Human Action Recognition Based On Manifold Learning

The research of human action recognition has become a frontier fields of thecomputer vision, while it has many applications in visual monitoring, motion analysis,human-computer interaction, virtual reality and so on. Video or image sequences areused for motion analysis as input. The extracted feature dimensionality is often too highto extract the intrinsic characteristics which could reflect the internal structure. And italso leads to the high time complexity and space complexity. Manifold learning is apopular nonlinear dimensionality reduction method in the field of machine vision andpattern recognition. The main purpose is to obtain the effective data information hiddenin the high dimensional space by looking for a mapping from high-dimensional space tolow-dimensional manifolds space.The main target of our study is to solve the non-supervisory issues of the classicmanifold learning methods by utilizing the category information of the samples. Thenexplore the combination between manifold learning and human action recognition.Finally, verify its effectiveness by the experiments on the action database.The main contributions of this paper are as follows:1. The Supervised-ISOMAP(S-ISOMAP, for short) algorithm is improved. It facesa problem that the algorithm could not get the explicit mapping in the classificationapplication when the new samples are added. Assume that the linear expressed relationsin the neighbor area are constant, then compute the neighbor expression by minimizingthe error of the linear expression. An supervised incremental ISOMAP(SI-ISOMAP, forshort) algorithm is proposed and it increases the recognition rate.2. An adaptive weighted neighbor distance factor is used to reconstruct theneighbor distance when compute the neighbor graph, different from S-ISOMAP tocompute the metric matrix using the category information. The distance from the targetpoint to other points in the local neighborhood is replaced by the average distance fromthe target point to the points of the category. The proposed algorithm is combined withGeneralization of ISOMAP(G-ISOMAP, for short) to improve the speed of the incremental learning. The experimental results prove that the recognition accuracy andcomputational efficiency have been greatly improved.3. Analyze the action sequence images to extract the single-cycle action sequence.The classic manifold learning methods vectorize the images and lose the spaceinformation. This paper proposes Distance Transform Local Maximum DissimilarityEmbedding(DT-LMDE, for short). The DT algorithm uses the distance to describe thepixel instead of the pixel value. LMDE algorithm makes the samples from differentclass away from each other in the neighborhood. Finally compute the similarity betweensequences using mean Hausdorff distance. The experiments show that the result of thismethod is effective.