Human action recognition based on Adaptive Distance Generalization of Isometric Mapping

Manifold learning could effectively represent human action which has Non-linear characteristics. Isometric Mapping (ISOMAP) is a classic unsupervised algorithm of manifold learning. However, ISOMAP couldn´t work well for the data with the class priori information. Moreover, the computational complexity of dimension reduction to the new data point is too high to be used in real time. Considering two shortages of ISOMAP, the Adaptive Distance Generalization of Isometric Mapping (ADGI) is proposed, using human action silhouette sequences as the features, in which the adaptive distance factor is introduced to combine with generalization of ISOMAP. Finally the nearest neighbor classifier is used for recognition. For the dimension reduction of human action features, ADGI is effective. Experiments in Weizmann database show the presented algorithm is better both in recognition ratio and in real time for human action recognition.