Computer vision research involving street and road detection methods usually focuses on driving assistance and autonomous vehicle systems. In this context, street segmentation occurs in real-time, based on images centered on the street. This work, on the other hand, uses street segmentation for urban planning research to classify pavement types of a city or region, which is particularly important for developing countries. For this application, it is needed a dataset with images from various locations for each street. These images are (...) not necessarily centered on the street and include challenges that are not common in street segmentation datasets, such as mixed pavement types and the presence of faults and holes on the street. We implemented a multi-class version of a state-of-the-art segmentation algorithm and adapted it to perform street pavement classification, handling navigation along streets and angle variation to increase the accuracy of the classification. A data augmentation approach is also proposed to use preliminary results from the test instances as new ground truth and increase the amount of training data. A dataset with more than 300,000 images from 773 streets from a Brazilian city was built. Our approach achieved a precision of 0.93, showing the feasibility of the proposed application. (shrink)