Korean Traffic Sign Detection Using Deep Learning

In this paper, we present a new optimized architecture modified from YOLOv3 to detect three different classes of challenging Korean Traffic Sign Detection (KTSD) dataset. We optimized the new neural network called TS detector with denser grid size, and optimized anchor box size to detect prohibitory, mandatory, and danger classes of KTSD dataset. We trained this architecture on our Korean traffic sign dataset to achieve the mAP value of 86.61%. Our results are significantly better than original YOLOv3 and D-Patches algorithm in terms of mAP value and CPU time.