A Study On Traffic Sign Recognition Based On Deep Learning

With the continuous improvement of people’s living standards,cars have become mass consumer goods and entered ordinary people’s families.The vast number of vehicles bring convenience to daily lives,but also bring great pressure to traffic order.The casualties and economic losses caused by traffic accidents remain high.Therefore,intelligent driving system has always been the focus of relevant researchers a nd government decision makers.Traffic sign recognition involves a wide range of traffic information,which can provide a key basis for the decision-making and planning of drivers and on-board computers.Accurate and real-time traffic sign recognition is an important guarantee for driving safety.In order to achieve accurate and real-time traffic sign recognition in various daily driving scenes,a native traffic sign dataset with comprehensive scene coverage and balanced category distribution was constructe d,and some target recognition methods based on deep learning were applied for research experiments.The related work was mainly carried out from the following three aspects:1)In a variety of daily driving scenes including nighttime and rainy weather conditions,videos containing road traffic signs were taken.The collected data were sifted and augmented,and manual classification and labeling were completed.In addition,aiming at the imbalance of classification performance caused by the sample insufficiency of individual category,the corresponding category was supplemented to achieve the relative category balance of the dataset.2)To solve the problem of image distortion under unfavorable illumination conditions,the CLAHE algorithm was used to enhance the image details.To weaken the image noise in night,rain and other weak light conditions,the Median Filtering algorithm was applied to denoise for restoring the feature information of the image.Aiming at the image blur and shadowing in moving state,the Laplace Filtering algorithm was utilized to sharpen the image to enhance the edge information of the object in the image.The target features were highlighted by preprocessing,which laid a good foundation for the effective implementation of the subseq uent recognition algorithm.3)In view of the performance of the YOLO-v3 algorithm and the Efficient Det algorithm in traffic sign dataset,the better one Efficient Det algorithm was selected as the baseline.The performance of the model was optimized throug h the equalization of datasets,the application of Mosaic and Mixup algorithms,and the adjustment of training strategies such as the Gradient Accumulation,the Transfer Learning,and the Reduce LROn Plateau.The optimized traffic sign recognition model had good robustness in sunny weather,rainy weather,night,partial occlusion,complex background,motion blur and other common traffic scenes.Finally,the recognition m AP achieved about 86%,and the reasoning speed reached over 20 FPS,taking into account both the accuracy and the speed requirements of traffic sign recognition task.