A new human action recognition approach was presented based on chaotic invariants and relevance vector machines(RVM).The trajectories of reference joints estimated by skeleton graph matching were adopted for represent...A new human action recognition approach was presented based on chaotic invariants and relevance vector machines(RVM).The trajectories of reference joints estimated by skeleton graph matching were adopted for representing the nonlinear dynamical system of human action.The C-C method was used for estimating delay time and embedding dimension of a phase space which was reconstructed by each trajectory.Then,some chaotic invariants representing action can be captured in the reconstructed phase space.Finally,RVM was used to recognize action.Experiments were performed on the KTH,Weizmann and Ballet human action datasets to test and evaluate the proposed method.The experiment results show that the average recognition accuracy is over91.2%,which validates its effectiveness.展开更多
A variation pixels identification method was proposed aiming at depressing the effect of variation pixels, which dilates the theoretical hyperspectral data simplex and misguides volume evaluation of the simplex. With ...A variation pixels identification method was proposed aiming at depressing the effect of variation pixels, which dilates the theoretical hyperspectral data simplex and misguides volume evaluation of the simplex. With integration of both spatial and spectral information, this method quantitatively defines a variation index for every pixel. The variation index is proportional to pixels local entropy but inversely proportional to pixels kernel spatial attraction. The number of pixels removed was modulated by an artificial threshold factor α. Two real hyperspectral data sets were employed to examine the endmember extraction results. The reconstruction errors of preprocessing data as opposed to the result of original data were compared. The experimental results show that the number of distinct endmembers extracted has increased and the reconstruction error is greatly reduced. 100% is an optional value for the threshold factor α when dealing with no prior knowledge hyperspectral data.展开更多
基金Project(50808025) supported by the National Natural Science Foundation of ChinaProject(20090162110057) supported by the Doctoral Fund of Ministry of Education,China
文摘A new human action recognition approach was presented based on chaotic invariants and relevance vector machines(RVM).The trajectories of reference joints estimated by skeleton graph matching were adopted for representing the nonlinear dynamical system of human action.The C-C method was used for estimating delay time and embedding dimension of a phase space which was reconstructed by each trajectory.Then,some chaotic invariants representing action can be captured in the reconstructed phase space.Finally,RVM was used to recognize action.Experiments were performed on the KTH,Weizmann and Ballet human action datasets to test and evaluate the proposed method.The experiment results show that the average recognition accuracy is over91.2%,which validates its effectiveness.
基金Projects(61571145,61405041)supported by the National Natural Science Foundation of ChinaProject(2014M551221)supported by the China Postdoctoral Science Foundation,China+3 种基金Project(LBH-Z13057)supported by the Heilongjiang Postdoctoral Science Found,ChinaProject(ZD201216)supported by the Key Program of Heilongjiang Natural Science Foundation,ChinaProject(RC2013XK009003)supported by the Program of Excellent Academic Leaders of Harbin,ChinaProject(HEUCF1508)supported by the Fundamental Research Funds for the Central Universities,China
文摘A variation pixels identification method was proposed aiming at depressing the effect of variation pixels, which dilates the theoretical hyperspectral data simplex and misguides volume evaluation of the simplex. With integration of both spatial and spectral information, this method quantitatively defines a variation index for every pixel. The variation index is proportional to pixels local entropy but inversely proportional to pixels kernel spatial attraction. The number of pixels removed was modulated by an artificial threshold factor α. Two real hyperspectral data sets were employed to examine the endmember extraction results. The reconstruction errors of preprocessing data as opposed to the result of original data were compared. The experimental results show that the number of distinct endmembers extracted has increased and the reconstruction error is greatly reduced. 100% is an optional value for the threshold factor α when dealing with no prior knowledge hyperspectral data.
