Recognition Of Traffic Signs Based On SSD Model

In recent years,my country's transportation network construction has developed rapidly,and traffic safety has been paid more and more attention.Traffic signs play an important role in road traffic and are of great significance to protect people's personal and property safety.With the development of computer science,the recognition of traffic signs has gradually become an important research direction in the computer field.Intelligently identifying traffic signs can not only lay the foundation for the coming of the future autonomous driving era,but more importantly,it can prevent drivers from missing important road information,thereby reducing unnecessary losses.Early research mostly used feature engineering methods,using specific shape information and color information of traffic signs,and extracting feature information through manual design to classify images.However,these methods cannot meet people's requirements in terms of accuracy and applicability.In recent years,with the rapid development of deep learning technology,a series of important breakthroughs have been made in the field of target detection,including R-CNN series,YOLO series and SSD methods.The SSD directly inputs pictures into the convolutional neural network to classify and regress target objects.It belongs to the "one stage" method.The most obvious feature is that the SSD framework can directly predict on feature maps of different dimensions,and is very sensitive to the size of different target objects in the image.The main work of this paper is to construct a recognition method based on the SSD framework for the problem of traffic sign recognition.The main innovation is to propose an adaptive negative sampling method based on distance factors,which is dynamically selected as the number of training iterations changes.For negative samples,use positive and negative samples that are far away from each other in the initial stage to make the algorithm quickly converge toward the global extremum.In the later stage,use positive and negative samples that are close to each other to improve the accuracy of the algorithm.Compared with the original SSD model,the improved SSD method proposed in this paper has a 4% improvement in m AP.In addition,according to the characteristics of how small the target of the traffic sign problem is,the structure design of the SSD model is discussed in detail and experimentally analyzed in three aspects,namely,the change of the number of layers of the SSD framework,The change of the parameter ? of the adjustment regression error term and the classification error term and the influence on the noise mixed into the data set.The original SSD framework uses six feature layers of different scales for target recognition.In this paper,a contrast experiment is made on the number of layers of the feature map,and 5,6,7,8,and 9 layers are used for analysis.Experimental results show that increasing the number of feature layers is conducive to improving the recognition of small target samples.For the traffic sign recognition task,the experimental results of setting the SSD model with 8 prediction layers are the best.Compared with the original 6-layer SSD model,adding two more layers of prediction to the previous convolutional network can refine the proportion of candidate frames and make the model size more suitable for learning small target objects such as traffic signs,thereby improving the model prediction effect.In addition,this paper also carried out a detailed comparative experiment on the parameter ?,which is used to adjust the regression error term and the classification error term in the loss function of the SSD algorithm.Experimental results show that the value of ? has no substantial effect on the learning effect of the model,but as ? increases,the regression loss is gradually strengthened,so the position prediction of the frame will be more accurate.In this paper,with 10% probability,two noisy data sets are generated by blocking 1/9 and 1/16 in the location label of traffic sign data.The results show that adding 1/9 occlusion noise to the data set has a 2% improvement in the model m AP.This paper mainly proposes an adaptive negative sampling method based on the distance factor based on the SSD model,and constructs a traffic sign recognition method based on the SSD model,which has achieved a certain accuracy improvement.At the same time,this article has conducted in-depth research on the model's structure,parameters of the model and the noise processing of the data set has been carried out,and some meaningful results have been obtained.