Research On Recognition Of Cross-View Human Gait Based On View-Invariant Features

With an increasing attention paid to personal and social security at home and abroad,biometric authentication technology has gained more and more attention.Gait,as a unique biological feature,has important significance in remote and uncontrollable situations,especially in the fields of intelligent security and criminal investigation,where its application value is more prominent.At present,human gait recognition still faces many challenges,such as view angles,clothing occlusion,and carrying objects.When the view angle changes,the gait and walking trajectory of the human body will significantly change,which will lead to a sharp decline in the accuracy of gait recognition.Aiming at the problem of cross visual gait recognition,this thesis summarizes the research status and related technologies at home and abroad,and then proposes two cross view gait recognition algorithms based on view-invariant features.The main research work of this thesis is summarized as follows:Firstly,this thesis researches on cross view gait algorithms based on hybrid attention model and generation of adversarial networks.By training a unified generated confrontation network,gait energy maps from different perspectives are converted to the same standard perspective,thereby avoiding errors caused by differences in perspective.By combining the advantages of spatial attention and channel attention,the model extracts gait feature information from spatial and channel dimensions.By introducing an identity discriminator,gait information can be effectively saved during the conversion process.At the same time,pixel loss is added to restore pixel level gait details,which makes the generator more stable during the training phase;A triple loss function is introduced into the generator to narrow the intra class distance and push the inter class distance further.After many experiments,the feasibility and effectiveness of this algorithm have been proven on two classical gait datasets.Secondly,this thesis raises a cross view gait recognition algorithm based on discriminant feature learning.Through discriminative feature extraction and spatiotemporal feature learning,this algorithm constructs a new cross perspective gait recognition network to extract perspective invariant identity features from gait contour sequences from multiple perspectives to accurately identify identities.This algorithm first addresses the shortcomings of existing loss functions and proposes a joint loss function suitable for cross visual gait recognition based on traditional loss functions.Then,the attention weight of each frame is learned through learning horizontal segmentation and long-term and long-term memory network attention models,and effective spatiotemporal features can be extracted to obtain discriminative gait features,further improving the performance of the network.After many experiments,it has been proved that this algorithm exhibits significant advantages when dealing with cross visual gait recognition problems.