deep learning methods,Piont Net and Point Net++,were used to carry out research respectively.The accuracy rates of natural forest environment recognition reached 62.4%and 70.8%,and the accuracy rates of artificial forest reached 72.5% and 82.2%.An improved RPN-Point Net++ algorithm was proposed.The detection accuracy rate in natural forests was not significantly improved,but the accuracy rate of harvesting and breeding features in artificial forests was increased by 7.9% to 90.1%.(3)The task of ground filtering,single wood segmentation,and stem extraction of the forest point cloud data of the sample plot was realized.By obtaining the parameters such as tree height,diameter at breast height and plant density,a decision tree model for selection of breeding targets was constructed.The method of regional growth and Euclidean distance was used to segment the sample plot,and the trunk of the tree was extracted based on the axial distribution density and axial similarity method.Three methods such as least square fitting method,Hough transform fitting method,and particle-based improved RANSAC fitting method were used to calculate and extract the diameter at breast height of standing trees,and the forest density of the plot was divided based on the DBSCAN clustering algorithm.Finally,a decision tree model for the selection of harvesting targets was constructed.(4)A method for extracting and breeding target features based on the backbone curve was proposed,and the volume was estimated using point cloud data.Through the layer-by-layer fitting method,the trunk shape was studied and the spatial curve parameter equation of the trunk shape was obtained.The spatial straightness,curvature and deflection characteristics of the trees on the sample plot were get,and the spatial comprehensive evaluation method for selection of harvesting targets was established.Based on the point cloud data,the tree trunk taper equation parameters are estimated,and the forest timber volume is estimated according to the round table fitting method,which is of great significance to the planning of forest harvesting and production.The research results of this paper prove that,relying on the point cloud data of forest areas,the intelligent selection of harvesting targets is of great significance to forest harvesting operations,which lays a foundation for in-depth research on intelligent forest machinery.