This paper addresses the challenges of real-time traffic sign recognition in diverse Indian environments, where low-resolution images and high-speed autonomous vehicles can lead to blurred and obscured traffic sign boards. We propose using a YOLOv8 model as a base and enhancing its performance by incorporating SPD-Conv and NAM layers. SPD-Conv, a new CNN building block, replaces strided convolution and pooling layers with a space-to-depth layer followed by a non-strided convolution layer, enhancing the preservation of fine-grained information and improving feature representations. NAM, a Normalization-based Attention Module, is added to the bottleneck layer of the YOLOv8 model to enhance training efficiency and reduce model complexity, ensuring high accuracy in lightweight models. Our approach, which includes a sample dataset of 11,200 images from 52 different classes based on Indian Road Congress(IRC) standards and guidelines, demonstrates significant improvements in real-time traffic sign recognition, achieving a mAP(Mean Average Precision)@0.5 accuracy of 94% for YOLOv8 with SPD-Conv and NAM layers.
Towards Real-Time Traffic Sign Recognition in Diverse Weather Conditions Through Adaptive Feature Learning
2024-09-24
7619344 byte
Conference paper
Electronic Resource
English
A Real-Time Traffic Sign Recognition System
British Library Conference Proceedings | 1994
|A real-time traffic sign recognition system
IEEE | 1994
|Enhanced R-CNN Model for Traffic Sign Recognition under Diverse Environmental Conditions
Springer Verlag | 2025
|Towards reliable traffic sign recognition
IEEE | 2009
|Towards Reliable Traffic Sign Recognition
British Library Conference Proceedings | 2009
|