| With the increasing development and improvement of surveillance networks in urban life,video reconnaissance technology has been widely used in real-world scenarios.At the same time,the research on pedestrian re-identification task technology has also achieved remarkable results.Existing pedestrian re-identification discusses how to extract more representative features of identity in order to achieve accurate pedestrian recognition from the aspects of illumination impact,occlusion degree,and perspective changes.However,this is all based on the assumption that pedestrians do not significantly change the appearance of their clothes in a short period of time.Once the target pedestrians change their clothes naturally after a long period of time,or change their clothes intentionally to avoid capture,the conventional methods that rely heavily on appearance characteristics are in this challenge.There will be a noticeable performance drop.In the past two years,research on long-term pedestrian re-identification has mainly focused on extracting identity-discriminative pedestrian features,including facial features,body shape/contour features,etc.However,these methods ignore the need for sufficient images of well-dressed pedestrians for training in order to obtain discriminative identity features.Different from short-term pedestrian re-identification,in long-term scenarios,it is very expensive and time-consuming to acquire a large amount of pedestrian data that changes appearances,resulting in the existing long-term re-identification datasets that are not rich in data and sample Space is sparse.Therefore,this paper conducts research on long-term person re-identification tasks from the following aspects:(1)Aiming at the problem of limited data volume and sparse sample space,this paper proposes a data augmentation method based on semantic awareness.Saving manpower and financial resources,this augmentation method greatly effectively enriches the amount of training data and the clothing styles.In this way,the model is guides to pay attention to the discriminative features of pedestrians outside the clothing area while reducing the model’s perception sensitivity to prominent clothing areas.(2)Aiming at the problems of sparse samples in distribution and missing samples out of distribution,this paper presents a re-identification method based on distribution complementarity.From the complementary perspective,this method designs out-of-distribution data samples without additional interference information on the basis of complementing the in-distribution sample space.With this strategy,the out-of-distribution space complementary to the in-distribution space is filled to guide the model to learn better class boundary information of person category.(3)Aiming at the problem that the appearance of pedestrians changes greatly in the changing scene,this paper proposes a multi-granularity feature learning method based on adaptive attention.From the perspective of extracting diverse pedestrian features,this method adaptively explores pedestrian information at different scales,guiding models to obtain person uniqueness information from pixel-level,part-level,and object-level.To verify the effectiveness of the proposed method,this paper uses two public long-term person re-ID datasets,LTCC-Re ID and PRCC-re ID,as well as two widely used short-term person re-ID datasets,Market-1501 and Duke MTMC-re ID.Extensive experiments conducted on them demonstrate that the method proposed in this paper can effectively alleviate the challenges brought by clothing changes in long-term re-identification.Compared with other existing methods in long-term pedestrian re-identification,the method proposed in this paper is more outstanding and more competitive. |
