Vehicle Re-identification Based On Attributes Guidance And Discriminative Feature Mining

In recent years,with the continuous advancement of urbanization,a large number of people are flooding into cities,which brings a series of safety problems to urban traffic.To this end,a number of video monitoring terminals of tens of millions of levels have been deployed throughout the country to collect massive data.How to fully mine and utilize these massive data is a difficult problem facing traffic safety in our country.As one of the most important modules in video monitoring,vehicle Re-ID plays a vital role in maintaining social order.The task of vehicle Re-ID is to quickly match a given vehicle to a mass of monitoring data.In recent years,the research on vehicle Re-ID are more and more,but there are still many challenges have not been able to deal with well.It suffers from large intra-class variation and inter-class similarity by diversified illuminations,occlusions,viewpoint variations,motion blur,low resolutions,the same type and so on.Therefore,this thesis puts forward the corresponding solution.The main work is as follows:First,the vehicle type,color and detailed feature of different viewpoints are important clues in vehicle Re-ID.Therefore,how to make use of vehicle type,color and view information is the problem that we should pay attention to in this work.In real life,if a police officer wants to look for a suspect’s vehicle under a surveillance camera,he may first rule out vehicles of different models,then rule out vehicles of different colors to narrow down the search area,and then look for the vehicle in this small area according to the detailed characteristics of the vehicle.It inspired us to propose a vehicle Re-ID algorithm based on attribute-guided feature learning.Specifically,we first extract the features of each vehicle view,type and color through progressive learning,and then effectively integrate these features as the final identified features.The experiment proves that our method has better performance than other methods.Then,with the continuous development of deep learning,supervised learning has enabled the model to get very good results in the training stage,but not satisfactory in the test.Therefore,in this work,we propose a new data augmentation method to improve the generalization performance of the model by increasing the diversity of the training set.First,the attention model is embedded into ResNet-50 to construct a vehicle Re-ID model and train the model until convergence.Then,the sample was occluded according to certain strategies.Finally,the original samples and occlusion samples are taken as new training sets and the model is retrained until convergence.Experiments show that the proposed method has good performance.