Attention-guided object and action recognition based on probabilistic learning and feature boosting for understanding human-object interaction

点击翻译按钮获取中文摘要

This research proposed probabilistic methods of attention-guided object recognition and object-oriented action recognition for understanding human-object interaction. In the proposed methods, attention-guided object recognition is performed in the context of co-occurring objects by using a classification tree which is learned based on the probabilistic latent component tree analysis and feature boosting. Also object-oriented action recognition is performed in the mutual contexts of objects and actions by using a probabilistic semantic network of visual motion classes and their semantic tags which is learned based on the incremental probabilistic latent component analysis. It was shown that the proposed method achieved high recognition accuracy through experiments using image data sets of plural object categories and also a set of video clips of object-oriented actions captured by a Kinect sensor mounted on a robot.