Research On Traffic Sign Recognition Based On Convolutional Neural Networks

Traffic signs contain a wealth of traffic information that guides drivers to drive safely.The accurate and rapid recognition of traffic signs by computer vision technology is of great significance for the realization of assisted driving and autonomous driving.However,in the real world,the scenes of traffic signs are complicated.The image obtained by the camera is easily affected by the factors such as lighting conditions,driving speed,shooting angle,etc.,which brings great difficulties to the recognition of traffic signs.The recognition algorithm based on the combination of manual design features and traditional machine learning is less adaptable and less scalable.How to design a more robust recognition algorithm is an urgent problem to be solved.At present,convolutional neural networks have achieved great results in the field of object detection and image classification.In this paper,the convolutional neural networks is applied to the field of traffic sign recognition,and a traffic sign recognition algorithm is proposed.This algorithm divides the recognition task into two stages: detection and classification.The details are as follows:As an important part of the Advanced Driver-Assistance System(ADAS),traffic sign recognition has attracted more and more attention from researchers.Accurate and rapid recognition of traffic sign has important significance for assisted driving and autonomous driving.At the same time,convolutional neural networks have achieved significant results in both object detection and image classification.A series of object detection models such as Faster R-CNN,SSD,and YOLO have been proposed,while image classification models such as AlexNet,VGG,GoogLeNet,ResNet DenseNet continue to refresh the error records of the ImageNet Large Scale Visual Recognition Challenge.This paper studies traffic sign recognition based on convolutional neural network,and divides traffic sign recognition into two stages: traffic sign detection and traffic sign classification.The details of this paper are as follows:(1)In traffic sign detection stage,YOLOv3 was first trained on the German Traffic Sign Detection Data Set(GTSDB),and the six models saved during training were tested on 15 input sizes.The model with highest recall was obtained.That is,the least missed.Then compare the performance of YOLOv3 and YOLOv3-tiny on GTSDB,and propose to replace the feature extraction network of YOLOv3 with YOLOv3-tiny feature extraction network to make the model complexity moderate,in order to achieve the best trade off of accuracy and speed.Finally,the experiment show that the recall of proposed YOLOv3-I only decreased by 2.2%,and the detection speed increased to 2.25 times of YOLOv3.(2)In traffic sign classification stage,a simple and efficient classification model is designed for the problem of high complexity of existing models.This model combines the low-level features and high-level features through shortcut,and uses different size asymmetric convolution kernels to convolve the inputs on multiple branches.At the same time,Dropout is used to reduce over-fitting,Batch Normalization is used to accelerate training,and global average pooling is used to reduce the amount of model parameters.Finally,experiments were carried out on the GTSRB test set.The results show that the designed model achieves an accuracy of 99.41%.Moreover,the model parameters are small and the recognition speed is fast,thus the model complexity is low.(3)When implementing the traffic sign recognition algorithm,traffic sign ROI ratio judgment and category logic judgment are proposed to reduce the false positive in the detection stage.The traffic sign recognition algorithm YOLOv3-R is realized by using YOLOv3 detection model and classification model and the traffic sign recognition algorithm YOLOv3-R-I is realized by using YOLOv3-I detection model and classification model.It is verified by experiments that the two algorithms have their own advantages in recognition performance and recognition speed.In actual application,they can be selected according to requirements.Finally,the interface written in Python for the recognition algorithm is completed based on PyQt5 under Ubuntu in order to visually visualize the recognition results.