Semantic segmentation of road furniture in mobile laser scanning data

Road furniture recognition has become a prevalent issue in the past few years because of its great importance in smart cities and autonomous driving. Previous research has especially focussed on pole-like road furniture, such as traffic signs and lamp posts. Published methods have mainly classified road furniture as individual objects. However, most road furniture consists of a combination of classes, such as a traffic sign mounted on a street light pole. To tackle this problem, we propose a framework to interpret road furniture at a more detailed level. Instead of being interpreted as single objects, mobile laser scanning data of road furniture is decomposed in elements individually labelled as poles, and objects attached to them, such as, street lights, traffic signs and traffic lights. In our framework, we first detect road furniture from unorganised mobile laser scanning point clouds. Then detected road furniture is decomposed into poles and attachments (e.g. traffic signs). In the interpretation stage, we extract a set of features to classify the attachments by utilising a knowledge-driven method and four representative types of machine learning classifiers, which are random forest, support vector machine, Gaussian mixture model and naive Bayes, to explore the optimal method. The designed features are the unary features of attachments and the spatial relations between poles and their attachments. Two experimental test sites in Enschede dataset and Saunalahti dataset were applied, and Saunalahti dataset was collected in two different epochs. In the experimental results, the random forest classifier outperforms the other methods, and the overall accuracy acquired is higher than 80% in Enschede test site and higher than 90% in both Saunalahti epochs. The designed features play an important role in the interpretation of road furniture. The results of two epochs in the same area prove the high reliability of our framework and demonstrate that our method achieves good transferability with an accuracy over 90% through employing the training data of one epoch to test the data in another epoch.