were applied separately to make a comparison. Further, to assess the effect of the gait features obtained on the classification performance, we trained a SVM classifier on the gait classification task. The selection of kernel function K( ) x, y is very important in SVM because it defines the nature of the decision surface that separate gait data. Here, we used the following kernels functions: (1) Linear: ( ) i j i j K x , x = x ⋅ x . (2) Polynomial (Poly): ( )( ) ( )d i j i j K x , x = x ⋅ x +1 , where d is the polynomial’s degree. (3) Gaussian radial basis function (RBF): ( ) ⎟ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎜ ⎝ ⎛ − = 2 2 2 , exp σ i j i j x x K x x , where σ is the width of RBF function The detailed training procedure of gait classification was as follows: firstly, we constructed an initial training set, that is, initialization of the training set or samples, and optimization of parameter such as regularization parameter C of SVM, the number of neighboring points k and the number of feature dimensionality n selected of ISOMAP and the kernel parameter d and σ of SVM. Secondly, we used the crossvalidation method to adjust the optimal parameters k , n , C , d andσ till the best classification performance can be produced. Here, with the small size sample data, cross-validation method was used to evaluate the gait classification performance[2]. Here, we designed a six-fold cross-validation scheme, that is, all 60 subjects’ data were divided into six segments. Each of the six cross-validation test segments contained 5 young and 5 elderly subjects while their respective training segment included the remaining 25 young and 25 elderly. Firstly, 5 out of the 6 segments were used to train and construct the SVM decision surface, and the remaining one was used for testing. Secondly, the above procedures were repeated for 6 times. Finally, the six classification results were averaged to obtain a final performance result. Table Ⅰ presented the optimum results from our experiment. From Table 1, we could see that the combination of ISOMAP with SVM reached the best performance( 88% accuracy), followed by the PCA-SVM and then the original SVM. These results demonstrated that the non-linear gait features obtained by ISOMAP could provide more additional discriminatory information for improving classification performance. Also, we found that three kernel functions used in SVM performed well, suggesting that the type of kernel function of SVM may have a slight effect on the performance of SVM classifier. Besides, in this experiment, because the optimal value of each parameter could vary with different values of other parameters, they must carefully be selected by trial and error method. The optimal parameters selected were presented in TableⅠ. In conclusion, these results demonstrate that ISOMAP-SVM is a robust classifier for gait data. V. DISCUSSION AND CONCLUSION The results of this research suggest that nonlinear gait features obtained by ISOMAP could capture useful information about intrinsic structure of gait data with individuals during walking, and they could be employed to train a machine classifier with learning algorithm to automatically classify human gait patterns with superior generalization performance. In this study, the aim of the application manifold learning algorithm in gait data analysis is to obtain more significant information about human gait change for improving the gait classification performance. In the experiment, we find that ISOMAP algorithm is capable of obtaining more nonlinear gait features, which contain more relevant information about the intrinsic non-linear dynamics of human movement, than PCA. The reason is that PCA only can find a low-dimensional embedding of the gait data points that preserve the all pairs of points distance based on Euclidean in the high-dimensional input space. However, unlike Euclidean distance in PCA, the geodesic distance in ISOMAP algorithm can reflect the actual low-dimensional geometry structure on manifold in gait data, that is, ISOMAP algorithm can discover the intrinsic geometry structure of gait data by preserving geodesic distance on the manifold, in the words, it can find the low-dimensional embedding manifold structure in high-dimensional gait data[7, 8]. As we known, the generalization performance of the SVM classifier depends primarily on the selection or extraction from input variables, that is, the selected or extracted features as the linear dynamics of human movement. Each subject was asked to walk on the straight laboratory walkway of approximately 10 m at a self-determined pace. When a foot stepped on the middle of the strain gauge force platform embedded in the walkway, the foot-ground reaction forces(GRF) were recorded during walk. Here, the sampling frequency was set to 400 Hz. In order to avoid the individual difference, the acquired forces and the occurrence of their corresponding times were normalized by subjects’ weight and by gait cycle respectively. Thus, we can represent a gait pattern as 101 dimensions vector by sampling at each 1% in a time normalized stance phase. Here, the vertical GRF gait data, as shown in Fig.1, was selected to analyze. Figure.1. The foot–ground reaction forces along the vertical directions during a normalized stance phase III. MANIFOLD LEARNING ALGORITHM FOR NONLINEAR GAIT FEATURE EXTRACTION To obtain the significance low-dimensional embedding associated with the human gait function in high-dimensional input space, the manifold learning algorithm was applied to perform dimensionality reduction. Presently, there are three usual manifold embedding algorithm for dimensionality reduction: Isomap, Locally-Linear Embedding(LLE), Laplacian Eigenmaps(LE). In this study, considering the effective utilization of the global relationship between all gait variables, we selected the Isomap algorithm to perform dimensionality reduction for the actual discovery of the intrinsic nonlinear structure of the analyzed gait data. Isomap algorithm is a global geometric framework for nonlinear dimensionality reduction, and it can effectively use the measured geodesic distance and classical multidimensional scaling(MDS) technique to find the intrinsic geometry of data(i.e. the low-dimensional embedding in the highdimensional input space). In the algorithm, the geodesic manifold distances between all pairs of data points can be obtained by constructing the graph, and the geodesic distance can be estimated by finding the shortest path in the graph representation of the manifold. Here, we firstly determined the neighboring points by Euelidean distance d ( ) i j X , between the arbitrary pointsi, j in the gait data input space X , and then we can obtain the neighboring relation by selecting the k neighboring points. Thus, we can construct the weighted graph G with edges connecting neighboring point on the manifold M . Next step, we can estimate the geodesic distances d ( ) i j M , by finding the shortest paths d ( ) i j G , between all pairs of points in the constructed graph . Finally, we can use the graph distance d ( ) i j G , replace the distance in MDS, and we can find the low-dimensional embedding by minimizing the following cost function: ()() E DG DY = τ −τ (1) where τ is an operator for efficient optimization of the determined distance matrix D by converting the distances to inner products. The detailed procedure of solution for finding the low-dimensional embedding can be found in[7]. IV. EXPERIMENTS AND RESULTS After The gait nonlinear features can be obtained from gait data by using ISOMAP algorithm to perform dimensionality reduction according to its algorithm step in section 3. As a result, the nonlinear information feature of gait can be obtained easily by a simple transformation for the projected data. In this experiment, in order to evaluate the ability of the extraction of nonlinear gait features, ISOMAP and linear principal component analysis (PCA)