Research And Application On Human Gait Recognition Based On Tensor And Deep Learning Of Visual Data

Gait characteristics can be collected at a long distance without cooperation and are hard to disguise.Therefore the researches based on gait are of great importance in the fields of crime investigation,identity recognition,clinical medicine and so on.However,the gait characteristics can be influenced by view angular variation,different clothing,diverse carrying bags and so forth,which could descend the performance of gait recognition.In addition,in the process of obtaining gait contours,the damaged gait images should be processed properly.Otherwise,the rate of gait recognition will decline sharply.According to the problems aforementioned,the main contributions of this thesis are as follows:1.Investigate the history and status of gait recognition research.This thesis investigates the origin of gait recognition and the scientific basis of using gait as the biological characteristic for human identification.At the same time,this paper also reviews the research status at home and abroad,analyzes the categories of existing gait methods,and explains their advantages and disadvantages,which could provide some supports for this paper and other relevant researches.2.Research on how to process gait data by tensor modeling.When obtaining the gait contour,the silhouette of human can be influenced by a lot of noise because of the variation of shadows or other factors during the walking of the testee.In this paper,tensor modeling is applied to optimize the gait images,which can reduce the noise of gait images and improve the representation ability of gait images.3.Research on cross-view and cross-condition transforming by generative adversarial networks for gait recognition.Gait contours can change greatly because of angle and walking condition variations.Without the corresponding treatment,the diversity of gait contour in the same person can be much greater than the diversity between the difference individuals.Hence the recognition rate will drop sharply out of question.In order to solve those problems,this thesis proposes the identity preservation cycle gait generative adversarial networks to realise the transformation between different views and conditions.On the basis of the original generative adversarial networks,this thesis uses cyclic generation to improve the capacity of cross-domain transformation.At the same time,the identity preserving discriminators are added to ensure that identity information would be reserved during the process of transformation.The experiments on the large gait data sets of CASIA-B and CASIA-C have proved that the proposed gait recognition algorithms framework have better performance when solving angle and walking condition variations.