Active Action Recognition in Humans and Robots

Active action recognition is the selection of the best viewpoints for more accurate and faster action recognition. The studies in this thesis aimed to examine whether and how humans and robots can process efficient active action recognition. Participants could either be humans or robots. The robots were recurrent convolutional neural networks trained to classify actions using supervised learning and select the next best view by deep Q-learning. Each participant classified human actions from different viewpoints in either active or passive conditions. The participants in the active condition could select the viewpoint movements, whereas those in the passive conditions had no control over their viewpoints. The passive conditions could either be no view movement (NM) or random view movement (RM). In the NM condition, the view did not change within trials. In the RM condition, the viewpoint moved randomly. The studies showed that humans were slightly more accurate and faster in recognizing actions in the active condition than in the RM condition. However, some studies did not replicate this advantage in humans. Nevertheless, the robots were more accurate in the active condition than in the passive conditions. The efficient viewpoints for action recognition of humans and robots were the ones from which their action recognition was more accurate or faster in their NM condition. The efficient views tended to be the top, the front, and the side views with respect to the actors. Humans in the active condition selected the efficient viewpoints more often than the others. However, robots in the active condition did not choose the efficient viewpoints more frequently than the others. Instead, the robots moved their viewpoints far from their starting positions, suggesting they learned to get more accurate action classifications by observing actions from many viewpoints at different timepoints rather than from a few efficient viewpoints.