Vehicle Re-identification For City OD Survey

City OD survey has great significance for both urban traffic management and city construction programming.OD is abbreviation of origin and destination.It includes traffic flow survey and vehicles travel path survey.In the past,people conduct city OD survey by handing out questionnaires or on-site interviewing.These methods will not only consume large amounts of resources but also waste a lot of time.With the development of science and technology,using machine helps people with work has become possible.This paper will research how to make use of computer vision technology and traffic monitoring equipment helping people with city OD survey.To conduct traffic flow survey,we can use the object detection technology based on deep neural network.The object detection can detect the objects of our interest.Training a deep neural network needs a lot of data,and labeling data is a boring and time-consuming work.Aiming at this problem,we design an online learning network which can help with labeling data.This network based on object detection network,and it can obtain the classes and regions of the objects which we are interesting in.The results obtained by this network can help with labeling data.Meanwhile,the artificial labeled data can used to train the network.After repeated training,the prediction results of the network is close to the artificial labeled data.To conduct the vehicles travel path survey,we can make use of vehicle re-identification technology.This technology can re-identify two vehicles appear in different areas,we can infer the travel path and travel time of the vehicle with this information.However,in real environment vehicle re-identification meets many problems such as lack of license plate information and vehicles of same model.These problems will interference the model.To solve these problems,we propose a feature fusion network extracting features for vehicle re-identification.This network can obtain weights according to different input images.These weights are used for the feature fusion.This network is specially designed for vehicle re-identification in real environment,it improves the re-identification accuracy greatly.We collect two datasets,and then we conduct several contrast experiments on our datasets to validate our methods.The result of these experiments proves the effectiveness of our methods.