Research On Pedestrian Identification Technology Based On Gait

As a new generation of biometric identification technology,gait recognition has unique advantages in long-distance pedestrian identification.It has a very broad application prospects and practical needs in the fields of military,criminal investigation,public security and smart home.However,gait recognition is easily affected by internal and external factors such as clothing,carry-on,change of perspective and illumination,which seriously hinders its commercialization process.This paper focuses on image preprocessing,gait feature image extraction,general machine learning algorithms and deep learning algorithms to solve the problem of gait pedestrian identification under complex covariate conditions.The main work and innovations of this paper are as follows:(1)This paper studies the image normalization method adopted by the traditional gait energy image,and proposes a normalization method based on the center of gravity alignment in the static region of the pedestrian contour image.The image of the improved gait energy image is clearer,and it can extract more abundant and discriminative static contour features.(2)The principle and calculation method of gait entropy image are studied.Shannon entropy is introduced to enhance the differentiation of dynamic and static regions in gait energy image.A general mask is designed,then a new type of gait energy image is obtained by "and" operation of part of the general mask and the improved gait energy image,which can effectively reduce the influence of pedestrian wearing and carrying changes on gait feature extraction,reduce intra-class differences,and thus improve the correct recognition rate of pedestrian identity.(3)The traditional machine learning algorithm is studied.By further extracting the hand features such as HOG and LBP from the proposed two gait feature images,and inputting them to the nearest neighbor classifier for distance measurement and identity matching,an ideal correct recognition rate can be achieved.(4)The composition structure,characteristics and back propagation algorithm of convolutional neural networks are studied.With reference to the existing excellent convolutional neural network model structure,a deep convolutional neural network model suitable for pedestrian identification tasks is designed and built.The high-order gait features extracted automatically are more efficient than the artificially extracted features.At the same time,it is concluded that the small sample training depth convolutional neural network model is prone to over-fitting.(5)Aiming at the over-fitting problem,the migration learning algorithm is studied.By fine tuning the pre trained Inception-v3 model,the CASIA-B gait database is used to train and optimize the model parameters,and a new model for pedestrian identification tasks is obtained.The model is used as an efficient feature extractor to extract the bottleneck features of gait feature image,inputting them to the nearest neighbor classifier for distance measurement and identity matching,and the recognition rate is further improved.