Research On Action Recognition Method Based On Complex Linear Dynamical Systems

At present,human behavior recognition plays an important role in all areas of society,such as sports video analysis,security monitoring,human-computer interaction,and virtual reality,which promote the development of various industries.Therefore,research on human behavior recognition technology has set off a boom in scientific research.Although a variety of behavior recognition algorithms have been proposed at present,due to the complexity of the database and human motion,there are still many areas where human behavior recognition technology needs improvement and research.In the study of human behavior recognition,timing information is often ignored.This paper focuses on the issue of loss of timing information in human behavior recognition and studies human behavior recognition.The main contents of this paper include:1.In order to extract the feature that can both represent human motion and be suitable for time-series modeling,we performed three commonly used features: histogram of oriented gradient,histograms of oriented optical flow,and motion boundary histogram.To find a better representation of features.Through experiments,we found that the motion boundary histogram is more suitable for the modeling of temporal information in human behavior recognition.2.Time-series data generated by the extraction and representation of video features usually has partial dimensions that are irrelevant or redundant.If the original sequence is processed directly,the performance of data processing algorithms will be greatly reduced,and an effective algorithm needs to be sought.To eliminate irrelevant dimensions.However,there are certain links between the dimensions of the multidimensional time series,and we must consider the dimensions of each dimension as a whole.In order to solve this problem,we use Locally Linear Embedding algorithm,Laplacian Eigenmaps algorithm and Locality Preserving Projections algorithm in time series miningto test and analyze these time series data.According to the experimental results,the Locality Preserving Projections algorithm is selected.3.In the human behavior recognition,there is usually a problem of missing timing information.Although the traditional linear dynamic system can capture timing information,the transfer matrix and the output matrix are limited by the arrangement,rotation and linear combination,so that each row in the output matrix is It is not possible to uniquely identify the characteristics of the corresponding system.In this paper,time series modeling using complex linear dynamic systems can not only extract timing information,but also solve the shortcomings of traditional linear dynamic systems.Experiments verify that the complex linear dynamic system is more suitable for time series modeling and improves the classification accuracy.