Unsupervised Person Re-identification Based On Deep Learning

With the rapid development of computer vision technology,video surveillance technology has been widely used in the field of intelligent security.As an important technology to solve the problem of cross-camera tracking in the field of video surveillance,person re-identification has a broad application prospect..In recent years,deep learning has developed rapidly.The research on person re-identification based on deep learning has gradually replaced the traditional method of extracting features by hand and become the focus of both industry and academic circles.Its task is to match the pedestrian query images input into the person re-identification system in the massive pedestrian database.In the field of pedestrian re-recognition based on deep learning,supervised person re-identification requires a large amount of tag information guidance model training.Due to the high cost of sample labeling in the actual situation,research on unsupervised person re-identification that does not require a large amount of tag information guidance has been paid more and more attention.Research on unsupervised person re-identification is in its initial stage.There is a huge difference between source domain data and target domain data,which is a key factor affecting the effect of model recognition.However,the research of unsupervised person re-identification based on clustering method ignores the influence of the huge difference between source domain and target domain on the performance of the model.In addition,the pooling strategy in the backbone network of the unsupervised person re-identification is too single,which may easily lead to the loss of a lot of characteristic information.The above problems,the author of this paper put forward a kind of style translation based on unsupervised clustering person re-identification algorithm,first of all,in order to reduce the influence of the difference of the model of the domain,the model based on clustering method will be introduced to the cross-domain style transfer method,reduce the influence of the difference of model recognition effect of the field,and improve the style transfer model,further enhance based on unsupervised clustering method of heavy person re-identification model recognition effect;Secondly,aiming at the problem that the pooling strategy of backbone network is too single,this algorithm proposes a multi-pooling fusion strategy based on the idea of segmentation to be added into the model testing process,so as to further improve the model recognition effect.The algorithm proposed in this paper mainly includes the following two aspects.(1)Aiming at the problem that the current unsupervised person re-identification method based on clustering ignores the huge difference between the source domain and the target domain,which affects the performance of the model,an unsupervised clustering person re-identification method based on style transformation is proposed.Firstly,aiming at the problem that the unsupervised person re-identification model based on clustering is greatly affected by the difference between source domain and target domain,a style transfer method for person re-identification is introduced into the clustering method model,and the style transfer method is used to directly reduce the difference between domains and improve the recognition performance of the model.Secondly,the generator of cross-domain style transfer model has the problem of single transfer scale and low efficiency of characteristics information transfer,using a new type of residual block to replace the original residual block and into the generator to up-sampling and the down-sampling process,forming more characteristics of the scale transfer and information transmission efficiency of the generator,make the style transfer model can better against the source domain and target domain to transfer the style,reduce the difference between the two domains,further enhance the effect of the whole model.(2)Aiming at the problem of the backbone network general has a single pooling strategy in the unsupervised person re-identification,the problem can lead to a large amount of feature information missing.In this paper,the unsupervised clustering person re-identification model based on style transformation is improved.A multi-pool fusion strategy based on the idea of segmentation is proposed for model testing to further improve the recognition effect of the model.First,the feature map extracted from the backbone network is copied and segmented by height to obtain the features of the upper body,the lower body and the overall pedestrian.Secondly,the multi-pool fusion strategy of average pooling and max pooling are used respectively for the three characteristics.Finally,the expanded features of each part after pooling are spliced together to represent the characteristics of pedestrians.At present,the effectiveness of the proposed algorithm is verified by experiments on the setting of universal data set for unsupervised person re-identification.The results show that the proposed improved method achieves better recognition effect.