Research On Traffic Sign Detection Algorithm Based On YOLOv4

With the rapid development and wide application of driverless technology,the accuracy and efficiency of traffic sign detection are crucial for the safe driving of driverless vehicles.Currently,traffic sign detection algorithms based on deep learning have become a research hotspot and application focus,and these models have the advantages of efficiency,accuracy and robustness.However,it still faces great challenges to achieve high speed and high accuracy traffic sign detection in complex and changing driverless scenarios.At the same time,traffic sign images in real scenarios often occupy only a small part of the entire field of view,which is a typical application scenario for small target detection,and the performance of mainstream image detection algorithms in small target detection still needs further improvement.To overcome these problems,this paper investigates the adaptable,efficient and accurate YOLOv4 traffic sign detection algorithm based on the YOLOv4 model by improving and optimizing the backbone network,neck connection network and detection head.The main research works are as follows.(1)A traffic sign detection algorithm with improved YOLOv4 is proposed to deal with the issue of insufficient detection accuracy due to the small position occupied by traffic sign images.The 13×13 detection layer is removed from the network and a104×104 detection layer is added to obtain more global feature information containing location information,and an attention mechanism is inserted into the algorithm to make the network focus on the target region and obtain more key features.The inclusion of dynamic residual connections in the backbone network promotes the propagation of well-performing signals and the use of Decoupled-head detection head,which accelerates the convergence of the network and improves the detection accuracy.The experimental results show that the m AP value of the improved algorithm is 4.78% higher than that of YOLOv4,and the AP value of the improved algorithm is 1.4% higher than that of YOLOX for the comparison on the mandatory class.(2)A lightweight network model is proposed for the problem of long training time when the number of traffic sign model parameters is large.The study of improving YOLOv4-based traffic sign detection model using Mobile Netv3 firstly replaces the backbone network based on the improved Mobile Netv3,using the depthwise separable convolution to reduce the number of parameters and computation of the model.Then a new SPP module is added to increase the perceptual field of the feature layer and capture the effective contextual features.The experiments initially use the K-means++ algorithm to set the prior frame to improve the detection accuracy of the algorithm.Finally,the comprehensive performance is better by comparing with other deep learning models.Moreover,the model size of the improved lightweight algorithm is only 19.4% of YOLOv4,and the m AP is improved by 1.7%,while the detection speed is also improved by 25%.(3)The improved YOLOv4-tiny traffic sign detection model is proposed to address the drawback that the YOLOv4-tiny algorithm is prone to the loss of key feature information during the feature extraction process.Using lightweight attention and new detection heads to optimize YOLOv4-tiny,firstly,a lightweight channel attention module is added to the backbone network to enhance the key features in the target region.Second,a new 52×52 detection head is added to obtain more location information from the shallow feature layer.The experimental results verify that the improved model can significantly improve the detection accuracy and robustness while maintaining the small model size and fast detection speed,and the m AP value is11.62% higher than that of the original YOLOv4-tiny.