Gender recognition is a key issue in biometric identification.It has broad application prospects in many fields such as military,public security,intelligent monitoring,big data analysis,and pedestrian recognition.Compared with the traditional gender recognition method based on facial features,the acquisition of gait features is more concealed,the target is not easy to hide,and the require of resolution of video images is low,which is very suitable for long-distance,uncontrolled open occasions.However,related research on gait recognition is still in infancy.Especially in real scenes,gait features are very susceptible to external factors such as wearing,carrying,angles and distances.Therefore,how to remove the influence of these disturbances has been the focus of research.In recent years,the rapid development of deep learning in the field of image recognition brings new technical support to the research of gait recognition based on video images.It provides a new solution to overcome the interference of external factors.This paper obtains the gait characteristics based on the pedestrian gait sequence.The paper focuses on gait feature image calculation,gait feature extraction,machine learning and other related content.The main work and innovation are as follows:(1)An improved gait energy map is proposed to eliminate the interference of wearing and carrying objects on pedestrian gait characteristics.Traditional gait feature images are usually processed for a single pedestrian image.Although they can express gait features well,they are very susceptible to external interference and lack the "common" feature of pedestrian gait of the same gender,suitable for pedestrian identity.Identification is not suitable for gender identification.Aiming at this problem,this paper uses Shannon entropy to calculate the irrelevance of each region in the universal gait template which can express the "commonness" feature of pedestrians of the same gender,and obtain the "common" regional mask version.Finally,it is not easy to extract the gait energy map.Feature areas that are subject to outside interference are used for gender identification.(2)Based on the requires of gender classification,the MyVGGNet model was designed based on VGGNet-16.The training of convolutional neural network is essentially the optimization of feature extraction ability.The richer the database used for training,the more perfect the network training,and the extracted image features are more "useful".This paper retains the parameters of the VGGNet model convolutional layer trained on ImageNet,and fine-tunes it with a new fully-connected layer to adjust the gender recognition task on the already perfect image features to improve the recognition accuracy.(3)The fusion of convolutional neural network and support vector machine.Aiming at the advantages of linear SVM classification in small samples and CNN's ability to extract deep abstract features of images,we combine the two methods under the conditions of feedback and no feedback to solve the gender recognition problem. |