Robust Human Action Recognition Using Dynamic Movement Features

Action recognition has been widely researched in video surveillance, auxiliary medical care and robotics. In the context of robotics, in order to program robots by demonstration (PbD), we not only need our algorithms to be capable of identifying different actions, but also to be able to encode and reproduce them. Dynamic movement primitives (DMPs), as a trajectory encoding method, are widely used in motion synthesize and generation. But at the same time it can also be applied to action recognition. With this idea, this paper extracts a kind of dynamic features from the original trajectory within DMP framework. The feature is temporal-spatial invariant. Based on the feature, FastDTW-KNN algorithm is proposed to solve the recognition task. Experiments tested on HAR dataset and handwritten letters dataset achieved an excellent recognition performance under a large data noise, which has verified the effectiveness of our method. In addition, comparative recognition experiments based on the original feature and our extracted dynamic feature are conducted. Results show that the dynamic feature is robust under temporal and spatial noise. As for classifiers, we compared our method with KNN, SVM and DTW-KNN followed with a detailed analysis of their advantages and disadvantages.