Real-Time Person Identification Combined With Multiple Features

Recognizing the identification of a person is an important content in the field of computer information security,and the advantages of gait-based identification,such as difficult to disguise,distance and non-contact,which are different from other biometric identification technologies attracted more and more researchers to explore.The main task of gait-based identity recognition is to fully mine the serialized gait data by using deep learning,computer vision and other theories and technologies,and on this basis,carry out identity recognition for the people in the video.This thesis is to identify the identity of the individual characters in the video as the research target.First of all,design a target segmentation algorithm which introduces human skeleton feature to separate the outline of the person,and then design a feature extraction algorithm which introduces the human feature to extract contour features and the local characteristics of the human body each place,finally design a video character identification algorithm based on feature fusion of video character identification.The work of this thesis is mainly described in the following three aspects:(1)Design a target figure segmentation algorithm based on human skeleton features.In the process of video character segmentation,combined with a bottom-up pose-to-pose estimation method,first get pose-to-pose key points,then clustering divides the human body posture template,then for video character alignment operation to deal with different situations character position,finally introduced the human body skeleton characteristics for target segmentation.Experiments show that designed in this thesis based on the character posture alignment operation effect is better when processing condition,through this action helps to form a distinctive ROI,in order to better distinguish between two highly overlapping characters,the human skeleton feature is introduced.The experiment shows that the target segmentation algorithm that introduces the human skeleton feature is more accurate,and the segmented body contour is clearer.(2)Based on the human body proportion,designing a feature extraction algorithm to obtain the gait energy map through the contour features of people,and improve the gait energy map based on the proportion relationship of human arms and legs in the whole human body.The experimental results show that the improved gait energy map has better feature expression,which improves the gait recognition rate.The character contour feature extraction based on the convolutional neural network takes into account that each person has different feature expression when walking,and extracts the common features of the character contour and the salient features of each person,which are combined to extract the character feature.Experiment results show that the contour feature extraction algorithm based on convolutional neural network can effectively characterize human contour information.(3)Due to the certain limitations of single feature,the feature of identifying the extracted features makes it more sensitive to changes in a certain feature,but insensitive to others.According to this,a fusion method is designed to realize the identification of a person,the extracted spatial-temporal features are fused with the acquired multi-features of the characters.Experiments are carried out on the CASIA-B dataset.The experiments show that the character identification based on fusion features designed in this chapter can effectively integrate spatial-temporal information and achieve higher accuracy rate.