MODELING OF ALGORITHMS FOR CLASSIFICATION OF OBJECTS BASED ON LASER SCANNING DATA

Background. The modern concept of the development of the agro-industrial complex is based on the implementation of high-precision autonomous robotic systems that are capable of carrying out precision monitoring of perennial plantations. The key task in this context is the creation of digital twins of gardens with intensive vegetation in real time. Laser scanning technology provides dense three-dimensional point clouds, however, the complexity of the architecture of tree crowns, the presence of technical structures and occlusion create significant obstacles to the automated interpretation of measured data. The need to develop robust mathematical models for semantic segmentation in such environments forms the relevance of this work.

Objective. The purpose of this work is to model the algorithmic basis for the classification of structural elements of the garden using laser scanning data.

Conclusions. The methodological basis of the work is a combination of RANSAC (Random Sample Consensus) methods for robust separation of the earth's surface plane from target objects and DBSCAN (Density-Based Spatial Clustering of Applications with Noise) for semantic grouping of vegetative mass and identification of vertical structures. For experimental verification of the algorithms, an open multi-scene HOPS dataset was used, containing point clouds (500,000 or more within each scene) of an intensive apple orchard with a high level of geometric complexity. Mathematical models were implemented in Python 3.10 using specialized Open3D and NumPy libraries. Critical areas of accuracy degradation were identified, which are associated with the morphological similarity of vertical branches to concrete supports and the influence of laser occlusions. The modeling results allow generating high-precision maps of leaf surface density and crown depth.