Glare sign recognition based on improved Yolox algorithm
Teng Xu, Ling Ren, Tian-Wei Shi, Yuan Gao, Jian-Bang Ding, Rong-Chen Jin · 2023
Traffic sign recognition is one of the most critical aspects of autonomous driving. To improve detection accuracy without increasing the number of model parameters too much, this paper uses Conditionally Parameterized Convolutions (condconv) to replace the convolutional blocks in Yolox-s. To understand the transfer of feature maps within the deep learning network, a gradient mapping using Gradient-weighted Class Activation Mapping (Grad-CAM) is used to generate a heat map. As the existing traffic sign dataset does not include data for glare situations, data augmentation was used to simulate traffic signs in a glare context. Experiments show that the proposed improvements can improve detection accuracy (+2.3%) with only a small increase in model parameters $(+0.1 \mathrm{M})$. The effectiveness of the proposed data augmentation method is demonstrated by using the trained model for real-world detection.