Research On Vehicle Re-identification Across Non-overlapping Camera Views

Vehicle re-identification(Re ID)is a new technology emerging in the field of intelligent video analysis,which aims to achieve accurate identity consistency association of vehicles in the network of surveillance cameras.Vehicle Re ID is utilized to distinguish whether the vehicle images captured by different cameras at different times are vehicles of the same identity.Therefore,it can expand single-camera surveillance to multi-camera collaborative surveillance system,and automatically analyze the surveillance data to improve monitoring efficiency for vehicle retrieval.In the research of vehicle reidentification,due to different shooting angles,low resolution,occlusion,and lighting changes in the actual monitoring video of road network,the appearance of the vehicle in the image may change significantly,making vehicle Re ID very challenging.Based on the above,this paper proposes multiple vehicle Re ID methods from different aspects.The main research work is as follows:(1)At present,a lot of research work on Re ID mainly focuses on designing feature representation or feature matching,which lacks consideration of the interconnection relationship between samples.Therefore,this paper proposes a novel Re ID model based on synergistically cascade forest to model the deep between-vehicle relationship.With the deepening of levels of synergistically cascade forest,the capacity of the model to distinguish vehicles captured from diffierent cameras is gradually enhanced.Furthermore,curriculum learning is introduced in the training process of synergistically cascade forest model.It paves an easy-to-difficult training framework for vehicle Re ID,and the model performance obtains significant improvement.Through deep learning of the interconnection relationship between samples,the number of parameters of the synergistically cascade forest model is small,but it has good Re ID performance in both small-scale and large-scale datasets,and even same model parameters for different datasets also achieve favourable generalization ability.(2)Most existing vehicle Re ID methods depend on a large number of labeled data.However,collecting and labeling large-scale vehicle Re ID datasets mainly rely on manual methods,which is very difficult and time-consuming.This paper utilizes one-view Cycle GAN to generate style-transferred image between different cameras.The generated multi-order images are used as augmented data.It not only can increase the diversity of data to reduce the risk of model overfitting,but also can avoid complex camera network matching problem.Multi-order images contain different interdomain information.Center Loss and Hard Triplet Loss are combined to learn the cross-distance between multi-order images,which can effectively reduce cross-camera discrepancy.Therefore,it is efficient to overcome the problem of large intra-variation and little intervariation,and improve the robustness of the Re ID model.(3)Aiming at the application prospect of cross-camera vehicle Re ID in the tunnel,this paper constructs a Tunnel-VRe ID dataset in order to enrich the diversity of scenes.Due to the complex illumination,the vehicle appearance in tunnel monitoring is often blurred and difficult to recognize.In view of the extremely challenging tunnel scene,this paper proposes a tunnel vehicle Re ID model based on temporal-spatial constrain.The temporal-spatial constrain,using the travel time between vehicles across cameras,is constructed by gaussian mixture model and bayesian inference.This paper introduces temporal-spatial information into vehicle Re ID task,which needs to satisfy similar visual appearance of vehicles and temporal-spatial constraint at the same time.Through mutual support of visual appearance and temporal-spatial constraint,it builds a more discrimin tunnel of vehicle recognition model.In summary,aiming at the problems existing in vehicle Re ID tasks,this paper progressively proposes three vehicle Re ID model,including synergistically cascade forest,multi-order deep cross-distance learning and temporal-spatial constrain.Extensive experiments on three vehicle Re ID datasets demonstrate the rationality and effectiveness of the three vehicle Re ID models proposed in this paper.It is helpful to promote the application of vehicle Re ID technology in traffic monitoring system.