Research On Image Recognition Algorithm Of Barefoot Pressure Under Natural Walking And Loading Conditions

With the rapid development of information technology,informatization,digitalization,mobility and remoteness have brought convenience to people’s lives,while at the same time,problems such as personal information security have emerged,which makes the application of biometric technology in criminal investigation,security,finance and other public areas more and more important.In criminal investigation cases,the means of criminal suspects are becoming more intelligent and complicated,and the useful traces left at the scene of the crime are becoming less and less,but footprints are important traces that people can not avoid leaving in their activities.As the identification of different individuals,footprints have their own inherent biometric attributes,and footprint-based biometric recognition technology plays an important role in criminal investigation and other fields.In real life scenes,most footprints are produced when people are walking,which is more disturbed by more factors and more characteristic changes than footprints when people are standing still.Footprint recognition in the state of walking is faced with the influence of many interference factors,such as the scene environment,the bearing body,the psychological state and camouflage of people,different walking speeds,different loading states and the change of human posture.Therefore,this thesis takes the plantar pressure image as the object,combined with deep learning technology,to study the footprint recognition algorithm under different walking speeds and different load conditions.The main work is as follows:(1)Barefoot plantar pressure image data sets in different walking states were collected and constructed.In view of the lack of open data sets for footprint research,this thesis uses the existing footprint perception and analysis laboratory to formulate a unified data acquisition process and construct a barefoot plantar pressure image data set suitable for footprint recognition research.In the second stage,barefoot pressure images of 40 people walking under different weight-bearing conditions were re-acquired.The plantar pressure images were cropped into single barefoot pressure image,and then the denoising,centralization,image enhancement and other preprocessing operations were carried out to improve the quality of footprint data,and finally 9176 barefoot footprint pressure data under different walking conditions were constructed.(2)A footprint recognition algorithm based on integrated multi-convolutional neural network is proposed.Because the footprint is easily affected by the walking speed,it shows a certain degree of variability,which makes the footprint within the class difference,and when using a single conventional convolutional neural network for recognition,the variance is large and the accuracy is not high,which brings difficulties to footprint recognition.Therefore,based on the ensemble learning method,this thesis uses Inception XNet,Res Net and Dense Net as the basic networks to extract and learn barefoot pressure features and predict them,and finally uses the weighted voting ensemble method to optimize the prediction results of the three basic networks to improve the final prediction results.Because each network model has its own characteristics,the prediction error of each model is different,so by training multiple models and combining the prediction results together,the variance can be reduced.The experimental results show that the proposed algorithm has good robustness in barefoot pressure image recognition at different walking speeds,and the recognition accuracy is higher than that of other traditional single networks,which effectively solves the problem of low accuracy of footprint recognition caused by the instability of traditional footprint features in natural walking state.(3)A footprint recognition algorithm based on selective fusion attention network is proposed.A selective fusion attention network model is proposed to solve the problem of the change characteristics of footprint features under different weight-bearing states and different body weights and the recognition accuracy of the integrated model under weight-bearing walking conditions.According to the model,firstly,an Inception XNet single network with the maximum distributed weight in an integrated model is used as a bottom-up path to extract multi-scale footprint features of different levels in a barefoot pressure image,an up-sampling operation is adopted in a top-down path to generate the multi-scale footprint features of different levels into features with higher resolution,and the features are fused.In the process of feature fusion,the selective fusion attention is used to fuse the stable and invariant footprint features with high discrimination in the deep and shallow feature information,and finally the footprint recognition is carried out through the stable and invariant plantar pressure distribution features after preferential fusion.Several groups of tests are carried out in three barefoot pressure image data under the conditions of large weight change,double shoulder load and single portable load.The experiments show that the accuracy of this algorithm reaches 94% on the data sets of different load States,and more than 93% on the data sets of large weight change.On the premise of ensuring the recognition speed,It can effectively identify footprints of different load States and different weights in natural walking,has high recognition accuracy and robustness,and can provide more extensive theoretical basis and technical support for later on-site footprint comparison and identification.