1. Introduction
Traffic signs play an important role in traffic flow, which contribute to the development of assisted driving and autopilot technology and accelerates the construction of intelligent transportation system. Currently, the main solution for traffic signs detection is to collect images by vehicle-mounted cameras and use computer vision and pattern recognition methods for detection and recognition [1]. With the rapid development of deep learning technology, target detection algorithms, such as Faster RCNN, YOLO, FCOS, etc. have been widely used in traffic signs detection.
In research on traffic sign detection, Li et al. [2] proposed an improved Faster R-CNN detection algorithm to deal with the problem of complicated background and small traffic sign detections in wide-field traffic scenarios. Luo et al. [3] introduced Bi-FPN for feature fusion and used GAM (global attention mechanism) [4] to enhance the feature extraction ability of the network to verify the effectiveness of the improved YOLOv5 algorithm for traffic sign detection. Chen et al. [5] proposed a multi-scale feature fusion detector based on the improved FCOS (full-convolution single-stage object detection) [6] to solve the problems of small traffic sign detection. The above-mentioned have achieved certain results in the detection of traffic signs, but there are still problems such as low detection accuracy and missed detection in actual detection. Therefore, further research is needed on traffic sign detection.
This paper based on the YOLOv8 algorithm proposed a data enhancement method by analyzing the characteristics of small objects, and introduced a new method of feature fusion to improve the accuracy. The improvements are as follows:
• Adding noise and blur processing on the original data augmentation method makes the network pay more attention to fine-grained features, which contribute to detecting small traffic signs.
• In the feature fusion stage, the asymptotic feature pyramid network (AFPN) is used to achieve direct interaction of non-adjacent layers and enhance the expressive ability of key features.
2. Algorithm Overview
The YOLO family is a classic object detector. Since the first edition of this algorithm was published in 2015, it has achieved a leading efficiency with a single-stage framework and has quickly become a mainstream detection algorithm. Through continuous research and innovations, different versions of YOLO have been proposed. The latest version is the YOLOv8 [7] algorithm, which was open-sourced by Ultralytics in January 2023. This algorithm introduces new features and improvements, becoming the best model in the YOLO family. YOLOv8 includes four parts: Input, Backbone, Neck and Output. The structure of YOLOv8 is shown in Figure 1.
2.1. YOLOv8 Algorithm Theory
Input mainly includes color perturbation, spatial perturbation, Mosaic, MixUp. Different numbers of pictures are spliced after processing a single one by combined data augmentation, which increases the multi-directional object perspective and also enriches the diversity of image backgrounds.
Backbone mainly includes convolution layer, C2F layer and SPPF layer. The C2F structure is different from the C3 module in YOLOv5 which draws on the idea of ELAN (efficient layer aggregation network) in YOLOv7 [8], which increases the efficiency of gradient propagation and enables the network to quickly converge. The SPPF layer maintains the design in YOLOv5.
Neck adopts a structure combining FPN (feature pyramid network) and PAN (path aggregation network) [9]. The features of adjacent layers are concatenate which are the input to the C2F module. As features are passed from top to bottom
and bottom to top, high-level semantic features with underlying features are combined.
Output realizes the decoupling of detection and classification. The bottom-level features are used to obtain the information of small target objects, and the top-level features are results of large target objects. Each detection layer outputs a result vector, which contains location and corresponding category information.
2.2. Improvements in Data Augmentation
DA (data augmentation) is widely used in deep learning data processing, such as image scaling, translations, random rotation, etc. The above methods can get different perspectives of the object. CutMix, Mixup, Mosaic, Copy-Paste, etc. combine different pictures, which can increase the diversity of backgrounds. Research on data enhancement for small target object detection [10] focuses on oversampling or resampling of small target samples, by transforming small target objects and pasting them in different positions in the image to increase the number of small target objects in training.
This paper uses the TT100K [11] dataset, which includes traffic signs produced by Tsinghua University using Tencent Street View panoramas. The pictures are obtained under different weather conditions, and the illumination changes greatly. Usually, the proportion of traffic signs in the picture is small, only 0.2% of the image size. The proportion of the target size in the training pictures of this data set is counted, and the result is shown in Figure 2. It can be seen that the proportion of traffic signs is mostly concentrated in the range of 0.01% to 0.05%.
DA, such as motion blur, fog and raindrop noise, is added on the basis of original method to simulate images acquired in different weather and traveling states. The effect after data augmentation is shown in Figure 3.
![]()
Figure 2. The ratio of image size in TT100K training set.
![]()
(a) ![]()
(b)
Figure 3. Comparison of DA. (a) DA in YOLOv8; (b) Ours.
2.3. Improvement of Feature Fusion
Feature pyramid is an important part of neural network, which is used to fuse features to obtain comprehensive description information. Neck in YOLOv8 achieves the consistency of the size of the adjacent feature maps by up-sampling or convolution, and then directly concatenates the channels of the feature maps to complete the preliminary fusion between the features. On the other hand, unidirectional information in concatenate layer can be used, or feature information in another direction can be obtained by using the feature pyramid after the feature fusion in the same direction is completed. Moreover, feature fusion cannot be directly performed between non-adjacent feature layers in the current network. This will lead to the loss of information when the features of non-adjacent layers are fused.
Therefore, this paper uses AFPN [12] to complete the feature fusion in Neck, and its structure is shown in Figure 4. This method combines the ideas of the pyramid idea and ASFF (adaptively spatial feature fusion) [13] to realize the fusion of features. Firstly, it is the fusion of adjacent features. By up-sampling or down-sampling between adjacent layers to get the unification of the size of the feature map in the current layer, and then uses the adaptive spatial fusion of features which pays more attention to key layers, as shown in formula1. Among them,
and
respectively represent the weight of the input during linear
combination, satisfying
,
represents the conversion of eigenvectors located at (i, j) from the nth layer to the lth layer. This formula is also applicable to linear combinations between multi-layer features. Secondly, it is the fusion of cross-layer features. Since the feature fusion already contains the adjacent feature information, the language gap between features is reduced when the cross-layer features are fused.
(1)
3. Model Training and Testing
3.1. Experimental Environment
This article builds an algorithm running environment based on Docker. The specific training environment is: Ubuntu20.04 system, Nvidia 3080Ti which memory size is 12 G, the Pytorch version is 1.10.0a0, and the CUDA version is 12.0.
3.2. Evaluation Criteria
Recall and mAP (mean average presicion) are used commonly to evaluate the model performance. The recall mainly focuses on the ratio of the number of detected targets to the total number of targets. The mAP indicates the average value of the average accuracy rate of each type of target under a certain IOU (intersection over union) condition for all categories.
3.3. Comparative Experiment
This paper compares and analyzes the comprehensive performance of the improved YOLOv8 network. Parameters of training such as learning rate and batch size remain unchanged except for the module improvement, and the number of epochs is 500. The experimental results are shown in Table 1, it can be found that compared with the standard network, the improved YOLOv8 has increased mAP by 3.31 percentage points, Recall has increased by 3.59 percentage points.
We also conduct ablation experiments on the influence of different improved parts on the model, and the results are shown in Table 2. Among them, “×”
![]()
Table 1. Comprehensive index test results.
![]()
Table 2. Ablation study of data augmentation and AFPN.
means that the corresponding improvement strategy is not used, and “√” means used. The recall of the ablation of DA-ours is improved by 1.94 percentage points, while the mAP is slightly lower than YOLOv8. The AFPN improves map and recall by 1.82 and 1.5 respectively, which enhances feature fusion.
4. Conclusion
Aiming at the problems of missed detection and low detection rate in traffic signs, this paper proposes an improved YOLOv8 algorithm to detect traffic signs, which uses the data augmentation method in a specific scenario to effectively enhance the diversity of the data itself, so that the network can learn effective features and apply AFPN for fusion features. Experiments show that the improved algorithm can reduce missed detection and improve accuracy.