Recent advances in computer vision segmentation for urban tree detection using street view imagery: a literature review

Abstract

Segmentation of trees in urban settings is important for tracking vegetation and managing green infrastructure in urban areas. However, there are numerous factors that make it difficult to obtain precise results, such as differences in tree species, season, and environmental variables. Street-view imagery can provide a valuable database to be used for the detection and identification of trees. Unfortunately, the process of labelling images manually is time-consuming and inconsistent. In recent years, deep learning models have provided better automation and precision. However, difficulties in obtaining results persist due to factors such as occlusion, background, and changes in lighting conditions. Therefore, this paper presents an overview of the current trends in the use of deep learning for the detection and segmentation of trees. Specifically, the following aspects will be discussed in this work: preprocessing of images, semantic and instance segmentation, models based on deep learning, and transformers. This literature review also analyzes current datasets, evaluation measures, and applications, offering knowledge about existing challenges and future directions for further studies. The article is designed to be a source of reference for scientists and practitioners working on problems related to urban vegetation classification and computer vision.