| Gait recognition technology has attracted much attention due to its ability to perform individual recognition at a distance without the need for subject cooperation.Compared with traditional biometrics such as fingerprints,faces,and irises,gait features contain not only spatial but also temporal information,thus having stronger information carrying capacity.However,the interweaving of spatial and temporal information also increases the difficulty of extracting gait features,which means that the full exploitation of spatiotemporal information in gait sequences is one of the key issues in the field of gait recognition.Existing gait recognition research has made certain achievements in the utilization of spatiotemporal information.However,how to cope with the spatial information changes caused by angle and appearance changes,and achieve efficient spatiotemporal feature extraction,is still a challenging problem.Therefore,this paper explores efficient utilization strategies for spatiotemporal information in gait sequences from the perspectives of temporal and spatiotemporal differences in gait sequences.Firstly,in response to the difficulty of gait recognition caused by spatial information changes,a cross-view gait recognition method based on regional dynamic differences is proposed.By analyzing the temporal differences in different regions,this paper finds that the temporal patterns of gait sequences are closely related to their spatial distribution.Therefore,a non-dynamic region suppression algorithm and a spatiotemporal feature extraction strategy based on regional temporal differences are proposed.The former reduces the impact of low information density regions in input frames on the model,while the latter extracts local spatiotemporal features based on the human body’s dynamic pattern to obtain finer local features.This method achieved an average recognition rate of 93.2%on the CASIA-B dataset,especially showing excellent robustness under cross-angle conditions.Secondly,in order to optimize the computational efficiency of spatiotemporal feature extraction for gait,a lightweight and efficient gait recognition method based on spatiotemporal decoupling network is proposed,which achieves parallel extraction of gait spatial and temporal features with very little computation.Moreover,by suppressing noise generated by spatiotemporal feature aggregation using a noise suppression function,this method can fully fit local temporal patterns of gait sequences through spatially shared and temporally independent feature extraction strategies.Therefore,while the model is lightweight,it can further achieve significant performance improvements.Specifically,this method achieved an average recognition rate of 93.7% and 90.5% on the CASIA-B and OU-MVLP datasets,respectively.Finally,to explore the application mode of gait recognition algorithms in practical scenarios,this paper designs and implements a desktop gait recognition system for pedestrian monitoring,rapid gait contour data collection,registration,and individual recognition.Testing proves that this system can effectively recognize the identity information of pedestrians in different backgrounds without restricting clothing and walking speed. |
