An Improved SSD Deep Neural Network Method For Traffic Sign Detection

Traffic sign detection is one of the key technologies related to smart cars.Accurate and real-time traffic sign information acquirements help to reduce even to avoid traffic accident occurring.Detection efficiency of traditional traffic sign detection methods are difficult to be fruther improved.However,deep learning usually has strong feature representation ability and good application potential in the field of traffic sign detection.Therefore,researches on traffic sign detection methods based on deep learning has great theoretical and practical significance.Aiming at the problem of target detection of traffic signs,herein it was designed and implemented a SSD deep neural network method based on window size clustering and multi-scale feature fusion.Firstly,an investigation research of traffic sign detection benchmark datasets was carried out,and the chosen datasets were augmented and preprocessed to improve the sample diversity.And then,an improved SSD model was proposed,which replaced the VGG16 basic network in the original SSD model with a deeper ResNet50 residual network in order to improve the feature representation ability further.The K-means clustering algorithm that replaced the blind search mechanism in original SSD was exploited to define sizes of the default windows in SSD.And in addition,in order to explore the shallow high-resolution features and deep semantic features to participate detection decision-making together,a multi-scale feature fusion mechanism was introduced.Furthermore,compared experiments on the detection performance of Faster R-CNN,FPN,benchmark SSD and proposed method were carried out systematically.Finally,as for the German traffic sign detection benchmark dataset,it obtained 97.96% mAP and 0.08 s detection speed of per image.As for the Chinese traffic sign detection benchmark dataset,it obtained92.37% mAP and 0.085 sdetection speed of per image.It demonstrates that the proposed method obtains the improved detection performance.The main work includes the following two parts:1.A survey of the traffic sign detection benchmark datasets.For the sample imbalance problem,a data amplification method was exploited by increasing the noise and changing the proportion of color components.2.Summarize the basic knowledge of convolutional neural network,residual networkand SSD model and analyze the advantages and disadvantages of each typical model from the view of traffic sign detection.An improved SSD model was proposed,which can integrate the advantages of residual network,K-means and multi-scale feature fusion.Compared with Faster R-CNN,FPN,and benchmark SSD,the proposed model obtains improved detection performance.