Research On Deep Learning Based Traffic Sign Detection

Traffic sign detection plays a very important role in autonomous driving tasks,and it belongs to the sub-direction of the object detection direction.Object detection has always been one of the key issues in the field of computer vision.With the rapid development of deep learning,the accuracy of general object detection has been continuously improved.However,in traffic scenarios,general object detection has great limitations.In actual traffic scenes,traffic signs have many characteristics,such as small size objects,large object scale changes,and image quality that is easily affected by the environment.In view of the shortcomings of general object detection methods in traffic sign detection,this paper studies the current general object detection algorithms,and proposes new solutions based on them.Firstly,this paper proposes a traffic sign detection network model based on receptive field fusion and sample optimization methods.Traffic signs are of various sizes and there are many small objects,but a single network can only extract features of a single scale.This paper introduces a feature extraction module that integrates multi-scale receptive fields into the feature extraction network,and combines the information of different ranges of target attachments.Thereby improving the feature expression ability of the network for targets of different sizes.In the region proposal stage,considering the large number of candidate samples and the low proportion of effective training samples,this paper designs a new sampling indicator to filter out more effective samples by combining the target score and intersection ratio.In addition,for the error-prone samples in the detection results,this paper designs a sample post-processing module to improve the recognition ability of the entire detection network for these difficult samples.Secondly,this paper proposes a multi-scale traffic sign detection network model based on self-attention.Aiming at the characteristics of the various sizes of traffic signs,this paper first adopts a pyramid-shaped feature extraction network structure.For objects with different sizes,this paper selects features with different scales.In order to solve the problem of uneven quality of candidate samples around the target and a large proportion of low-quality samples,this paper first proposes a center sampling method to restrict the radius of the effective sample to achieve the purpose of removing low-quality samples.Considering that the existence of traffic signs is closely related to the contextual information of the specific scene of the road,this paper introduces a self-attention module to calculate the contribution of global features to the local area,so that the network can use global information to analyze candidate samples.In order to verify the effectiveness of the method proposed,this paper has carried out detailed experimental comparison on the Tsinghua-Tencent 100 K data set.The experimental results show that in terms of detection accuracy,the two traffic sign detection methods proposed in this article have achieved excellent results.In addition,this paper also compares the two network models proposed above.Through analysis,it is shown that the difference in sample selection is the main reason for the difference in performance of the two types of methods.