CRS-NET: A CNN Network for Traffic Sign Detection

Abstract

Accurate traffic sign detection is a challenging task while very important for the safety of autonomous driving. Current learning-based traffic sign detection models generally can achieve the amenable detection results while enduring the large model size and heavy computation cost. Thus, these huge models may be unsuitable for mobile applications. This paper proposes an improved Yolov8s-based lightweight traffic sign detection model named as CRS-NET. In the proposed model, the Coordinate Attention (CA) module is added to the neck network of original Yolov8s to enhance the fusion of spatial features and channel features. Also, we introduce depthwise separable convolution to separate spatial and channel information, the Representation learning with visual tokens (RepViT) Block integrates a Light RepViT Block (LRB) module. It is used in the neck network to reduce the number of parameters and computational complexity. In the head, attention module Spatially Enhanced Attention Module (SEAM) is leveraged to reduce model size and locates the accurate position of the traffic sign. Last, the Wise- Intersection over Union (WIoU) loss function is employed to improve the performance of the detector. The CRS-NET model has been evaluated using the CCTSDB2021 dataset, achieving a 3.9% improvement in accuracy, a 1.1% increase in mAP, with a 1.8M reduction in Parameters and a 5.5G decrease in FLOPs.