Research On Traffic Sign Detection Algorithm Based On Deep Learning

The detection of traffic signs is a critical component in the fields of assisted driving and autonomous driving.In recent years,deep learning has made significant strides in various fields and has rapidly advanced.As a result,the use of deep learning algorithms and related technologies for traffic sign detection has emerged as a key area of research in this field.In real-world traffic scenarios,vehicles usually travel at high speeds.To detect traffic signs accurately and in a timely manner,the vehicle needs to capture images from a long distance,enabling the processing detection network enough time to analyze and process the traffic sign information and feed the results back to the driver.However,long-distance shooting results in a small proportion of traffic signs in the entire image,leading to the loss of small target features during the deep learning algorithm’s image processing,resulting in lower object detection accuracy and reduced performance of the detection system.Furthermore,the current mainstream convolutional neural networks used in deep learning have large numbers of parameters and computation,leading to large memory requirements that are difficult to embed in the hardware platform used in vehicles.This article aims to address the above issues and undertakes the following work:(1)This study addresses the issue that traffic signs in driving images are often small and obscured by background interference.To solve this,we propose an AME-YOLOv4 detection algorithm that is tailored to the characteristics of traffic sign images taken in real-life scenarios.The algorithm enhances the spatial SAM attention mechanism and introduces an EM feature fusion module that strengthens the feature extraction network’s ability to capture small traffic sign information.Additionally,the attention mechanism of SENet channel is improved by introducing the R-SE attention module,which recalibrates the channel weights of features extracted by the backbone network.This enhances the overall network’s detection performance on small traffic sign targets.To efficiently fuse features of different scales,we introduce the adaptive feature fusion module ASFF,which assigns weights to features of each scale.Data from the CCTSDB dataset are processed for data enhancement,and the network’s generalization ability is improved.Experimental results show that our improved AME-YOLOv4 algorithm achieves a detection accuracy of 96.1% on the CCTSDB traffic sign dataset,which is 1.8 percentage points higher than that of the original YOLOv4 network.The algorithm also demonstrates a significant improvement in detecting small targets.(2)The existing YOLOv4 model is limited by its complex structure and large network memory,making it unsuitable for real-time traffic sign detection in practical scenarios.To address this issue,Chapter 3 proposes a lightweight algorithm.Specifically,the Ghost Net was selected as the lightweight network and further optimized by replacing the original Darknet53 backbone network with an improved R-Ghost.This reduced the network parameters and improved the SPP module by using faster serial pooling instead of parallel pooling,which increased the speed of the overall network model.Additionally,depth-separable convolution was introduced to replace ordinary convolution in the Neck structure,further reducing the model parameters.These improvements enabled the construction of a lightweight model,which is only 45.5M,one-fifth of the size of the original model.Experimental results show that the improved model achieves an accuracy of 94.16%,which meets the requirements of practical application scenarios.Figure[32] table[10] reference[64]...