Research On Person Re-identification Based On Deep Learning

With the continuous development of smart and connected cities,computer vision technology based on deep learning has been widely used in many practical scenarios.Person re-identification,as an important branch in computer vision,aims at the extraction of target person under the non-overlapping area of multiple surveillance cameras,i.e.,cross-camera tracking and retrieval of person.The research of person re-identification is important for many fields such as smart security,smart business and epidemic trajectory tracking.However,the person images captured by each surveillance camera have a series of problems such as different illumination,different viewpoint poses and different image resolutions,which bring great challenges to the person re-identification.So it is crucial to extract person features with strong discriminative power from the images.In addition,for the situation that the person images obtained from each surveillance camera have very different styles,eliminating the domain differences among person datasets can improve the accuracy of person re-identification across scenes.After analyzing the current status of person re-identification research and related technologies,the following research work and algorithm innovation are focused in this paper:Aiming at the problem that most of the existing person re-identification algorithms cannot obtain features with strong discriminative power of person,a person re-identification algorithm based on multi branch with fine-grained feature fusion is proposed.Firstly,Res Net50 is used as the backbone network to extract person shallow features,and then a dual-attention global feature extraction module and a local feature extraction module are constructed.The dual-attention global feature extraction module enhances global feature extraction by adding spatial attention mechanism and channel attention mechanism.The local feature extraction module is divided into feature vector horizontal two-slice branch and horizontal four-slice branch,focusing on local detail information by different granularity and level of feature slices.The model is jointly trained by using softmax loss function and triplet loss function after adding batch normalization layer.Extensive experiments show that this model achieves 89.8%,95.9% and 80.5% of Rank-1 accuracy on Duke MTMC-re ID,Market1501 and CUHK03 datasets,proving the effectiveness and advancement of the proposed person re-identification algorithm based on multi branch with fine-grained feature fusion.To solve the problem of severe performance degradation when person re-identification algorithm is applied unsupervised learning across scenes,a structure-invariant cross-domain adaptive person re-identification algorithm is proposed in this paper.After analyzing the Cycle GAN,considering that general generative adversarial network can only achieve style transfer between two camera domains,improved Star GAN is used to achieve image style conversion for multiple camera domains of person dataset.The multi-scale structural similarity loss function is used to ensure that achieving image style transfer in the target domain do not cause person structure distortion.Through mutual style transfer learning experiments between Market1501 and Duke MTMC-re ID datasets,the results of Rank-1accuracy of 58.9% and 44.7% are obtained,which verifies that the unsupervised domain adaptive model proposed in this paper has good generalization ability in the new scene,and the problem of domain gap between person re-identification camera domains is solved to some extent.Finally,the complete end-to-end cross-domain person re-identification is achieved by cascading the multi branch with fine-grained feature fusion algorithm and multi-scale structure invariant Star GAN proposed in this paper.When using Market1501 and Duke MTMC-re ID as the target domain,compared to the direct transfer of the model,the Rank-1 accuracy is improved by 22.5 and 25.1 percentage points,which effectively improves the accuracy and robustness of person re-identification across scenes.