Research On Low-frame-rate Gait Recognition Based On Generative Adversarial Networks

Gait recognition,as a new biometric recognition method,recognizes the identification of individuals through their walking postures.It does not require the cooperation of the detected object,and has the recognition potential under long distance conditions.With the development of intelligent monitoring,gait recognition has broad application prospects.Gait recognition uses video sequence data to identify people's identity information over long distances,and video sequence data will be limited by frame rate issues,which will directly affect the performance of the recognition model.Most existing gait recognition methods require proper data preprocessing.The quality and continuity of gait sequences will be affected to a certain extent.Gait contour sequence diagrams with poor segmentation quality are often discarded to reduce potential noise.Based on the generative adversarial network algorithm,this paper conducts in-depth research on the low frame rate of gait recognition cross-view problem.It proves the effectiveness of our method and improves the accuracy of gait recognition.The research innovations are as follows:1.An algorithm to increase the gait frame rate using a generative adversarial network is proposed.The generative adversarial network of this method consists of a generator and a discriminator.We introduce the WGAN structure in order to stabilize the training of generative adversarial network.The frame sequences with good quality are generated through the mutual game learning of the generator and the discriminator.The frame interpolation effect on the original gait contour sequence is realized,which enriches the gait sequence information and improves the accuracy of gait recognition.2.An algorithm for introducing Margin Ratio Loss function in gait classification and recognition network is proposed.The loss function of the experimental benchmark classification recognition network cannot effectively classify the samples.We are inspired by metric learning,which makes the samples distance of different classes large and the samples distance of the same class small.Therefore,the Margin Ratio Loss function is introduced,which further improves the classification recognition accuracy.In order to verify the effectiveness of our methods,our experiments are performed on the commonly used CASIA-B dataset and OU-ISIR dataset,and compared with other representative algorithms.Experimental results show that the above two algorithms have a significant improvement over the benchmark.