Research On Traffic Sign Detection Algorithm Based On Deep Learning

As a major component of intelligent transportation,traffic sign detection can identify and classify traffic signs timely and accurately in the face of complex traffic conditions,providing an important guarantee for people’s safe travel.However,the existing traffic sign detection algorithms often face the problems of small traffic signs and poor detection effect due to other factors,and slow detection speed due to the pursuit of high accuracy detection.Therefore,this paper addresses the above problems by using the SSD model as the underlying model,and through a series of improvements and optimizations to meet the demand for high-precision realtime detection.The main research contents of this paper are as follows:(1)To address the issues of low detection accuracy and missed detections caused by small traffic signs or other factors in real traffic scenes,an improved traffic sign detection algorithm based on SSD is proposed.First,the K-means++ clustering algorithm is used to adjust the size of the default window,effectively avoiding the problem of mismatching due to the default window being too large for small traffic signs.Secondly,a Res Ne St network is used as the backbone feature extraction network,which incorporates group convolutions and channel attention mechanisms to enhance the strong representation ability of weak target features.Additionally,RFB modules are added in extra layers to increase the receptive field of small targets.Finally,the Bi-FPN network is used to effectively combine the semantic and positional information of the model’s deep and shallow feature maps,enabling the model to display rich feature information at different scales.By combining the above improvements,the detection performance of small traffic signs in real scenes is effectively improved.(2)To address the problem of large parameter size and slow detection speed in the improved SSD model,which hinders efficient operation on memory-constrained and computationally limited in-vehicle devices,a lightweight improved SSD traffic sign detection algorithm called Bi Mb3-SSDlite-KD is proposed based on knowledge distillation.In this paper,the improved SSD model is used as the teacher network,and the lightweight network Mobile Netv3-SSDlite,which has a similar structure to the SSD model,is selected as the student network.The model compression technique of knowledge distillation is employed to guide the student network in learning rich intermediate feature representations from the teacher network.This approach aims to achieve model lightweighting while maintaining high accuracy,thereby meeting the real-time detection requirements.To verify the effectiveness of the proposed algorithm,this paper conducted extensive experimental analysis on the CCTSDB dataset.The improved SSD model achieved a m AP of95.33%,which is a significant improvement over the original SSD model and outperforms other cutting-edge learning methods of the same period,making it better suited for traffic sign detection in natural scenes.In addition,the m AP of the Bi Mb3-SSDlite-KD model reached91.37%,which,compared to the improved SSD model,reduced the model size by 74.25 M while sacrificing only a small amount of accuracy,and increased the detection speed by 88.9%.This demonstrates that the model has effectively balanced the speed and accuracy of traffic sign detection,achieving fast and accurate detection.