A Robust Real-Time Anchor-Free Traffic Sign Detector With One-Level Feature

Traffic sign detection has a wide range of applications. In the wild, the existing detectors are difficult to balance the requirements for robustness, real time, and accuracy. In this paper, we propose a one-level feature anchor-free traffic sign detector based on adaptive image enhancement and explicit objectness branch. Firstly, aiming to weather variations in outdoor scenes, we design an Image Enhancement Module to achieve strong robustness, including exposure, tone, and luminance three filters. We integrate a small convolutional network to predict filter parameters, which is jointly trained end-to-end and does not exist during inference. Secondly, we propose a Feature Encoder which consists of a One-Level Feature Extractor and a Receptive Fields Expander. In One-Level Feature Extractor, we introduce the Multilayer Perceptron (MLP) Block and design a Feature Aggregator (FA) to fully fuse feature maps of different resolutions. Receptive Fields Expander is used to make up for the deficiency of one-level feature map and generates output features with various receptive fields. Finally, to improve the accuracy based on the pixel-wise prediction, we train an explicit objectness branch of the detection head and improve the dynamic selection strategy of positive and negative samples. We weight Generalized Intersection over Union (GIoU) and Smooth L1 to propose a new localization loss. During inference, we multiply the predicted objectness value by the classification score as the final classification confidence to alleviate sample imbalance. To comprehensively evaluate the proposed method, we conduct experiments on GTSDB, TT100 K and CCTSDB 2021 datasets, and all achieve state-of-the-art results.