Gait Recognition Robust To Covariate Factors

Gait recognition has been paid increasing attention because it’s non-offensive, easy to collect, difficult to disguise and the whole system can be operated from a distance. It’s widely seen that gait recognition module can act as a part of a multimodal biological system. However, gait recognition would be greatly affected by some covariate factors including clothing type and carrying objects. Finding an approach robust to these covariate factors is the most challenging problem. In this paper, some related research and work have been done on pre-process of gait sequence, moving object detection and segmentation, morphological processing, feature extracting, feature selection and classifiers.In this paper, a method based on canonical correlation analysis(CCA) was proposed to model the correlation between gait sequences on normal condition and condition with covariate factors. GEIs are partitioned into several parts based on local information entropy value, with each part selected as a sub-pattern. Then canonical correlation analysis(CCA) is applied to each part of GEIs on two different walking conditions. Finally, to reduce the effects of covariate factors, overall correlation strength is used for classifying, which is the weighted sum of all the partial correlation strengths.To avoid judging existence of covariate factors which must be done in above method, a new method based on collaborative representing classifier(CRC) was proposed. To eliminate the negative effects of covariate factors during the process of refactoring, an expanded dictionary was constructed, adding some images of covariate factors into the original one.Experiment results on CASIA-B gait database show that the proposed methods outperform other classical methods over all views. Finally, a gait system was realized on Windows platform, with the implementation of OpenCV, taking Visual Studio 2012 as development platform. It has been validated that the system performs well in the aspect of correct recognition rate and real-time capability, but still need to be promoted.