Research On Gait Recognition Method Based On Pedestrian Appearance Information And 3D Pose Key Points

Gait recognition is an emerging biometric recognition technology.Compared with other traditional biological features,such as human face,fingerprint,iris,etc.,gait has received extensive attention from researchers in the field of biometric recognition due to its easy collection,difficult to camouflage,and non-contact characteristics.Gait recognition is one of the best long-distance identification technologies.With a high-definition camera,its recognition distance can reach up to 50 meters.Although gait recognition has many advantages,the research on this technology is still in its infancy,and there are some unsolved problems that hinder the development of gait recognition.These problems include different conditions when walking,such as whether you are wearing a jacket,the ground conditions are good or bad,whether you are backpacking,and many other factors.These frequent changes in the real world affect the accuracy of gait recognition.Aiming at this kind of problem,this paper is proposed to combine the pose estimation algorithm with gait recognition technology.The pose estimation algorithm is a classic algorithm in the field of computer vision.Its task goal is to obtain the key points of the human body in the picture.This paper uses the key points of the human body to re-model the gait features and construct a 3D spatial-temporal map of the key points of the human body.It is proposed to use the graph neural network to process the gait features.This method uses the temporal information in gait sequence to reduce the influence of covariates such as appearance.It is proposed to use graph attention network to replace the attention mechanism in the original network to strengthen the role of key nodes.The experimental results verify the effectiveness of the new attention network.Three neighbor node grouping strategies are used to redistribute the weight of neighbor nodes.The experimental results show that different strategies will cause different effects of neighbor nodes related to gait characteristics on the original node.The experimental results of gait recognition based on graph neural network obtained by using two improved strategies are compared with other gait recognition methods using 3D key points.The advanced nature of this method is obtained.Regarding the problem that key points of the posture are insufficient to characterize the contour features,this paper proposes to use gait contour sequence to construct frame-difference energy map,using three frame-difference energy map construction methods.Three fusion methods are used to fuse it with the original contour map,and each frame in the obtained feature sequence has some features of the remaining frames.In this way,the recognition method that takes the gait contour as the input not only pays attention to the appearance features,but also pays attention to the walking features contained in the gait image sequence.The advantages and disadvantages of three frame difference energy map construction methods and three fusion methods are verified by experiments.Three types of attention modules are used to improve the model,respectively activating the spatiotemporal attention,channel attention,and framedifference energy attention in the original features to strengthen the network’s ability to perceive important features.Experiments are also used to obtain the effectiveness of the attention module.Through comparative experiments,the experimental results of this method have reached the level of the most advanced methods.The above two methods are gait recognition methods based on image sequences,so a model based on the gait energy map method is designed and experiments are performed to compare the experimental results of two types of different inputs.In order to explore the performance of multi-source information fusion gait recognition algorithm,the three models are divided into two groups for experiments,and the complementarity between different models and characteristics is obtained.