Human Identification Based On Fusion Features Of Gait Silhouette And Skeleton Under Incomplete Information

Gait-based identification is a non-intrusive and convenient biometric technology that has garnered significant attention from researchers and is increasingly being employed in practical applications due to its resistance to imitation and adaptability to varying environmental conditions.It has been gradually deployed in practical applications.Existing gait identification technologies usually extract frame-level features first and then extract sequence-level features with global information.Sufficient sequence information is crucial for gait identification.However,in practical applications,factors such as bandwidth,energy consumption,memory,and real-time requirements will limit the number of frames within an identification cycle,resulting in a problem where identification models can only perform calculations based on incomplete information.To address the above problems,this thesis seeks to enhance the accuracy of gait identification algorithms under incomplete information conditions by researching on improving the learning ability of feature extraction networks,integrating multi-modal gait features,and supplementing sequence information based on motion prediction methods.Specifically,this thesis proposes a multi-feature fusion-based gait identification method that improves the information mining ability of the identification network under given conditions.A human gait prediction method is then proposed to expand the learning information that the identification network can learn under incomplete information conditions.Lastly,a real-time gait identification application is implemented to put the proposed methods into practical use.The main work and contributions of this thesis are:(1)A two-branch gait identification model is proposed,which combines appearance-based and model-based methods by first extracting features of the two modes separately and then fusing them using a cross-attention mechanism.This model uses the complementary properties of these two modal features to eliminate the impact of appearance-based methods on clothing and backgrounds as well as to compensate for the limitations of model-based methods caused by inaccurate posture estimation algorithms and limited initial data amount.The proposed gait identification algorithm based on feature fusion obtained an average accuracy 94.9% on the CASIA-B dataset,which outperformed the existing single-modal algorithms and multi-modal algorithms based on simple fusion strategies.(2)A general non-autoregressive model-based framework for predicting gait silhouettes and 3D skeletons is proposed to supplement data by predicting subsequent frames effectively and relieve the challenges brought by incomplete available information in practical gait identification scenarios.The proposed gait identification algorithm based on information enhancement through feature fusion performs better than other methods based on other information enhancement strategies or methods without enhancement under incomplete information conditions.The algorithm obtained an average accuracy of 91.3% with an input sequence length of only 10 frames.(3)A real-time solution for gait-based identification is designed by reasonably breaking down the requirements and formulating system modules.Relevant technologies and middleware are introduced to decouple the modules and implement an easy-to-deploy,scalable,and easy-to-maintain real-time identification application.The proposed algorithms in this thesis are deployed in the application to provide users with scene monitoring and identity inference functions.Tests on functionality and performance show that the system can meet the functional requirements and realtime requirements of gait-based identification.The identification accuracy can reach 86.9% in practical tests with an input sequence of only 10 frames,demonstrating good practical application effectiveness.