Adaptive LLE Method Based Global Distance And Its Application In Human Activity Recognition

Human activity recognition has become one of the most active topics in the artificial intelligence and pattern recognition field. For the high dimension of human activity data, an effective method to reduce dimension is hardly needed. The original linear dimensionality reduction method is useless to process the complex human activity data. Recently locally linear embedding as an effective manifold-learning method has been widely used because of its low computational complexity, few configuration parameters and excellent robustness. But the original LLE required intensive and smooth sample data, while human activity data is usually sparse and unsmooth. Meanwhile, the intensive sample data count against distinction of human activity. As a result, the original LLE perform poorly in reducing dimension of human activity data. Based the discuss above, we have done the work as follows:(1)This article compared some classical lineal and nonlinear manifold-learning methods. Starting from their theoretical framework, we focus their scope of utilization, time complexity and strengths and weaknesses. As the purpose of improve dimensionality reduction method to applicable to human activity data;(2)In this paper, we studied original LLE based Euclidean distance selecting neighbors and finished further improvements. The improved method is adaptive LLE based global distance. It improved the method for calculating the nearest neighbors with introducing a global factor in original formula. It can shorten the distance of samples and redesign samples of large difference of distribution to get tight and smooth samples. It is useful to overcome the disadvantage of original LLE. The method can adaptively determine the intrinsic structure of human activities by residual variance of sample in input space and embedding space. It improved dimension reduction and computing performance;(3)By comparison of several motion capture technology, we adopt the advanced optical motion capture system to sample human activity data. We use adaptive LLE based global distance to process sparse sample and finally complete human activity recognition. Experimental results validate the method perform excellent in application of activity recognition and have achieved higher recognition rate than the original LLE.