Gait-based Pedestrian Identification Research

With the rapid development of technologies such as artificial intelligence and computer vision,pedestrian identification based on gait has received extensive attention and development.However,despite the high availability of gait data,the recognition rate may be affected by unfavorable factors such as shooting angle and walking state.In order to eliminate the influence of unfavorable factors and extract effective gait features,this paper proposed a pedestrian identification algorithm based on GAN view conversion.The algorithm could convert gait feature maps from different views and states into a 90° side view of normal clothing without objects.In addition,in order to improve the authenticity of the generated gait image and make it contain more identity information,an identity discriminator was introduced based on the traditional generative adversarial network(GAN).After analyzing the GEI features of different viewing angles and states,it is found that the recognition performance of the model at three typical angles gradually decrease with the increase of the viewing angle span,and the highest recognition accuracy is achieved at 90° viewing angle.The experimental results on the CASIA-B dataset show that it is feasible to realize pedestrian identification based on the GAN view conversion network.Based on the GAN view conversion network,this paper further studied and proposed a pedestrian identification algorithm based on multi-view fusion.By coordinating one generator and three discriminators(true and false discriminator,identity discriminator,and view discriminator)of GAN,the view transition and state transition of three target views(54°,90°,126°)were realized.In addition,in order to reduce the loss of feature information during the view conversion process,this paper introduced a residual structure into the network to retain more identity information.In order to determine the identity of pedestrians,identity recognition models ware established for three conversion views,and the recognition results of the models were fused by majority voting in the decisionmaking stage.In order to verify the effectiveness and generalization of the model in this paper,CASIA-B and self-built small gait dataset were used to evaluate the performance of the network.The experimental results show that the model in this paper has achieved better recognition results in the state of backpack and coat.Compared with other methods,the model in this paper has better performance,and it is more robust to changes in view and clothing,and can identify pedestrian identities more accurately.