Point-of-interest (POI) data contains rich semantic and spatial information, having a wide range of applications including land use, transport planning and driving navigation. However, urban POI mapping traditionally requires a lot of manpower and material resources, which only few institutions or enterprises can afford to. With the increasing amount of street-level imagery, it is possible to directly extract POI-related information from such data and automatically map the distribution of urban POIs. In the pilot study, we mainly focused on extracting POIs from billboards in street-level imagery. Firstly, the you only look once (YOLO) algorithm was considered to locate billboards in the imagery, then an optical character recognition (OCR) model was adopted to extract POI-related semantic information from the detected billboard, and finally the extracted semantic text was further processed to obtain POI results. The preliminary study shows that it is a promising way of mapping urban POIs from crowdsourced street-level data using deep learning techniques.