Traffic Sign Detection and Recognition Based on An Improved YOLOv5
Zhaoxia Yang, Jiaxuan Yu · 2023
Intelligent driving requires the perception of the surrounding environment and the recognition of road signs to make informed driving decisions. Accurate and real-time traffic sign recognition is a crucial component of this process. Addressing this challenge, this study focuses on utilizing the TT100K dataset and employing lightweight YOLOv5s training, along with data augmentation techniques, to enhance data diversity. The resultant model achieves an average precision of [email protected] reaching 0.848 on the PC platform, with a detection speed of 4ms on an RTX 4090. For localized deployment, a PC GUI application has been developed, offering functionalities such as model loading, image and video recognition, and real-time camera-based recognition.Considering the small-scale nature of traffic signs in driving scenarios, an improvement is proposed based on multiple comparative experiments. This enhancement involves computing new anchor boxes using K-means and adopting the Weighted IoU (WIoU) as a loss function. Furthermore, a Channel Attention (CA) mechanism is incorporated above the Spatial Pyramid Pooling Fusion (SPPF) layer of the backbone network to enhance the accuracy of detecting small-scale targets. Experimental results demonstrate that the improved YOLOv5s algorithm achieves an [email protected] of 0.868, while maintaining the same detection speed.