Tree detection in urban environments is an essential stage for applications in urban mapping, city management, and forest inventory. In recent decades, airborne LiDAR data have gained prominence owing to their ability to represent objects with high geometric quality. In this paper, we propose a straightforward geometry-based tree detection approach exploring the local omnivariance feature and K-means clustering. The main contribution can be summarized as an automatic tree detection approach based on point cloud geometry, which does not require sample data for training, as in machine-learning approaches. In addition, the approach uses a set of parameters/thresholds that can be intuitively defined. The experiments were conducted across three datasets with distinctive urban characteristics. Obtained results indicate that the proposed approach has strong potential for detecting trees from airborne LiDAR data with different point densities and acquired in urban environments of varying complexities, yielding an average Fscore of around 95%. Compared with related approaches, the results are similar to machine learning-based strategies and, generally, resulted in superior performance in terms of completeness, as evidenced by the low occurrence of omission errors.
Santos et al. (Mon,) studied this question.