Instance Segmentation of Trees in Urban Areas from MLS Point Clouds Using Supervoxel Contexts and Graph-Based Optimization

Abstract

In this paper, an instance segmentation method for tree extraction from MLS data sets in urban scenes is developed. The proposed method utilizes a supervoxel structure to organize the point clouds, and then extracts the detrended geometric features from the local context of supervoxels. Combined with the detrended features of the local context, the Random Forest (RF) classifier will be adopted to obtain the initial semantic labeling results of trees from point clouds. Afterwards, a local context-based regularization is iteratively performed to achieve global optimum on a global graphical model, in order to spatially smoothing the semantic labeling results. Finally, a graph-based segmentation is conducted to separate individual trees according to the semantic labeling results. The use of supervoxel structure can preserve the geometric boundaries of objects in the scene, and compared with point-based solutions, the supervoxel-based method can largely decrease the number of basic elements during the processing. Besides, the introduction of supervoxel contexts can extract the local information of an object making the feature extraction more robust and representative. Detrended geometric features can get over the redundant and in-salient information in the local context, so that discriminative features are obtained. Benefiting from the regularization process, the spatial smoothing is obtained based on initial labeling results from classic classifications such as RF classification. As a result, misclassification errors are removed to a large degree and semantic labeling results are thus smoothed. Based on the constructed global graphical model during the spatially smoothing process, a graph-based segmentation is applied to partition the graphical model for the clustering the instances of trees. The experiments on two test datasets have shown promising results, with an accuracy of the semantic labeling of trees reaching around 0.9. The segmentation of trees using graph-based algorithm also show acceptable results, with trees having simple structures and sparse distributions correctly separated, but for those cramped trees with complex structures, the points are over- or under-segmented.