Research On Traffic Signal Recognition Algorithm Based On Deep Learning

Traffic signal recognition technology is an important part of driver assistance systems.It can not only assist drivers to correctly identify traffic signals,effectively reduce traffic accidents,but also provide safe technical support for unmanned driving and accelerate the development of smart cities.Over the past ten years,researchers have proposed a large number of computer vision-based signal recognition algorithms.However,the existing traffic light recognition algorithms based on graphic image processing still have the following problems:(1)Most of the existing algorithms only detect traffic signals in the daytime condition,and the detection performance of the signal lights in the night condition is not good.(2)Existing methods can easily detect tail lights,pedestrian lights,and street lights as signal lights,which cannot effectively reduce misrecognition.(3)Existing models based on traditional convolutional neural networks have a poor ability to recognize small targets.The recognition accuracy rate needs to be improved.In view of the above-mentioned shortcomings,this paper proposes a method of traffic signal recognition based on deep learning.The method includes two parts:candidate region extraction and traffic signal recognition.This paper first uses a simple but very effective image pre-processing method to extract traffic sign candidate areas in the driving scene,and then proposes an improved Faster R-CNN algorithm to further detect and identify candidate areas.1.Extraction of candidate areas for traffic lights:Starting from the difference between traffic lights during the day and night,different extraction methods for candidate areas during the day and night are proposed.Firstly,by calculating the average brightness of the image,the image is quickly discriminated day and night.The black rectangular back plate of the traffic signal in the daytime has obvious characteristics.The rectangular area in the image can be detected as a candidate for the traffic light based on gray threshold segmentation and morphological filtering method;At night,the color characteristics of traffic lights are more obvious and the lights are whiter.A color segmentation algorithm based on RGB space is used to extract red,yellow,green and white areas in the image as signal light candidate areas.2.Traffic Signal Recognition Part:In view of the problem of the traditional Faster R-CNN model’s insufficient ability to recognize small targets,this paper proposes a Detail Attention Faster R-CNN(DA-Faster R-CNN)network to improve the accuracy of traffic signal recognition.The contribution lies in the following three points:(1)Use the improved ResNext network to replace VGG as the network backbone to improve the ability of the neural network shallow feature layer to extract detailed information.(2)The feature fusion module is introduced to fuse the information of two different scale feature layers output from the basic network to generate a new high-level semantic feature layer,which improves the detection performance of small targets.(3)Atrous Convolution is applied to the high-level semantic feature layer.The hollow convolution can increase the receptive field without making pooling loss information,so that each convolution output contains a larger range of information.This paper conducts comprehensive experiments on the proposed algorithm on two different data sets.The experimental results show that the traffic signal recognition algorithm proposed in this paper can adapt to two different situations,day and night,and can effectively improve the accuracy of traffic signal recognition.In addition,the algorithm meets the real-time requirements and is superior to other algorithms.