Action Recognition Algorithm Based On The Improvement Of Extreme Learning Machine Or Multilevel LBP Algorithm

Machine learning's applications are becoming more and more hot in real life,and machine learning based on big data is more and more popular.In the theoretical study of machine learning,the research of classifier occupies an incomparable position,and most of the research problems can be transformed into classification problem.The performance of classifier is often the key to the success of the results of the study.Now human behavior recognition used for detection and identification mainly through uncomplicated video surveillance.How to get the computer to be able to achieve human behavior,learning and recognition.what characteristics to express the behavior of human is still a research focus.Behavior recognition is also a wide range of knowledge related to image processing,computer vision,machine learning and pattern recognition.With the advent of the high-solution camera,the key technology is a moving target feature extraction and machine learning algorithms.And most studies mainly stay in a invariant scene,for the study of human behavior recognition is a very important social value.In this paper,under the circumstances of the disadvantages of the traditional extreme learning machine,using several relatively better improved algorithm optimize the performance of ELM.The main contents of the paper are as follows:(1)Based on wavelet multi-level LBP algorithm,the behavior of image extraction is more significant.This paper proposed an algorithm of extracting feature based on Gabor wavelet and LBP algorithm for the efficient extraction of behavioral characteristics and the ability to effectively classify different human behaviors in behavioral images.Because features of Gabor have strong robustness to the change of action behaviors and environmental conditions.We firstly use the multi-scale local features of the samples to obtain wavelet of multi-level decomposition and filtering analysis.And then define a number of different sizes of the filter,filtering and calculating the gradient vector.Finally,the uniform pattern of LBP algorithm is used to get the vector of the sequence different image to merge the sequence image features,and then obtain the eigenvector of the behavioral image.(2)Based on Cholesky decomposition of the extreme learning machine algorithm,the accurate classification of behavioral images have an ideal experimental effect.The traditional neural network of BP gradient algorithm needs to set a large number of training parameters,and easy to fall into the local optimal,training time is too long,excessively fitting and so on.ELM only needs to set the parameters of middle layer nodes without manually setting a large number of training parameters and it can produce the optimal value of the solution.And the algorithm is easy to implement,less time of consuming and generalization performance good characteristic,so action recognition algorithm has been put forward.Firstly,a method based on Cholesky decomposition to seek the calculation of ELM is introduced into the algorithm.Secondly,according to the characteristics of kernel function updates during Extreme Learning Machine(ELM)in online learning,partitioned Cholesky decomposition algorithm is utilized for online solution of ELM,which results in on-line updating for the triangular factor matrix.Finally,we can obtain a new online learning(ELM-Cholesky)algorithm,making the recognition efficiency has been improved.