Study On Traffic Object Recognition Based On Deep Learning Technique

With the continuous development of technology,intelligent transportation system has been a necessary part of human life.GPS system,intelligent travelling system,intelligent supervising system and so on has been everywhere in our daily life.Unlike those systems,unpiloted system is still in developing stage.A set of technology magnates in our world have been devoted themselves on this field for many years.They also get some progresses.Computer vision,as an important part of unmanned driving technology,is also attracting more and more attention.How to use the image information in the driver's view to analyze the objects in the environment is an important proposition in the present.On the other hand,With the help of artificial intelligence these problems seem to getting closer and closer to the best solution.The main content of this paper is as follow:(1)Based on previous researches on deep learning,this paper explores the possibility of the application of deep learning recognition algorithm in object recognition under the driver's field of vision.Through the experiment of udacity's unmanned data set,it is found that the existing single-phase detection algorithm SSD and two-phase detection algorithm,Faster-Rcnn,have shortcomings and shortcomings.The single-stage algorithm can not match the two-stage algorithm in the accuracy rate,and the two-stage algorithm can not meet the real-time requirement in the computation speed.(2)based on SSD algorithm,this paper aims to improve the application of this method in the field of unmanned driving.This improvement is reflected in two points,one is the addition of the FPN feature pyramid structure on the basic network,and the other is the method of series two SSD algorithms in the overall algorithm framework.The cascaded SSD algorithm performs well on the experimental data set,reaching 84 on the Map value,and only 3FPS slower than the SSD algorithm in the prediction speed.In addition,this algorithm has a significant improvement over the other two algorithms in the false positive rate,only around 5%.In addition,in order to make the cascaded SSD algorithm more practical,this chapter presents the method of image segmentation.The image segmentation method utilizes the mechanism of parallel processing,so that the algorithm can make full use of the performance of the hardware,so that the hd video data can be processed in real time,and the operation speed can reach 26FPS.(3)the traffic signs are specific to the traffic signs.This paper also studies how to identify traffic signs based on relevance.In the driver's view,the driverless system is more concerned with its own traffic instructions.However,it is found that the existing network structure can not be convergent under the task of this correlation.For this reason,this paper introduces the non-local network module to improve the existing structure.Finally,a good effect was obtained on the experimental data set,and the Map value reached 87.(4)in terms of algorithm implementation,this paper mainly adopts the mxnet deep learning framework,and the innovation in some structures is mainly realized by modifying the cuda code at the bottom of mxnet.The goal is on the one hand,in order to make the new structure embedded in mxnet training and library easy to use,the other side in order to make the algorithm more practicality can be flexible with mxnet library for migration.