An improved interest point detector for human action recognition

In this work, we present a method of human action recognition based on detection of interest points by spatial and temporal constraints. Firstly, the improved Harris-Laplace algorithm is proposed to solve the problem of multi-scale. Then, the bag-of-visual features (BoV) model is used for feature extraction, and is built the visual dictionary with K-means clustering. We train the Support Vector Machine (SVM) for action classification and prediction. In order to validate our approach, we verify our approach on existing benchmark datasets and some more complex scenes. Experiment results demonstrate that the human action recognition method we proposed shows state-of-the-art performance in complex scenes.