Improved LLE algorithm for motion analysis

Improved LLE algorithm is proposed. Usually there is translation in practical data. The form of translation relies on the type of data. Data alignment is a very important step for linear or nonlinear dimensionality reduction methods. Original LLE algorithm uses Euclidean distance to find neighbors, and Euclidean distance is not a good measurement for translated data. Euclidean distance, Hausdorff distance and SSP distance are discussed, and SSP distance is used for improved LLE algorithm. Two methods are put forward for improving LLE algorithm. One is aligning input data of original LLE algorithm, the other is modifying LLE algorithm itself. SSP distance is used to find neighbors, translated data based on SSP distance is used for gaining reconstruction weight, and the plot for each dimension of LLE representation is used for visualization of LLE representation. Motion analysis experiments results show, improved LLE algorithm is better than original LLE algorithm for translated data, and obtains better visualization of LLE representation.