| Wi-Fi-based gait recognition is a passive user identification method that does not require users to wear devices,but the path sensitivity of Channel State Information(CSI)hinders its application in multi-path environments,which aggravates the sampling and Deployment cost(i.e.large number of samples for training and multiple specially placed devices).While video-based ideal CSI generation method are expected to significantly reduce samples,the lack of environment-related information in ideal CSI makes it unsuitable for general indoor scenes with multiple walking paths.In this paper,we propose WiVi,a multi-path gait recognition framework based on cross-modal fusion of WiFi and video,which has the advantages of few training samples and low device deployment cost.The specific method is to judge the walking path by introducing WiFi-based human tracking results and represent the gait features by introducing simulated WiFi CSI based on video,that is,path recognition based on human tracking and WiFi-video Cross-modal fusion based multi-path gait recognition system.(1)Path recognition based on human tracking.The path judgment based on human tracking can realize the path recognition of a small number of samples,but due to the influence of WiFi noise,it is difficult to find the starting point of walking and the deviation of walking tracking is large.For the walking detection problem,we use principal component analysis to denoise CSI,then use short-time Fourier transform to generate a spectrogram,extract the variance caused by walking,and finally use an exponential moving average algorithm to detect the starting point of walking.For efficient classification problems,we utilize SG filtering to smooth walking trajectories and extract trajectory trends to eliminate bias.Finally,the K-nearest neighbor classifier is used for path recognition to determine whether the user is walking on a predefined path.(2)WiFi-video Cross-modal fusion based multi-path gait recognition system.Video-based simulated WiFi CSI can retain more comprehensive gait information to reduce the number of training samples required,but faces two main problems.First of all,the existing methods only support the simulation of CSI signals that walk on a specific path perpendicular to the WiFi propagation link,and directly applying to other paths will lead to excessive signal errors.We optimized the simulation signal generation process,extracted more comprehensive gait information features,and used a variety of filtering techniques to remove noise from real WiFi data.Finally we employ two separate neural networks(NNs)to enable context-aware comparisons between ideal and measured CSI.The first network is supervised by measured CSI samples and learns to obtain semi-ideal CSI features that incorporate room-specific "accents",i.e.long-term environmental influences typically caused by room layout.A second network is trained for similarity evaluation between semi-ideal features and measured features to mitigate short-term environmental effects such as channel variations or noise.In this paper,the prototype implementation of the path and gait recognition system is deployed on a commercial WiFi device,and the performance of path recognition and gait recognition is evaluated through experiments.For the path recognition problem,compared with directly using the walking trajectory for classification,our system can improve the accuracy percentage by a maximum of 21.0%and an average of 13.3%.When the distance between the center points of two paths exceeds 2 m,the system can accurately and efficiently identify different paths,and its accuracy rate will exceed 90%.And in the multi-path gait recognition system deployed in reality,the average path recognition accuracy can be as high as 86.9%.For the gait recognition problem,performance evaluation results show that the recognition accuracy of WiVi ranges from 85.4%for a 6-person group to 98.0%for a 3-person group.Compared to single-path gait recognition systems,we achieve an average 113.8%performance improvement.Compared with other multi-path gait recognition systems,we achieve similar or even better gait recognition performance while requiring 57.1-93.7%fewer samples. |
