Vehicle-mounted point cloud clustering method based on context characteristics and graph cut algorithm

Abstract

The invention provides a vehicle-mounted point cloud clustering method based on context characteristics and a graph cut algorithm, and the method comprises the steps of segmenting the point cloud to form the super voxels, segmenting the point cloud data according to the spatial density connection to form different connection regions, wherein each connection region becomes the super voxel; calculating the context correlation characteristics among the the super voxels, wherein the context correlation characteristics among the super voxels comprise the spatial correlation characteristics and thesemantic correlation characteristics, the spatial correlation comprises the direction, distance and topology aspects, the semantic correlation comprises the dimension and shape aspects, and the multi-mark graph cut cluster comprises forming a graph model formed by taking the super voxels as nodes and the connecting lines among the supervoxels as edges, evaluating the connectivity of the super voxels integrally by using the graph cut algorithm, and recombining to form the new point cloud cluster. According to the method, the priori knowledge is not needed, the over-segmentation phenomenon in point cloud clustering can be effectively improved, the precision of a point cloud clustering result is greatly improved, and the high-quality basic data is provided for the subsequent target recognition.

Description

technical field [0001] The invention relates to a vehicle-mounted point cloud clustering method, in particular to a vehicle-mounted point cloud clustering method based on context features and a graph cut algorithm. Background technique [0002] The mobile vehicle-mounted scanning system provides a quick way to acquire urban street scene cloud data. Accurately analyzing and interpreting scene objects from a large number of unordered point cloud data is the goal of point cloud clustering and classification research. The accuracy of point cloud clustering directly affects the results of subsequent classification, information extraction, and target geometric model reconstruction. The features used for clustering in existing point cloud clustering methods are usually calculated based on points or voxels, and point-based features will Affected by the quality of point cloud data such as uneven density, outliers, and noise, voxel-based features will be affected by the discretization...

AI Technical Summary

Problems solved by technology

The constraints of clustering gradually shift from the local scope to the consideration of the connectivity between data under global optimization conditions. Global optimization models such as conditional random fields and graph cuts have been used for point cloud data clustering, where the edges expressing the correlation between nodes Weight determination is very important. In existing algorithms, the distance constraints between point neighborhoods are often used to define the weight of edges, but the correlation between data points is not considered.

Images