Research On Occluded Person Re-identification Based On Fine-grained Feature Learnin

Person re-identification(re-ID)is to matching pedestrian images in scenes across multiple non-overlapping cameras,which is widely used in smart security,smart cities and other fields,and thus has received massive attention.In recent years,with the continuous development of deep learning,the performance of person re-ID algorithms in general scenes has been significantly improved.However,in real scenes,pedestrians are often occluded by different obstacles or other pedestrians,which leads to the reduction of visible regions and feature information loss and introduces occlusion noise.Occlusion can cause interference to the trained model,leading to difficulties in learning robust feature representations for person re-ID methods in general scenes,making the algorithm performance significantly degraded.Therefore,how to effectively learn robust pedestrian feature for occlusion is of great practical value for person re-ID tasks in realistic scenarios.In response to these problems,the specific research conducted in this topic is as follows:(1)Existing occluded person re-ID methods rely heavily on additional cues such as pose estimation and semantic parsing models to extract features,which inevitably brings problems such as complex model design and high computational cost.To address these problems,we construct a dual-branch feature aggregation network.First,pedestrian features from coarse-grained to fine-grained are extracted based on a chunking strategy,which makes the information contained in the features more comprehensive.Then,the spatial transformer network locates the upper body region of the pedestrian and applies a relation-aware attention module to it to explore the discriminative information in the upper body region and enhance the adaptability capability of the local features to the occlusion scene.Finally,feature fusion is performed to obtain a pedestrian representation that is robust to occlusion.Extensive experiments verify the effectiveness of the proposed method on the occluded dataset and its generalization on the holistic dataset.(2)Most existing Transformer-based occluded person re-ID algorithms learn global features,local features or cascading of local features directly for pedestrians while ignoring the relationship between features,which lead to incorrectly retrieving different pedestrians with similar attributes in corresponding parts as the same pedestrian.Therefore,we propose a dual-branch feature learning model based on Transformer.The global-local feature interaction module is used to learn the relationship between local features and global features,so that each local feature contains information about itself and relationship with other body parts,thus obtaining a more distinguishable representation.To further enhance the robustness of the model in the training phase,we simulate the occlusion of real scenes by randomly erasing local areas of the input image with the random erasure occlusion enhancement module.In addition,we retain some auxiliary recognition cues while learning the local discriminative features through the spilt group module.Experiment results on several public datasets demonstrate the effectiveness of the proposed method for the occluded person re-ID task.