Research On Cross-view Gait Recognition Algorithms Based On Deep Learning

With the improvement of graphics computing capabilities and the advent of the era of big data,some research fields such as pattern recognition and computer vision are getting more and more attention and development.Recently,some biometric recognition technologies such as face recognition,fingerprint recognition and vein recognition are becoming more and more demanding and important in various scenarios of our daily lives.As one of the biometric recognition technologies,gait recognition distinguishes different individuals based on their walking postures.Different from face recognition,fingerprint recognition and vein recognition,gait recognition has the advantage of being able to identify subjects under long-distance,uncontrolled and low resolution condition.What is more,gait is hard to change and disguise,which makes it more and more valued in recent years.However,the gait samples captured in the monitoring system often differ to the registered gait samples in view,which seriously weakens the effectiveness of the traditional gait recognition algorithms.Therefore,it is very significant to solve the problem of cross-view gait recognition and improve the accuracy and real-time performance of gait recognition products.The difficulty of cross-view gait recognition is that the pedestrian gait has huge visual differences in different views.In addition,the gait feature only contains the information of pedestrian's shape and walking posture,but does not include color or texture features,which makes it difficult to extract the view-invariant features that represent the identity of the pedestrian from different gait views.In order to solve the low accuracy problem of cross-view gait recognition,this thesis studies the view-invariant gait feature extraction methods based on deep learning for cross-view gait recognition.Specifically,the main contributions are summarized as follows:(1)A cross-view gait recognition method based on the decomposition of identity information and view information in gait features is proposed.A new auto-encoder is designed to detach the identity feature and the view feature,which consists of two encoders(the view encoder(VE)and the identity encoder(IE))and a gait decoder(GD).Specifically,a view regression loss is designed for the output of the view encoder to learn the view features of input gait samples.For the output of the identity encoder,the triplet loss is used to metric the similarity of different identities.In order to rid the view encoder output of identity information,an identity fuzzy loss and an identity classification loss are designed for the view encoder and identity encoder,respectively.In addition,a decorrelation loss function is designed to minimize correlation between the output of VE and IE.The experimental results on the CASIA-B and OU-ISIR databases show that the proposed method can detach the view information and identity information of the gait sample.(2)A Block-wise Spatial-temporal Representation Aggregation(BSRA)framework is proposed to extract view-invariant gait features for cross-view gait recognition.The proposed framework includes a Block Horizontal Pyramid(BHP)module and a Spatial-Temporal Feature Aggregation(STFA)module.Specifically,BHP is built firstly in order to extract part-wise gait features from gait silhouettes of multiple scales.STFA is then used to aggregate these features and produce final representation,which includes a block attention splicing module to aggregate spatially and a maximum gait energy module to aggregate temporally.In addition,a multi-view loss based on Chebyshev distance is presented to force the proposed BSRA learning view-invariant features and accelerate training convergence.Extensive experiments are conducted on two multi-view gait benchmarks,i.e.,CASIA-B and OU-ISIR.The results demonstrate the effectiveness of the proposed method,which outperform state-of-the-art methods in most cases Experiments under backpacks,different dresses and short-sequence conditions show that the proposed method is still robust and reliable in challenging environments.In addition,multi-view experiments show that the proposed method can achieve good performances under different number of gallery views and the trend of multi-view average rate is obtained.