Small Traffic Sign Detection and Recognition in High-Resolution Images

Lei You, Yu Ke, Hongpeng Wang, Wenhu You, Bo Wu, Xinghao Song

Abstract: Traffic sign detection and recognition is a research hotspot in the computer vision and intelligent transportation systems fields. It plays an important role in driver-assistance systems and driverless operation. Detecting signs, especially small ones, remains challenging under a variety of road traffic conditions. In this manuscript, we propose an end-to-end deep learning model for detecting and recognizing traffic signs in high-resolution images. The model consists of basic feature extraction and multi-task learning. In the first part, a network with fewer parameters is proposed, and an effective feature fusion strategy is adopted to gain a more distinct representation. In the second part, multi-task learning is conducted on different hierarchical layers by considering the difference between the detection and classification tasks. The detection results on two newly published traffic sign benchmarks (Tsinghua-Tencent 100K and CTSD) demonstrate the robustness and superiority of our model.