Detection and Classification of Vegetation Growth in Public Spaces

Abstract

This master's dissertation addresses the problem of automatic detection and classification of vegetation growth in public spaces using computer vision methods. The research aims to develop AI models capable of analyzing street-view images in the visible spectrum to identify vegetation and develop methods to classify unwanted vegetation. Problematic vegetation includes cases where growth compromises safety, damages infrastructure, or deteriorates the aesthetic appearance of public areas. Two datasets with street-view images, Auto Arborist and Cityscapes, were utilized for training models to achieve this objective. Modern deep learning models, YOLOv8 and detectron2, were employed for vegetation detection and segmentation. Detectron2 image segmentation task achieved a high accuracy of 0.95 in vegetation detection. For solving the unwanted vegetation classification task, it has been hypothesized that pixels of unwanted vegetation are located next to pixels of objects such as walls, roads, sidewalks, traffic lights, and road signs. Two approaches to classifying vegetation as good or unwanted based on this hypothesis and using the detectron2 model were tested in different images. All detection and classification methods were tested on street-level images from Funchal, Portugal. The conclusion discusses future work and improvements, such as adaptation of the hypothesis to the geographical location of the target images, collection and labeling of a custom dataset for vegetation detection and classification, and improvement of the evaluation of the performance of the classification method.