Human Action Recognition: A Dense Trajectory and Similarity Constrained Latent Support Vector Machine Approach

In this paper, we propose a novel approach based on two ideas, one using the trajectory extracted from videos as the local descriptors, and, second, using a similarity constrained latent support vector machine approach, to enforce the consistency of the selection of regions of interest generated by such trajectory based local extractors. We compare the performance of the proposed approach with other cognate state-of-the-art approaches, e.g., trajectory based approach, but using histogram matching techniques for classification, and, constrained similarity latent support vector machine approach, but, using appearance based local descriptors, on the UCF Sports dataset, and found that the proposed approach gives better recognition results.