| Recognizing human action has been an important research area in computer vision while enabling wide applications, e.g., surveillance, human-computer interaction and so forth. Most research mainly focused on videos containing simple actions recorded in controlled settings.However, there is still a great challenge to recognize actions from more realistic datasets due to background clutter, viewpoint changes and occlusion. An efficient and robust action recognition method is in urgent need with the rapid growth of the number of videos.Most action recognition approaches ignored the relationships among multiple action categories. However, action categories are correlated and share common features. Training classifiers with category preserving could take full advantage of common features shared between categories, while improving the generalization performance with less videos. In this paper, we explore to share knowledge in action categories in a multi-task learning framework.Firstly, we propose multi-task action recognition based on category boundary distance. The proposed method leverages support vector machine to measure the degree of similarity between each pair of categories and obtains the implicit group structure by clustering. In order to encourage the categories in the same group to share feature and vice versa, the implicit group structure is used as the prior regularization in multi-task learning. Our experiments on datasets HMDB51 and UCF50 show that the proposed method has achieved the comparative even higher accuracy over several state-of-the-art approaches.Secondly, we propose multi-task action recognition based on category Gaussian distribution.We employ Fisher Vector for video representation, which is concatenated by the gradients of likelihood with respect to the mean vector and covariance parameters of GMM and that the contribution of each Gaussian component to different action classes can be different.The proposed method employs mutual information to evaluate the importance of a Gaussian component in classifying a certain category. Only feature dimension of the Gaussian with a higher score remains. The number of Gaussian components that two categories hold at the same time is counted to measure the degree of similarity between them. Based on the similarity relationship the accurate implicit group structure of categories would be obtained by clustering and combined with multi-task learning framework to recognize actions. Experimental results on datasets HMDB51 and UCF50 show that the proposed method achieves an obvious improvement of accuracy over several state-of-the-art approaches. |
