In urban road scenes, due to the small size of traffic signs and the large amount of surrounding interference information, current methods are difficult to achieve good detection results in the field of unmanned driving.
To address the aforementioned challenges, this paper proposes an improved E-YOLOv4-tiny based on the YOLOv4-tiny. Firstly, this article constructs an efficient layer aggregation lightweight block with deep separable convolutions to enhance the feature extraction ability of the backbone. Secondly, this paper presents a feature fusion refinement module aimed at fully integrating multi-scale features. Moreover, this module incorporates our proposed efficient coordinate attention for refining interference information during feature transfer. Finally, this article proposes an improved S-RFB to add contextual feature information to the network, further enhancing the accuracy of traffic sign detection.
The method in this paper is tested on the CCTSDB dataset and the Tsinghua-Tencent 100K dataset. The experimental results show that the proposed method outperforms the original YOLOv4-tiny in traffic sign detection with 3.76% and 7.37% improvement in mAP, respectively, and 21% reduction in the number of parameters. Compared with other advanced methods, the method proposed in this paper achieves a better balance between accuracy, real-time performance, and the number of model parameters, which has better application value.