Human robot interaction in a crowded environment

Human Robot Interaction (HRI) is the primary means of establishing natural and affective communication between humans and robots. HRI enables robots to act in a way similar to humans in order to assist in activities that are considered to be laborious, unsafe, or repetitive. Vision based human robot interaction is a major component of HRI, with which visual information is used to interpret how human interaction takes place. Common tasks of HRI include finding pre-trained static or dynamic gestures in an image, which involves localising different key parts of the human body such as the face and hands. This information is subsequently used to extract different gestures. After the initial detection process, the robot is required to comprehend the underlying meaning of these gestures [3]. Thus far, most gesture recognition systems can only detect gestures and identify a person in relatively static environments. This is not realistic for practical applications as difficulties may arise from people‟s movements and changing illumination conditions. Another issue to consider is that of identifying the commanding person in a crowded scene, which is important for interpreting the navigation commands. To this end, it is necessary to associate the gesture to the correct person and automatic reasoning is required to extract the most probable location of the person who has initiated the gesture. In this thesis, we have proposed a practical framework for addressing the above issues. It attempts to achieve a coarse level understanding about a given environment before engaging in active communication. This includes recognizing human robot interaction, where a person has the intention to communicate with the robot. In this regard, it is necessary to differentiate if people present are engaged with each other or their surrounding environment. The basic task is to detect and reason about the environmental context and different interactions so as to respond accordingly. For example, if individuals are engaged in conversation, the robot should realize it is best not to disturb or, if an individual is receptive to the robot‟s interaction, it may approach the person. Finally, if the user is moving in the environment, it can analyse further to understand if any help can be offered in assisting this user. The method proposed in this thesis combines multiple visual cues in a Bayesian framework to identify people in a scene and determine potential intentions. For improving system performance, contextual feedback is used, which allows the Bayesian network to evolve and adjust itself according to the surrounding environment. The results achieved demonstrate the effectiveness of the technique in dealing with human-robot interaction in a relatively crowded environment [7]