Research On Vehicle Re-identification Algorithms For Complex Surveillance Scenarios

Vehicle re-identification is a cross-view vehicle image retrieval task,which is one of key technologies of intelligent video analysis.However,in practical unconstrained surveillance scenarios,vehicle re-identification has great difficulties due to a series of unfavorable factors,such as camera viewpoint variations,illumination channges,and vehicles’ high-speed movements.Recently,with the release of vehicle datasets and the rapid development of deep learning technologies,the research of vehicle re-identification has an opportunity.For that,this paper studies deep learning based vehicle re-identification algorithms for complex surveillance scenarios,which mainly includes three aspectes,as follows.(1)An efficient multi-resolution network(EMRN)based vehicle re-identification algorithm is proposed.Most existing algorithms usually resize vehicle images to the same resolution for feature learning,which could make high-resolution images lose detailed information or introduce noise when up-sampling low-resolution images,harming the feature learning effect.In our EMRN method,the deep network can be directly driven with multi-resolution images due to the specifically desinged multi-resolution feature dimension unified module and the multi-resolution image randomly feeding strategy.(2)An all-to-one attention(AOA)mechanism based vehicle re-identification algorithm is proposed.Because appearances of vehicles of the same model are highly similar,there are only differences in local parts,such as wheels,lights and ornaments on front windows.Existing algorithms use the attention mechanism to learn the differences in local parts,but they often learn the attention mechanism only according to the own local part’s appearance and ignore other parts’ appearances,causing a risk of bias in the attention learning from each part.In our AOA method,the importance of a single part is evaluated by all parts,so that all part could cooperate effectively in the process of attention learning,so as to improve the attentional learning effect.(3)A hybrid pyramid graph network(HPGN)based vehicle re-identification algorithm is proposed.Vehicles have different appearances at different spatial locations,therefore vehicle feature maps naturally contain different spatial significance.For that,our HPGN re-weights a node to learn the spatial significance of the corresponding position.Furthermore,The HPGN stacks multiple spatial graph networks to comprehensively learn the spatial significance of feature maps at multiple scales.At last,extensive experiments on multiple large scale vehicle datasets demonstrate that the proposed algorithm is superior to state-of-the-art vehicle re-identification approaches.Therefore,the study of this paper can provide a certain theoretical basis and technical support for the development of vehicle re-identification technology.