Research On Gait Recognition Method Based On Vision Transformer

In recent years,the country and the world have endured the adverse effects of the new coronary virus,and the development of society and economy and the integration of culture have been hit.At the same time,this virus’ s strong spread also provides more thoughts for researchers of biological characteristic identification.Today’s biological characteristics identification technology mainly studies the individual characteristics of the facial features,fingerprints,iris,and walking gestures,which are not easy to change and camouflage and achieve identity recognition effects through the differences of biological characteristics.Among them,facial,fingerprints and iris recognition are most widely used,but these methods not only have high requirements for the light source positions and light strength of the collection environment,but also need to cooperate with the target to recognize the target.Effective identification can be performed in a near range.The gait recognition is different.With a long recognition distance and the characteristics of no contact,it provides a new idea for biological characteristic identification research,which has received widespread attention from researchers.However,the two advantages of the gait recognition also brought two problems for research gait recognition:1.Because gait recognition is a medium and long distance biological characteristic recognition technology,the environmental impact of gait recognition will be more complicated.Such as the impact of weather.2.In order to maintain the advantages of no sense of recognition,it is necessary to overcome the issue of the randomness of the target movement of recognition.Because current gait recognition usually needs to use a fixed location video collection device for data collection,Therefore Refined into a multi-view gait recognition problem.The main focus of work in this article is based on the two gait recognition research problems.The specific content is as follows:First of all,in response to the effects of the extreme environment brought about by long-distance recognition,this article uses a single-angle complex scene step-by-step recognition as a data experimental background,and uses Vision Transformer(Vi T)as a technical basis.The main ideas are: using infrared gait data to overcome the adverse effects of insufficient light sources and extreme weather on gait data collection.Considering the timeliness and part of body characteristics of gait,combine the characteristics of the sequence of the walking posture synchronization and the part of body characteristics.And cut the blocks of different body outlines to build of the independent characteristics of the segmentation block of each gait contour;In order to solve problems such as easy saturation accuracy and low-to-gait fitting efficiency in the experimental process.Put forward the effects of building dual attention modules to enhance the effects of the attention module in the fitting gait characteristics,and use the alien migration learning method to further improve the characteristic bidding efficiency.The model was applied to multiple simulation experiments in the CASIA C infrared body gait library of the Chinese Academy of Sciences.The results show that the models of this article are better than the traditional Vi T model and the CNN comparison model in terms of stability,data fitting speed,and accuracy of recognition.Secondly,because the current gait recognition and collection equipment requires a fixed position to collect data,and it is random to walk direction to identify the target.Therefore,the problem of gait recognition from multiple perspectives is proposed.This article has found that the traditional one-angle gait identification research is often used90° as the preferred angle,because other angles will often cause the human contour characteristics coincide and low accuracy.In order to improve the accuracy of multi-perspective gait recognition,this article proposes a method of gait feature extraction method based on Siamese neural network and Vision Transformer(Vi T).This method uses Siamese neural network relationship analysis,calculates the gait characteristic relationship between 90° and other perspectives,and generates the perspective characteristic relationship factors.The advantage of this method is that it does not generate universal gait feature diagrams,nor does it have to be converted through a complex perspective conversion matrix.The recognition accuracy can be significantly improved by only computing the gait feature relationship between viewpoints for highdimensional gait feature tensors.In addition,the Siamese Mobile Vision Transformer(SMVi T)constructed in this paper not only focuses on the local spatial features of human gait,but also takes into account the long-distance attention posture correlation features.It can extract multi-dimensional gait identity features,ensure that the gait feature tensor obtained can better reflect human recognition information,and is conducive to obtaining reliable perspective feature relationship factors.Finally,a gradual perspective training method is proposed to further improve the fitting efficiency of gait features through the flexible transition of adjacent perspectives.The experimental results on CASIA B dataset of the Chinese Academy of Sciences show that SMVi T can achieve the most advanced performance compared with many advanced gait feature recognition models from multiple perspectives.To sum up,through the research and experiment on two key problems derived from the advantages of current gait recognition,this paper,based on the two models proposed by Vision Transformer,has solved these two problems to a certain extent,and provided new ideas and methods for the application of Vi T in gait recognition.