D-FCOS: Traffic Signs Detection and Recognition Based on Semantic Segmentation

Considering the bright future of intelligent driving and automatic driving technology, many researchers pay attention to the significance of traffic signs detection and recognition. In this paper, a fast and accurate traffic signs detection and recognition model was proposed, which uses a detector similar to the fully convolution network and one stage method in semantic segmentation problem to localize the traffic signs in a pre-pixel fashion. Compared with other state-of-the-art object detection networks such as Faster R-CNN [1] and YOLOv3 [2], the semantic segmentation network proposed in this paper does not need predefined anchor, so it completely avoids the complex calculation and super parameter settings related to anchor boxes, which greatly reduces the workload of training and accelerates the model inference procedure.