Re-ranking Vehicle Re-identification With Attribute And Background Segmentation

With the improvement of living standards,the number of private cars is growing as a mainstream means of transportation.At the same time,the increase in the number of vehicles has become more and more obvious to public transport and social security,making the need to monitor and track vehicles even more urgent.The traditional method of monitoring and tracking vehicles by hand not only wastes a lot of manpower and resources,but also generates a lot of false alarms and missed alarms.Therefore,the research of vehicle re-identification was born.The essence of vehicle re-identification is cross-camera target identification,by identifying the same target vehicle through images acquired by different cameras.As vehicles are rigid objects,there are large similarities between different vehicles in the same series,while for the same vehicle,factors such as different viewing angles can cause large differences in vehicle appearance information,resulting in existing vehicle re-identification algorithms facing greater challenges due to the too small interclass gaps between vehicles as well as large intra-class gaps.At the same time,the background information in the original vehicle image often interferes with the performance of vehicle re-identification.Current research has achieved good results in this area by removing or re-fusing background interference information through background segmentation algorithms.However,such methods rely heavily on the actual effect of the background segmentation algorithm,and poorly performing algorithms may result in useful foreground information being removed or useless background interference information still being retained.Therefore,to address the above difficulties,the main research work in this paper is as follows:First,to address the problem of large intra-class gaps in vehicle re-identification across cameras,a feature fusion module is proposed in this paper.Considering that attributes are good vehicle auxiliary information,after first extracting the corresponding global features,vehicle color features and vehicle model features from the original vehicle image,the vehicle color features and vehicle model features are weighted and summed,and then concatenated with the global features as input to the feature fusion network to obtain vehicle features with strong discriminative power.A distance metric is performed by this feature to improve the ranking of positive samples in the result list.Second,to address the influence of background interference information on vehicle re-identification,this paper proposes a re-ranking vehicle re-identification algorithm with attribute and background segmentation.The algorithm first extracts the background image through the background segmentation module,then obtains the background image features through the feature extraction module,which is used to assist in optimising the ranking results of the fused features obtained by the feature fusion module,i.e.,finally,in the inference stage,the ranking results obtained based on the strong discriminative features are re-ranked through the differences between the backgrounds,so as to reduce the influence of the background differences of the original vehicle images on the recognition results and improved the vehicle re-identification performance.Experiments on the VeRi-776 and Veri-Wild datasets show that the proposed algorithm achieves 96.0% accuracy in Rank1 on the Ve Ri-776 dataset,an improvement of 1.7%?3.1% compared to other algorithms.The Rank1 accuracy on the three subsets of Veri-Wild reached over 90%,an improvement of 1.2%?2% compared to other algorithms.It shows that the algorithm can effectively reduce the negative impact of vehicle appearance differences and background interference information on vehicle reidentification across different camera views.Finally,this paper builds a corresponding vehicle re-identification system based on the proposed above algorithm.