A Research On Occluded Person Re-Identification Under Deep Learning

With the increasing demand for video surveillance and the continuous expansion of camera network scale and application scenarios,person re-identification(Re-ID),as an important biometric identification technology,has become an important part of intelligent video surveillance systems and has been extensively studied.As an effective identification technology,Re-ID mainly solves the problem of pedestrian retrieval under cross-camera and cross-scene.However,the real scene often contains a large number of unavoidable occlusions.Thus,the camera cannot capture the complete body,and the collected pedestrian images are blocked by obstructions,making the Re-ID method ineffective.To solve the problem of occlusion in noncontrollable scenes,this paper builds a basic framework for occluded person re-identification(Occluded Re-ID)based on data preparation,global feature extraction,fine-grained feature mining and fusion,and visible feature matching and alignment under deep learning.The following work was carried out for the research subjects:1.Based on the analysis of the dataset related to Occluded Re-ID,the effects of different data augmentation method combinations are studied,and an modified random erasing method is proposed to introduce simulated occlusion;a global perception residual network model is constructed as backbone network to extract more informative and robust global feature.2.We propose a keypoints-based information fusion and alignment method.From the perspective of keypoints information,this method combines global information fusion and local information mining to achieve fine-grained feature extraction,and completes the suppression and filtering of the occluded part according to the confidential information of the keypoints,and realizes the localization by matching the visible shared area for feature alignment,the recognition accuracy is better than the current mainstream Occluded Re-ID methods.3.We propose a method based on multi-attention and part alignment.From the perspective of semantic regions,this method uses an additional multi-attention module to suppress the influence of pose estimation errors caused by domain gap;on the other hand,it uses the keypoints based semantic regions of the human body to replace the original horizontal partition method,which improves the discrimination of local features.The recognition accuracy has been further improved,and it owns strong effects for solving occlusion problems.4.We propose a graph attention based information fusion method.From the perspective of high-order information aggregation,this method introduces the graph attention module and constructs an efficient neighborhood connection pattern according to the relationship between different local features.Through the fusion of the effective information of the neighborhood in the graph attention network,the final identity representation owns good discrimination and robustness,it realizes the suppression of the occlusion information and the domain gap introduced by the semantic model.This method could achieve good results on Occluded Re-ID datasets and holistic Re-ID datasets.