Research On Pedestrian Pose Transformation Method Using Generative Adversarial Networks

Person Re-identification is one of the tasks of pedestrian monitoring derived from the popularity of today's camera networks.With the rapid development of deep learning,pedestrian recognition technology has also evolved into an important challenge in the field of computer vision.In the application of camera network,the search of missing persons,the pursuit of criminals,and the investigation of suspects all have important practical significance.However,pedestrian images and videos acquired by camera networks generally face a problem that they have low pixels,so the face recognition method for person recognition is not applicable.Thereby the Person Re-identification method is proposed for a low-resolution pedestrian image or video.At present,the Person Re-recognition algorithm based on deep learning has achieved good results on some pedestrian datasets,which has been greatly improved compared with the traditional methods.However,in reality,when faced with more complex scenarios of the person,there are still many problems with pedestrian recognition technology.First of all,the existing pedestrian datasets are still not large enough,and the number of pedestrian samples is not enough to achieve better Person Re-recognition.Secondly,there are multiple poses among pedestrians.The pose misalignment of pedestrians of same identity makes it difficult for deep neural networks to Extract uniform feature information to ensure correct classification results.Therefore,in recent years,the Generative Adversarial Network is used for the pre-processing of pedestrian samples,and the above problems of Person Re-recognition are solved by expanding the datasets and the method of pedestrian pose normalization.Using the Generative Adversarial Network to generate specific pedestrian samples is one of the auxiliary technology of Person Re-recognition.Firstly,the pedestrian key point prediction and pedestrian attitude estimation method are used to obtain the pedestrian's attitude information.Then based on the Generative Adversarial Network,the pedestrian samples and the pedestrian attitude are used to generate the false pedestrian samples of the same pedestrian with different pose to expand the pedestrian datasets or achieve pedestrian pose normalization.However,the pedestrian samples generation work is also facing an urgent problem to be solved.First of all,the pedestrian samples in the pedestrian datasets have complex background information in addition to the pedestrian part.This part of the information may interfere with the generation of pedestrian samples.Secondly,there is a certain error rate in the pedestrian pose obtained by the existing pedestrian attitude estimation method.This can seriously affect the reliability of generating and combating network training effects.In this paper,the pedestrian segmentation method based on semantic segmentation technology is used to obtain the segmentation mask of the pedestrian samples in pedestrian datasets.The pedestrian samples segmentation mask is used to optimize the method of obtaining pedestrian pose and the optimization of pedestrian pose transformation samples generation.In this paper,we first obtain the binarized segmentation mask of pedestrian image in pedestrian dataset by using pedestrian segmentation technology.After obtaining the pedestrian image without background by using the pedestrian segmentation mask,a pose conversion work using the Generative Adversarial Network for the pedestrian image without background is proposed.Then,in order to solve the problem of pedestrian pose mask error acquired by the pedestrian pose estimation method,a pedestrian pose mask correction method guided by a pedestrian segmentation mask is proposed.By using the pedestrian contour information contained in the pedestrian segmentation mask and the implied pedestrian pose information to correct the wrong pedestrian pose mask,and using the mapping relationship established by the Generative Adversarial Network to achieve obtaining more correct pedestrian pose mask by input any correct pedestrian segmentation mask.After that the generated image is used instead of the original incorrect pedestrian pose mask.For the problem of inaccurate stature of pedestrian images generated by the pedestrian attitude transformation method guided by the pedestrian pose mask,this paper proposes a pedestrian attitude transformation scheme guided by the pedestrian segmentation mask.We can generate more accurate pedestrian contours by using this method.And for the generated samples which has been removed the background we can use the background mask to restore the background to achieve the effect of retaining more effective information.Finally,for implementing the optimization of Person Re-identification technology,the paper proposes to connect the Generative Adversarial Network and the Person Reidentification network.The generated fake samples input the Person Re-identification part can produce a Re-identification loss(Reid?loss)which will guide the generator to optimize itself to generate a better pedestrian sample that is more conducive to Person Re-identification network.Then based on the method proposed in this paper,using the extended datasets and pose normalization method to optimize the Person Reidentification method results.In a word,the paper proposed some new ideas for using the Generative Adversarial Network to realize pedestrian pose transformation,and the reliability of the proposed methods are verified by a large number of experiments.