Research On Human Action Prediction Method In Video

Action prediction is a challenging task in the computer vision field and has many potential applications in video surveillance,auto driving,and human-computer interaction.Different from the after-the-fact recognition of actions,action prediction takes incomplete videos only containing the beginning part of actions as input and aims to predict the categories of actions as early as possible.Since actions can present different characteristics at different temporal stages,a key issue of action prediction is how to capture the temporal progress state of an ongoing action.To this end,we propose a novel multi-task deep forest,in which action progress analysis and action category analysis are treated as two relevant tasks.The proposed multi-task deep forest is a cascade structure of random forests and multi-task random forests.Unlike the traditional single-task random forests,multi-task random forests are built upon incomplete training videos annotated with both action category labels and action progress labels.Meanwhile,incorporating both random forests and multi-task random forests can increase the diversity of classifiers and can improve the discriminative power of the multi-task deep forest.We further present a multi-task deep forest fusing regression and classification.The proposed method utilizes regression model to handle action progress analysis,while simultaneously adopting classification model to predict the action category.In essence,human motion is a continuous process,thus modeling the action progress analysis as a regression problem can better express the evolution of human actions.Moreover,a novel mid-level feature is constructed from the prediction results of action progress and category at each level of the deep forest.Extensive experiments were conducted on the UT-Interaction and the BIT-Interaction datasets to verify the effectiveness of the proposed methods.Detailed analysis and discussions of the experimental results are presented in this paper.