| As the country with the largest number of cars,China has a growing demand for intelligent transportation and car intelligence.Traffic signs are important auxiliary information on the road.Realizing the real-time detection and recognition of traffic signs has become the key to auto-driving and intelligent driving of automobiles.However,the current traffic signs of various countries in the world are quite different.Traffic signs are susceptible to environmental factors such as light,wind,rain,etc.The detection algorithm can not detect and recognize traffic signs quickly and accurately.Traffic sign detection is different from simple recognition of traffic signs.It includes two parts: marking traffic signs and recognizing traffic signs.The detection algorithm needs to filter out traffic signs from real-time images before identifying them.Therefore,it can be seen that the real-time detection research of traffic signs has important theoretical value and practical significance.The traffic sign detection algorithm based on deep learning network has the characteristics of fast detection speed and high recognition rate,which is the current mainstream research direction.Through the comparison and analysis of the current mainstream detection algorithms,we found that YOLOv4 has high detection accuracy,fast speed and good application prospects.We uses label Img to manually label the traffic sign data set.A total of 8000 images are labeled as the experimental data set.We compare the data set based on the latest YOLOv4 algorithm and YOLOv3 algorithm to verify the detection performance of the YOLOv4 algorithm on traffic sign targets.In order to improve the algorithm training speed and optimize the recognition rate of small targets,we propose improvements to the backbone network and detection layer in the YOLOv4 algorithm.We combine the residual network structure of the backbone network with the down sampling structure to reduce the feature redundancy and improve the training time.In addition,a shallow-deep feature transfer network is added to the detection layer to enrich the feature information of the detector and optimize the detection of small targets.Based on the Darknet framework,we verify the effects of the two improvements.The improvement of the backbone network shortens the training time of the algorithm from 370 minutes to 330 minutes,and the improvement of the detection layer increases the average detection accuracy from 93.35% to 94.43%.After that,we combined the two improvements and re-trained and tested based on the Darknet framework.The two improvements increased the map of YOLOv4 from 93.35% to93.87%,and the training time was shortened by 40 minutes.In order to verify the effect of the improved algorithm,we expanded the data set and replaced the detection framework with Tensorflow.We used a variety of algorithms for detection under the visualization platform tensorboard and got a more clear comparison result.The results show that the combination of the two improvements makes the training time of the improved algorithm better than the other three algorithms.The average accuracy of the improved algorithm for traffic sign detection is 92.53%,which is higher than 90.37% of Faster R-CNN,87.60% of SSD and 91.36% of YOLOv4. |
