Vague Traffic Sign Detection Based On Improved YOLOv4

As an important component of advanced assistance systems and driverless technologies,the traffic sign detection plays a sugnificant role in reducing road traffic accidents and improving the driving experiences.Traffic signs collected in real road scenes are usually disturbed by unfavorable factors such as light,occlusion,and imaging angle,resulting in blurred images,which further increases the difficulty of traffic sign detection.However,the traffic sign detection is mostly used on mobile devices with very limited computing resources,which requires detection algorithms with real-time and accurate performance on these mobile devices.YOLOv4,as a single-stage target detection algorithm,achieves a good compromise in terms of detection speed and accuracy.However,it is found that it is difficult to achieve ideal results for detecting difficult objects such as vague traffic signs by using YOLOv4 directly.Therefore,in this dissertation,YOLOv4 is improved based on both detection accuracy and speed,so that it has better performance for the vague traffic sign detection.In order to make YOLOv4 have the better detection effect on vague traffic signs,this thesis improves YOLOv4 network from the perspective of enhancing feature extraction ability and retaining more effective feature information.First,it suppresses the cluttered background information in the traffic sign images by embedding coordinate attention and makes the network focus more on the feature information of the traffic signs.Secondly,it uses that the common downsampling methods tend to ignore the variability such as shifts that cause traffic sign features to oscillate as the input is fine-tuned,and it applies Blur Pool to alleviate the loss of distortion such as shifts to a certain extent.Finally,upsampling is performed by it using DUpsampling to establish the correlation between the newly inserted pixels and the original pixels.The results of the ablation experiments show that the improved YOLOv4 improves the detection accuracy and speed of the original network by 1.34 and 2.02 on the CCTSDB dataset,respectively.The performance of the improved YOLOv4 and other traffic sign detection algorithms are compared,and the new method outperforms all of them.To verify the generalization ability of improved YOLOv4 for the vague traffic sign detection,this study sets up experiments on the enhanced GTSDB dataset.The experimental results show that the improved YOLOv4 algorithm still has good robustness.Although the detection accuracy and speed of the improved YOLOv4 algorithm proposed above have been improved,it is considered that the application equipment for traffic sign detection has very limited computing resources.A faster detection algorithm is more in line with real application scenarios,so this paper makes YOLOv4 have a faster detection speed through a lightweight network model.Firstly,the backbone of Ghost Net is used as the feature extraction network of YOLOv4 for the purpose of lightweight network through setup experiments.Secondly,in order to reduce the problem that repeated convolution requires more computational resources and there is a risk of overfitting,which makes it difficult to detect vague traffic signs,the SNL module is utilized in the feature fusion stage to reduce the computational resources and improve the feature fusion capability of the network.Finally,considering that the feature fusion part of YOLOv4 is also a deeper network in disguise,the residual edges are added between the output of the feature extraction network and each of the three detection heads to better preserve the location information of the shallow feature maps.The experimental results on CCTSDB and GTSDB show that the improved YOLOv4 sacrifices a small amount of detection accuracy while the model parameters and FLOPs are greatly reduced,and the detection speed is much increased than the original YOLOv4 does.Compared with other lightweight detection algorithms,the new YOLOv4 has a better balance between speed and accuracy,and it is more compatible with the use of real application scenarios.