Research And Implementation Of Traffic Sign Detection Platform Based On Deep Learning

Traffic Sign Detection is an important part of intelligent transportation system,which is widely used in real-time environment perception,high-precision map construction and other tasks.However,the existing traffic sign detection algorithms are often based on general object detection,and do not fully consider the characteristics of traffic signs.Traffic sign detection faces the difficulties of three aspects:(1)Traffic sign images are often collected based on on-board camera,so most of the traffic signs accounts for only a small proportion in the original image,which makes it is very difficult to detect such small traffic signs.(2)The same kind of traffic signs are very similar,so the accurate classification of the traffic signs is a very challenging.(3)In order to obtain better feature representation,the object detection model based on deep learning often uses a very heavy backbone,resulting in a large number of parameters and runtime delay,which is difficult to meet the requirements of real-time deployment.To solve above problems,this paper did the following:(1)This paper proposed the Feature Aggregation MultiPath Network FAMN,which can aggregation multiscale features and build the fine-grained features of traffic signs.On average,FAMN is 2.9%higher than state-of-the-art.(2)This paper proposed Attention Based Adaptive Dynamic Network AD-RCNN,which consists of Dynamic Region Proposal Network,Visual Attention Branch and Adaptive Dynamic Training.Compared to the baseline method,AD-RCNN improved the precision of 8%.Besides,this paper build the AD-RCNN Lite to meet the demand of real time traffic sign detection.(3)Based on the above networks,this paper implements a traffic sign detection platform,verifying the effectiveness of proposed methods.We encapsulate the method proposed in this paper as API services,so that users without computer vision background can easily build their own traffic sign detection applications.