Research On Key Technologies Of Video-based Human Face And Action Recognition

This paper expounds the research background and significance of video face recognition and behavior recognition.On the basis of relevant theoretical study and research,the current mainstream video face recognition and behavior recognition methods are deeply studied,and some algorithms of video face recognition and behavior recognition are proposed.It mainly includes video face recognition algorithm based on QPSO optimization of Riemann manifold,video human behavior recognition algorithm based on multi core 3D-CNN and video face recognition algorithm based on local spatiotemporal continuity polymerization.The main work of this paper includes the following three aspects:1.The unconstrained video scene is complex and changeable,the amount of data is large,the equipment is not consistent and the environment is uncontrollable.The video is flooded with a large number of occlusion and face rotation,which results in the low accuracy of face recognition,unstable performance and large computational complexity.In order to solve these problems,this paper proposes a face recognition algorithm based on QPSO optimization and Riemann manifold.The algorithm regards video face recognition as the similarity measure of image sets.After image alignment,texture features are extracted and fused,and then the intrinsic representation of video faces is obtained by using a large and simple dimension of Riemann manifold with QPSO optimization.The similarity measure uses the convex packet distance calibration method,and finally uses SVM.The best classification results were obtained.By comparing with several current mainstream algorithms on Youtube Face database,the validity of the algorithm is verified.The proposed algorithm has high recognition accuracy,low error and strong robustness.2.Considering the complexity of preprocessing of traditional human action recognition methods based on feature extraction,the accuracy of human face recognition is not high enough.In this paper,a video human behavior recognition algorithm based on multi core 3DCNN is proposed.The algorithm extracts the gray features,the gradient features and the characteristics of the optical flow respectively.At the same time,the convolution operation is carried out with different convolution kernel for the three feature sequences to obtain the multi angle features of the image sequence,and then the CNN network is constructed for video.Human behavior recognition.By comparing with several current mainstream algorithms on KTH database,the effectiveness of the proposed algorithm is verified.The proposed algorithm has high recognition accuracy,low error and strong robustness.3.How to use the temporal and spatial continuity information in video to design an effective video texture description operator is an important way to realize video face recognition and video analysis and understanding.In this paper,a video face recognition algorithm based on local spatial-temporal continuum aggregation description is proposed.The algorithm uses the image set to represent the video,divides the image sets into blocks,extracts the local temporal texture features of the video by the LBP-TOP operator,and uses the k-means algorithm to obtain the clustering center of the video description set.The resulting portrayal of the local information of the video.The similarity between feature vectors is defined by Euclidean distance,and the corresponding weights are assigned to different components.Finally,the nearest neighbor method is used to get the best classification result.The recognition performance of the proposed algorithm is verified by comparing experiments with several mainstream video descriptors in Honda/UCSD database.In this paper,we summarize the current biometric identification methods,and study the problems of video face recognition and video human action recognition.On the basis of related research,we propose several video face recognition algorithms and human behavior recognition algorithms,and make progress on the existing problems in this field.Finally,the future development of video analysis and processing is prospected.