Symbol Detection with Deep Learning

Abstract

This study shows the feasibility of detecting some symbols from a distance. Therefore, it will be possible to use these symbols as identification of things, e.g. people, vehicles, objects, information labels etc. There is such a necessity especially for construction sites to know the rough location of workers. In order to achieve this goal, firstly some two-dimensional symbols that are distinguishable from each other by both human eye and computer are generated. Afterwards, they are put on a three dimensional surface and annotated. Lastly, the labeled data are trained by the deep learning method YOLO version 5. Results indicate that it is highly efficient to use the chosen symbols for recognizing from a distance. For now, up to 10 meter images are taken into account. Real time tests are taken and the accuracy is about 85%. As the dataset will be combined with real-life images, the accuracy results will be even higher.