Traffic Sign Small Object Detection Algorithm Combining Multi-feature and Accurate Location

• Published in: 2024 IEEE International Conference on Cognitive Computing and Complex Data (ICCD) pp. 131 - 138
• Main Authors: Ke, Zunwang, Chen, Shuaibo, Du, Minghua, Zhou, Jisheng, Zhang, Yugui
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.09.2024
• Subjects: Accuracy; Autonomous vehicles; Cognitive systems; Data visualization; Feature extraction; Fuses; Normalized Gaussian distance; Object detection; Robustness; Sensitivity; traffic sign detection; Triple feature fusion; YOLOv8
• DOI: 10.1109/ICCD62811.2024.10843612

In the rapid development of autonomous driving technology, the accurate detection of traffic signs has become an indispensable part. Traffic signs are usually small in size, which leads to low detection accuracy, missed detections, and false detections with general object detection algorithms when capturing small, low-pixel traffic signs. To tackle these issues, this paper proposes a traffic sign small object detection algorithm combining multi-feature and accurate location. The algorithm utilizes a Triple Feature Fusion Module (TFF) to fully integrate features of different scales, providing rich shallow details for small-scale object detection; In addition to the original three detection layers, we have added a dedicated layer for detecting small-scale objects in YOLOv8, enhancing the algorithm's capability for small-sized traffic signs; By combining the normalized Gaussian distance (NWD) and CIoU, the loss function is optimized to tackle the problem that CIoU loss is overly sensitive to the position shifts of small-scale objects. Tested on the TT100K2021 and CCTSDB2021 datasets, the results showed that the mAP of our algorithm reached 82.8% and 81.5% respectively, which is an increase of 9.6% and 2.2% compared to the original YOLOv8. It is proved that the proposed algorithm has remarkable performance in traffic sign detection. https://github.com/ShuaiboChen/SmallObjectDetection.git