Person Re-identification Based On Multi-scale Feature And Siamese-GAN Network

Pedestrian re-identification is the extended direction of pedestrian detection and recognition in cross-camera,its results are close to real life.Unlike single-camera pedestrian recognition,cross-mirror tracking brings more difficulties to recognition because of its incoherence.Cross-camera tracking brings more challenges to recognition due to its discontinuity than single-camera pedestrian recognition.Therefore,this thesis has carried out some integrated research on pedestrian re-identification algorithm based on twin generation confrontation network which of multi-scale feature fusion,and has achieved certain research results.This thesis descript the technical research and concrete implementation of pedestrian reidentification technology based on deep learning in detail.Firstly,this research tends to continuously updating the structure of deep networks or adding network modules to networks,which increases the depth of the network and the cacualtion cost.This article proposed to add the SE(Squeeze and Excitation)module to the residual network,using its compression and excitation methods to selectively enhance and suppress the feature channel,in order to improve the accuracy of pedestrian re-identification.At the same time,shallow features and deep features are merged and deleted the highest-dimensional feature extraction module.The model calculation has been reduced;network layers are reduced,the calculation cost is saved.Secondly,a detailed ablation experiment was conducted on the influence of the size selection of the convolution kernel on the network operation speed and time consumption,compared with kernel size,model calculation and time consumption,and found the best parameters value to balance the time consumption,the amount of calculation and accuracy.The method of mixing human pose estimation and multi-scale feature,firstly the positive and negative sample pairs are input to the twin generation adversarial network,and the Siamese network perform the following operations on the positive and negative sample pairs: Using the human pose estimation model to segment body parts based on their posture,then extract local feature for each body part and extract global feature of suspicious target pedestrians,and then use multi-scale feature fusion method weights fusion local features and global features of pedestrian.After extracting the fused features,puting it into the adversarial network for image generation.Finally,the discriminator determines whether the images generated by the positive and negative sample pairs are from the same pedestrian.Through the above experimental comparison,it is found that feature extraction has the higher impact on the accuracy when combined with first module and the third module,as well as the fourth module is been deleted,while the convolution kernel size is 7.Human pose estimation and multi-scale features for feature extraction can highlight the object feature details,which makes the pictures by the generation adversarial network became more realistic,ultimately improves the accuracy of the algorithm.In general,this topic has made some improvements in the application of pedestrian reidentification technology on the classic deep network,and carried out some fusion experiments on the frontier technology to provide a positive theoretical and practical basis for subsequent research.