Research On Person Re-Identification Method Based On Target Alignment

Person re-identification comes from multi-camera tracking,and is used to judge whether person in different images captured in non-overlapping surveillance areas belong to the same person,which can be regarded as a sub-problem of image retrieval.Person re-identification technology has very important theoretical significance and application value.Person re-identification has become a valuable and challenging research topic in the field of computer vision due to the varied shooting angles of person images,inconsistent lighting conditions and diverse person postures.In recent years,with the continuous development of deep learning technology and the continuous introduction of large-scale person re-identification datasets,the accuracy of person re-identification has been greatly improved.In this paper,the person re-identification based on deep learning is deeply analyzed.On the basis of this,the idea of target alignment is used to introduce the person alignment operation in the person feature extraction stage,and the person image is converted to the standard image to realize the person image feature point.The mutual correspondence between them improves the detection performance of person re-identification.This paper proposes a person alignment method based on deformable convolutional network.By introducing a deformable convolution module in the feature extraction network,the sampling points can be dynamically adjusted to achieve person target alignment.Person re-identification frameworks that incorporate variable convolutional networks can achieve person alignment during end-to-end training without additional data annotation.Finally,the re-ranking is introduced in the person re-identification framework,and the position of the sample in the initial sorlting is adjusted according to the re-measured similarity to obtain a new search sequence,which further improves the retrieval precision.In this paper,we use the public datasets Market 1501 and CUHK03 to verify the person alignment-based re-identification framework proposed in this paper,and develop a person re-identification demonstration software based on Python.The experimental results show that compared with the original person re-identification method,the detection accuracy of the person re-identification framework based on variable convolution alignment has improved significantly on both datasets(mAP is respectively 63.63%/37.4%raise to74.60%/39.1%).It can be shown that the person alignment method proposed in this paper can effectively improve the detection performance of person re-identification and has better robustness.