Traffic Sign Detection With Multi-scale Feature Fusion Network

Traffic sign detection is an important part of intelligent transportation,and also an important module of assisted driving and even automatic driving.With the development of urbanization,road traffic problems are getting worse and worse.Nowadays,traffic sign detection has become a research hotspot at home and abroad.The key to the application of intelligent transportation system is how to improve the detection ability of traffic sign detection system in natural scene.Most of the existing traffic detection algorithms are based on shape and color features.These methods have great limitations in the natural scene.During the day or night and in different weather conditions,the color features of traffic signs will change,and the detection rate will decrease if the traffic signs are blocked or incomplete.The purpose of this paper is to improve the recall rate and accuracy rate of traffic sign detection in natural scenes.The specific work is as follows:(1)Based on GPU server computing platform,several frontier target detection algorithms based on deep learning are used to carry out the research and experiment of traffic sign detection.Faster r-cnn algorithm was used to detect traffic signs,and the algorithm was modified by the deep residual network ResNet to improve the recognition accuracy of the algorithm.Afterwards,r-fcn algorithm was used for the experimental research of traffic sign detection,and then the experimental results were compared on the same data set,and the reasons for the experimental results were analyzed.(2)The work propose a multi-scale convolutional neural network,which combines two algorithms to construct our whole detection system.In the feature extraction stage,u-net architecture is adopted to fuse multi-layer feature mapping.Finally,we make use of the idea of densebox,which can directly predict the confidence of boundary box and object class through all positions and proportions of the image.The procedure have designed an improved LocalityAware NMS(non-maximum Suppression),which is an important part of the target detection system.The method reduces the detection time and improves the accuracy of the final detection.(3)In our proposed algorithm,we adopted a novel objective function based on dice coefficient which is capable of establishing the right balance between foreground and background voxels.Consequently,we experimentally observed are much better performance.(4)The method evaluated the two benchmark data sets of GTSDB and Tsinghua-tencent 100 K and compared them with the classical algorithm.The results show that our traffic sign detection system has a good performance in terms of accuracy and recall rate.