Privacy free indoor action detection system using top-view depth camera based on key-poses

In this paper, we propose an indoor action detection system which can automatically keep the log of users' activities of daily life since each activity generally consists of a number of actions. The hardware setting here adopts top-view depth cameras which makes our system less privacy sensitive and less annoying to the users, too. We regard the series of images of an action as a set of key-poses in images of the interested user which are arranged in a certain temporal order and use the latent SVM framework to jointly learn the appearance of the key-poses and the temporal locations of the key-poses. In this work, two kinds of features are proposed. The first is the histogram of depth difference value which can encode the shape of the human poses. The second is the location-signified feature which can capture the spatial relations among the person, floor, and other static objects. Moreover, we find that some incorrect detection results of certain type of action are usually associated with another certain type of action. Therefore, we design an algorithm that tries to automatically discover the action pairs which are the most difficult to be differentiable, and suppress the incorrect detection outcomes. To validate our system, experiments have been conducted, and the experimental results have shown effectiveness and robustness of our proposed method.