Research On Cross-view Gait Recognition Algorithms Based On Temporal-spatial Feature Learning

With the continuous development of science and technology,biometric recognition technology,which is in the field of computer vision and pattern recognition,has developed rapidly.At present,face recognition,DNA recognition and other recognition technologies have been widely used in daily life,national security and other fields.Gait recognition technology is an emerging biometric recognition technology.Compared with traditional biometric recognition,gait recognition has the advantages of non-control,and is not sensitive to interference such as environmental light intensity.Gait samples can be acquired at a long distance,which means high concealment,so it can reduce the possibility of camouflage and has the excellent discriminability.Based on the above advantages,gait recognition technology has been paid more and more attention and has broad development space.However,there are still some problems that restricting the development of gait recognition technology.In the real world,the camera is generally used to collect gait features,however,these cameras are often fixed in the different locations,which leads to the acquisition of different views of gait.And with the change of view,gait samples will also change significantly,which brings a great challenge to gait recognition.Therefore,how to effectively solve the cross-view gait recognition problem determines whether gait recognition technology can be further developed and applied.Aiming at the problem of cross-view gait recognition,this thesis studies the cross-view gait recognition method based on temporal-spatial feature learning,and designs the highperformance deep learning networks to extract robust view-invariance gait features.In this thesis,the main contributions are as follows:First,the cross-view gait recognition based on temporal augment and spatial multi-scale fusion is proposed.First of all,the gait silhouettes are sliced into the whole body,upper body and lower body and sent into three branch networks to extract the gait feature information from different parts and scales.For each branch network,a Temporal Augment and Spatial Multiscale Fusion(TASMF)network is proposed to extract multi-scale features.Specifically,the TASMF network firstly uses the multi-stage feature extraction part to extract gait features from shallow layers to deep layers.Then,aiming at sequence information,the Temporal Attention Extraction(TAE)module is proposed to assign weights to different temporal features,which can make the network focus on the more important temporal information;Last but not least,the Multi-Scale Pyramid(MSP)feature fusion module,which employs the dilated convolution,is proposed to realize multi-scale fusion of gait features in spatial dimensions,and the multiple stages of gait feature horizontally pyramid feature mapping.Cross-view gait recognition algorithm based on temporal augment and spatial multi-scale fusion can effectively make use of time and space information,a series of experiments on the CASIA-B,OU-ISIR and OUMVLP gait datasets show the effectiveness of the algorithm.Second,cross-view gait recognition algorithm based on fine-grained temporal-spatial feature fusion is proposed.First of all,we propose the temporal-spatial fusion gait recognition structure,which can be summarized as three layers according to the hierarchical extension order.Specifically,the backbone is used to extract the basic gait features in the basic feature extraction layer.Then,aiming at the temporal feature dimension,a temporal feature fusion layer is constructed and a Motion Texture Mixture(MTM)module is proposed.MTM module can realize frame-level feature mixture,which makes each frame gait feature fuse the gait information between adjacent frames in order to extract the motion texture and enrich the temporal feature.Last but not least,refocus on the spatial-level,Fine-grained Multi-stage Feature Extraction(FMFE)is proposed in the spatial features mapping layer.It realizes further refining gait feature in the spatial-level.The addition of a spatial attention module to the FMFE module allows for more useful fine-grained gait features.A series of experiments on the CASIA-B,OU-ISIR and OU-MVLP gait datasets show excellent performance of the proposed algorithm.