Traffic Sign Detection in Complex Environment based on Multi-Scale Feature Enhancement and Group Attention

Since traffic light detection is essential for autonomous driving, it is studied intensively. However, traffic sign detection is difficult, especially in a complex environment. The traffic signs should be located first. Their unique features should be extracted next and fed into the classifier subsequently. In this paper, we adopt the current mainstream deep neural network-based object detection method for traffic sign detection. In our work, add specific environmental noise features to the dataset. A lightweight network, YOLOv4-Tiny, is chosen as the baseline network, and a multi-scale feature fusion module is designed to improve the performance of the network model. A lightweight group attention module is also designed. Experiments are carried out using the GTSDB dataset and the result shows the proposed model outperforms the other models in terms of precision and mAP.