AUT-UTP: Urban Traffic Panel detection and recognition dataset

Recently, detection and recognition of traffic panels andtheirtextualinformationisstudiedincreasinglytobecomethe next working part of driver assistance systems and autonomous cars. These information are especially useful when other facilities fail to provide enough information about routes and places, like when Global Positioning System (GPS) gets blocked in high density urban areas. However, previous researches in the field are only on road panels, which are in much more simpler conditions than the ones in high density urban areas. As there is no such dataset available to address this problem in real world application, like cluttered urban scenarios, this paper presents an extensive dataset of videos recorded on a moving vehicle from traffic panels in urban areas with ground truth data refined by human observation. This paper presents an analysis on quality and quantity of the data and provides a statistical report. Video sequences are categorized according to their different characteristics so that researchers can choose the part of data which benefits their research goals. Ground truth data and characteristics of each scenario are produced using semi-automatic methods, followed by more than 10000 times of manual refinement for bounding boxes around the panels, also for both English and Persian text inside them.