Research On Traffic Sign Recognition For Autonomous Driving Scenario

The research and application of autonomous driving systems and assisted driving systems have become a hot topic of concern for researchers and enterprises in the automotive field today.Automatic detection and recognition of traffic signs is a very important part of this.Traffic signs contain a lot of important information,and correctly recognizing traffic signs can reduce many unnecessary traffic accidents and protect the safety of passengers’ lives and property.However,there are currently two factors affecting the further improvement of recognition rate in the field of traffic sign recognition,which restrict its practical application.The first factor is that during actual driving,changes in the actual scene will affect the correct recognition of traffic signs,such as: 1)weather changes,sunny days may cause reflections in the traffic signs in the captured images,and shadows on cloudy days;2)angle changes,different shooting angles when driving at different positions will cause deformation of traffic signs in the image;3)motion or vibration effects,when the vehicle is moving at high speed,motion blur will occur in the captured image.The second factor is that the images collected while driving will produce a phenomenon of near-large and far-small like human eyes.However,it is obviously better to recognize traffic signs as far away as possible,that is,the smaller the traffic sign that can be recognized,the better.In order to recognize traffic signs earlier and give drivers or autonomous driving systems more reaction time,the performance of small traffic sign recognition is also very important and is a focus of attention in today’s traffic sign recognition and detection algorithms.Therefore,in order to solve the above problems,this paper proposes two aspects of work improvement:(1)In response to the problem of changing scenes during actual driving,this paper proposes a data enhancement strategy based on multiple image linear transformations to improve the generalization ability of datasets under different scenes and improve the robustness of models.In response to the problem of weather brightness and darkness,a random brightness change method is used to enrich the brightness features in the dataset;in response to the shooting angle problem,random scale transformation and perspective transformation methods are proposed to enrich scale features in the dataset;in response to motion blur caused by high-speed movement,propose a motion blur method to enrich the dataset.In response to the sample imbalance problem in the TT100 k dataset,use a random sticker method to balance the dataset.Then,the enhancement efficiency of different enhancement strategies was compared,and the enhancement method with the highest increase in information was found to be random scaling and viewpoint transformation.This proved that the model trained only on the TT100 k dataset mainly lacks traffic sign information observed from multiple angles.In addition,by combining multiple linear transformations,the number of samples in the dataset was increased by five times.Under the Faster R-CNN+FPN algorithm,the recognition rate was improved from 86.4% to92.9%.(2)In response to the problem that most traffic signs are small targets,an attention mechanism-based bidirectional fusion feature pyramid network is proposed to improve recognition accuracy.The algorithm framework used in this paper is a two-stage object detection algorithm Faster R-CNN and improved by various means.Firstly,for VGG16’s poor feature extraction ability for Faster R-CNN’s backbone,VGG16 was replaced with Conv Ne Xt＿base with stronger feature extraction ability.Secondly,in response to low recognition rate for small targets,a bidirectional feature fusion network based on attention mechanism was proposed to better fuse low-level features with high resolution and weak semantics with high-level features with low resolution and strong semantics.This makes it possible for models to improve their ability to recognize traffic signs of different scales,especially for small targets.In the end,model overall accuracy increased by 1.6%,reaching94.5%,nearly 7% increase for small targets.Finally,because 45 classifications are not enough to reach practical application,so TT100 k dataset was enhanced by Maps method,making all signs appear more than 500 times,restoring original 232 classifications,achieving 78.5% accuracy.Achieving the forefront of accuracy in traffic sign recognition today,it plays an important role in the field of traffic sign recognition.