Research On Multi-Person Behavior Recognition Based On Video Features

Collective activity recognition is one of the hottest discussed and the most difficult research areas in computer vision. Nowadays, most recognition algorithms rely on the pre-processing algorithms and complicated models. Poselet is an efficient feature to describe the distribution of key points and action attributes are high-level semantic feature which can describe collective activity well. Both poselet and action attributes are good description of collective activity, based on which two recognition algorithms are proposed in this thesis.One is an improved recognition algorithm for collective activities based on poselet feature. As the poselet feature is lack of describing group behaviour recognition, we improve the description of poselet feature and add a complementary feature-gradient feature descriptor. During the training phase, the video sequence is divided and the picture frame is divided into grid cells. We use the gradient feature of grid cell cuboid as a useful complementary feature to poselet feature. The combined feature is better to describe the behaviour of the collective action.The other is a recognition algorithm for collective activities based on action attributes. As directly associating the activity feature of the video from the pose feature to the class label is insufficient of description, a high-level semantic concept- action attribute is utilized to represent the action. Action attributes used here are mainly divided into two parts, one is observed in the video manually and the other is automatically learned from data based on the information theory. We combine the manual attributes with data-driven attributes to represent the video sequence.Experimental results on Collective Activity Dataset 1 and Collective Activity Dataset 2 show the stability and effectiveness of proposed algorithm.