An Automatic Point-of-Interest Generation Method Based on Open Street View Images

Point-of-Interest data is of great significance for building geospatial data infrastructure and providing public service, but the collection and updating of POIs still heavily rely on manual work. In this study, we propose an automatic POI generation method based on Google street view images. Firstly, we define the signboard entity by detecting signboard object from street view images and recognizing text lines in signboards. Secondly, we align signboard entities using multi-modal features to reduce data redundancy. Then we propose POI-GLM to extract POI name from signboard entities. In order to obtain accurate POI positions, we calculate signboard 3D positions based on multi-view street view images. To verify the effectiveness of our method, we conduct experiments on Hong Kong street view images. The result shows our method can generate accurate POI dataset with the recall rate of 59.42 % and average distance error of 19.15m, exceeding the previous works.