Small Traffic Sign Detector in Real-time Based on Improved YOLO-v4

As one of the indispensable parts of the smart transportation system, traffic sign detection is still an open challenge for the reason that traffic sign detection, one kind of small object detection, requires high precision, high reliability, and real-time speed. In this paper, a real-time detector for small traffic sign improved from YOLO-v4 is proposed. First, a transpose convolution layer is added into the backbone of YOLO-v4, named CSPDark-53, to get larger feature maps with richer semantic information of small objects. Secondly, aiming at fusing multi-scale feature maps more effectively, we present two novel attention modules, the upsampling attention module and the downsampling attention module, to calculate the channel weights and spatial masks to help the network focus on useful features and suppress invalid features. The experiments show our detector outperforms the compared models in Chinese Traffic sign dataset TT100k.