Gait Recognition Based On The Fusion Of Contour And Attitude Features

Gait recognition is an emerging biometric recognition technology that has attracted attention in the field of computer vision and biometric identification in recent years.It is designed to authenticate or identify people based on their walking postures.Because each person's physiological condition(including bone density,center of gravity,etc.),visual acuity and coordination ability are different,gait recognition has individual uniqueness,which is the theoretical basis for gait recognition.Compared with traditional biometrics(such as fingerprints,palm prints,etc.),gait recognition has the advantages of long distance,non-contact,not easy to hide,difficult to imitate and camouflage.As a method in the field of pattern recognition,gait recognition has extremely important practical value and theoretical research significance.Although the topic of gait recognition has made great progress in various aspects since its introduction,the technology is still in the preliminary research stage,and there are still many difficulties to be solved,such as walking roads,different perspectives,different costumes and different carriers make the recognition rate in the complex real environment combined with such influencing factors still low.Aiming at the difficulties in the practical application of gait recognition technology,this paper proposes a gait recognition method based on the fusion of contour and attitude features.Trying to solve the influence of covariate changes on gait recognition performance,making it more robust to self-occlusion,viewing angle,clothing and carrier changes.Firstly,the gait energy image with better robust performance is selected as the contour feature.In order to solve the problem that the feedforward neural network structure leads to insufficient expression characteristics,three kinds of network frameworks based on multi-resolution fusion are proposed innovatively.Combining high spatial resolution and rich semantic features to comprehensively improve the predictive ability of the model,and through comparison experiments in performance,the best performing network which through global average pooling to achieve multi-resolution fusion is used in subsequent fusion experiments.This lays a solid framework for feature fusion experiments.This paper also improves the performance of skeletal key points' prediction network in the aspect of attitude feature extraction.By combining the human analysis network PSRNet to create the skeletal key points' relocation system,it tries to restrain the change of attitude position by segmenting each body part.By this step,more accurate and robust attitude features are extracted for feature fusion experiments.The experiment proves that the skeletal key points' relocation system proposed in this paper can correct the position of key points with large deviation and replenish the joints that are missed.Through comparison experiments,the performance of Pose Relocation proposed in this paper is better than the independent skeletal key points' relocation model Pose Fix and the multi-stage end-to-end skeletal key points' relocation model CPN which are both leading in the field.Finally,the feature fusion and decision fusion of the contour feature and the attitude feature extracted from the gait sequence are studied.Through a series of contrast experiments,it is found that the fusion of the gait energy image which includes the dynamic and static information of the gait sequence and the spatiotemporal characteristics of skeletal key points which can weaken the effect of the carrier or costume changes could improve the accuracy of gait recognition,which confirms the rationality and effectiveness of the subject.