Research On Road Traffic Sign Detection Method Based On Deep Learning

With the development of artificial intelligence technology and the application of 5G technology,the unmanned vehicle technology based on video-based road environment sensing technology has become a new research hotspot.The traffic sign information and the traffic signal state on the road of the vehicle are informationized in real time,and the prompt information is automatically recognized and the road ahead information is warned to provide a decision basis for the driverless car system.How to eliminate interference from environmental factors as much as possible,and accurate and rapid detection and identification of traffic signs is a key technical problem that needs to be solved urgently.It has great research value and application prospects.Chinese traffic signs include restrictive traffic signs,banned traffic signs,warning traffic signs,and indicative traffic signs.The existing methods of traffic sign recognition are usually limited to those of relatively restrictive,prohibitive,and warning signs.The shape is rectangular,triangular,and circular.For small-distance traffic,Insufficient research on signs.Moreover,the number of samples of the indicative signs in the existing traffic sign image data set is seriously insufficient,and does not even include such samples.Road traffic conditions in the real environment are complex and changeable,with small targets,insufficient data volume,high sample similarity,and uneven sample distribution.These problems make traffic sign detection research face many difficulties,and the actual application is far from mature.Based on the above problem,this thesis studies the road traffic sign detection method based on the YOLOv3 method and the Tsinghua-Tencent 100K(TT-100K)traffic sign Dataset with comprehensive and close to the real environment.The main work of this thesis has the following aspects:(1)For the insufficient number of data samples,this thesis enhances the data based on the TT-100K Dataset.Improve the balance of data distribution by mirroring,rotating and scaling.Pre-processing such as denoising,defogging,and contrast enhancement is carried out for effects such as blur,fog,and night traffic lights.(2)For the difficulty and real-time requirements of small target detection,this paper chooses YOLOv3 with a large number of residual structures as the basic framework,and adopts a multi-scale feature fusion strategy to enhance the detection ability of small targets,replacing the upsampling layer of the original network with a convolution Layers to make training more efficient.(3)For the problem of difficult classification of fine-grained features,the network structure of the detection model was improved by introducing an attention mechanism to improve the classification effect,and the training model was optimized by fine-tuning the model and adjusting parameters.(4)For the problem that the classification effect caused by uneven sample distribution is accurate enough,the gradient loss coordination mechanism is adopted to improve the network loss function,and the weight distribution of simple-difficult samples is adjusted to make training more efficient and improve classification accuracy.(5)For the problem of missed detection of small targets that are too close,a soft-NMS algorithm is introduced to improve the original NMS algorithm and improve the detection accuracy.Using the A-G-YOLOv3 method proposed in this paper,the Tsinghua-Tencent 100K(TT100K)Dataset was used for testing,and the German Traffic Sign Detection Benchmark(GTSDB)Dataset was used for comparison experiments.The experimental results show that the proposed two data and the mAP of the above data are 81.16%and 98.76%,respectively,which are 12.49%and 5%higher than the improvement,respectively,3.93%and 3.14%higher than the comparative experimental method Faster R_CNN.And the detection speed can reach up to 20FPS,indicating that the method proposed in this thesis has certain applicability.