Learning and Detection of Road Signs

Xinyu Liu


Supervised by Shancang Li; Moderated by Tingting Li

Traffic sign recognition plays a crucial role in intelligent transportation systems, and its accuracy and real-time performance directly affect road safety and traffic flow. At present, deep learning based object detection algorithms have made significant progress, with YOLOv5 being widely used in the field of object detection as an efficient real-time object detection framework. However, there are still some challenges in small target detection and target feature extraction of traffic signs.

This paper introduces two important improvement methods to address the shortcomings of YOLOv5: BiFPN (Bi directional Feature Pyramid Network) and CBAM (Convolutional Block Attention Module). BiFPN effectively integrates features of different scales by introducing a bidirectional feature pyramid network, enhancing the model's ability to detect small targets and improving recognition accuracy. At the same time, CBAM introduces an attention mechanism, which enables the model to adaptively select key information in the feature map, further improving the model's expression ability. This article first introduces the background and significance of traffic sign recognition, and outlines the current application status of deep learning in the field of traffic sign recognition. Subsequently, the principles and working principles of YOLOv5, BiFPN, and CBAM were introduced in detail. On this basis, an improved YOLOv5 model integrating BiFPN and CBAM is proposed, and the network structure, training strategy, and evaluation method are described in detail.

The experimental results show that the improved model proposed in this article has achieved significant performance improvement on the public traffic sign dataset. Compared with the traditional YOLOv5 model, the improved model has achieved significant advantages in small target detection and target feature extraction of traffic signs, with significantly improved recognition accuracy. In addition, this article also analyzes the robustness and real-time performance of the model, and explores possible directions for further optimization in the future. In summary, the research in this paper proposes an effective improvement plan for deep learning methods in the field of traffic sign recognition, which has important theoretical and practical value and positive significance for improving traffic safety and optimizing traffic flow.

Final Report (07/10/2023) [Zip Archive]

Publication Form