The Application of Improved GG Clustering Algorithm in View-irrelevant Behavior Recognition

Abstract

When cluster descriptors of behavior feature in the analyzing the behavior feature data of behavior under different view, the traditional FCM algorithm can not determine the number of clusters to the data with spherical structure, so this paper proposes an improved GG clustering algorithm to solve this problem. This algorithm determine the optimal cluster number by the indexes of inter-cluster compactness and the separation of clusters. Then model behavioral descriptors that have been clustered to reach the purpose of improving behavior recognition accuracy. The experimental results show that: the improved algorithm can classify and model behavioral descriptors better and improve the recognition accuracy. Introduction There are many different methods about behavior recognition and understanding in recent years[1-3],in the method of extracting behavior features, establishing behavior feature descriptor and analyzing the descriptors with clustering method to build behavior recognition model to complete the behavior recognitionthe effect of clustering and modeling has an important influence on behavior recognition. Over the years many researchers has study deeply with the theoretical basis of validity index of clustering, study and improve the basic requirements the validity index of clustering should satisfy, and proposed a set of basic axiom that the validity index of clustering must satisfy. In 1965, the founder of fuzzy set theory Zadeh proposed a validity function of clustering: Separation function, but the judgment of fuzzy clustering validity is not very ideal. In 1974, Bezdek put forward the concept of partition coefficient, it constitutes the first practical clustering validity index PC[4], and then the concept of partition entropy is proposed PE[5]. In 1987, Davies and Bouldin[6] proposed separability measure based on Fisher distance between one cluster and another. In1991, Xie et al.[7] used objective function of fuzzy clustering, along with two important factors separation and compactness, proposed Xie-Beni index, but the evaluation standard does not consider the structure of data set. Kim et al.[8] proposed validity Kim based on overlapping degree among clusters, but the same as partition coefficient and partition entropy, it will monotonously change with the increase of the cluster number. In 1998, Rezaee proposed using linear combination to zoom compactness and separation by scale factor, thus make up for the defects of the difference measurement to a certain degree[9]. Although the index has a great improvement on overall performance, the structure is very complex, and always gives out the result that opposite the facts. Considering that XB index does not consider the structure of data set and Kim index monotonously changes with the increase of the cluster number, this article uses the sum of weighted square errors within the cluster to measure the compactness, and uses the couples of fuzzy clustering to measure the separation of clusters. So defines the validity index(CS) based on these two metrics, effectively overcoming the shortcoming of the traditional FCM clustering algorithms on determining the initial parameter. The experimental results show that the improved algorithm achieves a more stable and effective clustering result. International Conference on Advances in Mechanical Engineering and Industrial Informatics (AMEII 2015) © 2015. The authors Published by Atlantis Press 22 Traditional GG Clustering Algorithm Gath-Geva(GG) algorithm is an improvement of FCM algorithm. Fuzzy Cmeans clustering algorithm can only reflect the standard distance norms of super spherical data structure, so the FCM algorithm is only suitable for data structure with the same shape and direction. Thus GG clustering algorithm uses distance measure based on fuzzy maximum likelihood estimation, and can detect and adapt to sample data with different shape, size, density, at the same time, it makes the clustering no longer be limited by the sample data distribution volume and can improve the accuracy of clustering. Suppose } , , , { 2 1 n x x x X  = is a data set with n metadata, i x is sample data with p dimension, fuzzy clustering divide the sample set X into c clusters (suppose 2 , , , , j c c c c   ) according to fuzzy partition matrix ij U u   =   , [ ] 0,1 ij u ∈ represents the degree that the sample i x belonging to class j , and the sum of membership degree of sample i x belonging to all classes is 1, according to minimum distance quadratic sum that the sample point to the cluster center, define the objective function: