Research On Pedestrian Attribute Recognition Method Based On Gait Analysis

With the development of society,video surveillance has achieved significant development.Nowadays,important public areas,such as railway stations,airports,schools,and shopping malls,are all within the scope of video surveillance,ensuring the harmonious development of society.People are the main body of society,and most of the events happening in society are closely related to people.Therefore,image recognition technology that automatically judges pedestrian biological attributes(gender,age,etc.)from massive videos has made great progress,but previous technologies used Face image recognition,when the face image is occluded and the resolution is low,it will seriously affect the recognition work.Gait is a pedestrian's walking posture and is a behavioral feature.The pedestrian's gait can be represented by the outline of the pedestrian.This indicates that the gait has lower requirements for image resolution and can obtain the pedestrian's gait at a long distance.High concealment.Therefore,research on pedestrian attribute recognition methods based on gait analysis has attracted wide attention.This article first studies the research background and significance of pedestrian attributes and the current status at home and abroad,and analyzes the current research status of gait recognition.Secondly,it details the related aspects of obtaining pedestrian binary contour maps,determining the gait cycle,and generating gait energy maps.Method and detailed research on related theoretical knowledge in deep convolutional neural network.Based on this,the following research contents have been completed:(1)A gender recognition method based on deep features is studied.Aiming at the perspective,the impact of pedestrian walking status on recognition,first extract the HOG features from the gait energy map to complete the recognition.Experiments show that although the HOG features are robust to the perspective,they cannot effectively describe backpacks and coats on pedestrian gait.influences.Therefore,a residual expansion module was proposed.This module was used to build a residual expansion network.The network was used to extract depth features and combined with support vector machines to complete classification.Dataset B of the CASIA gait database was used for experimental evaluation,indicating the depth of extraction.Compared with other algorithms,feature pairs have a high degree of robustness to viewing angles and pedestrian walking states.Compared with other algorithms,it proves that depth features can describe the differences between men and women's walking attitudes.(2)A multi-class dynamic cross-entropy loss function is designed.In real life,the data collection of young adults is relatively easy compared with the minors,and it is relatively easy for the elderly.This results in an extremely uneven distribution of the number of age samples,which is not conducive to the learning of neural networks.Aiming at this problem,dynamic weights are used to constrain the neural network to predict the recognition results as multi-sample classes.The OULP-Age dataset is used for performance verification.Experiments show that dynamic weights impose greater constraints on neural networks than fixed weights.Combining the two algorithms,the work of this paper realizes the recognition of the gender and age of pedestrians.Evaluation experiments on the data set prove the effectiveness of the two algorithms.