Tesis doctoral de Juan Pedro Bandera Rubio
Robotics research has evolved in the last decades towards agents more complex, versatile, and capable. Thus, robots begin to appear that are able to work in dynamic, unpredictable environments such as houses, restaurants or museums. It is important for these robots to interact with people in their surroundings in an easy and efficient way. These requirements lead to the concept of social robot. One of the characteristics of such a robot is its required ability to learn from others. In this sense, learning by imitation appears as one of the most powerful mechanism a robot can use to learn socially from a human teacher. Speech learning reinforcement is useful but not always available, thus it may be worthy to consider a learning-by-imitation system that is based only in vision. This thesis presents a system that meets these requirements. The system itself is one of the main contributions, as its architecture includes novel concepts both in its structure and the implementation of its components. the sensory input for the proposed system is restricted to the data provided by a pair of stereo cameras. The upper-body movements of the human performer are extracted from these data using a novel human motion capture system. Once captured, human motion is segmented into discrete gestures. Then, these gestures are codified using a proposed representation based on local and global features. Encoded perceived gestures are compared against the gesture repertoire of the robot, that contains all gestures the agent has already learnt. Local features are compared using dynamic programming alignment techniques. These local results are reinforced by the comparison of global features, performed using analytic algorithms. The final results are fed to a learning component, that determines whether the gesture repertoire should be modified or not. Uncertain situations are solved by asking for a small degree of human supervision. it is important to consider that all previous modules are independent from the particular motor abilities of the used robot. This is an important difference respect to other approaches. In the proposed architecture, the physical characteristics of the robot do not constraint its perceptual abilities, thus perceived human movements can be represented, recognized and learnt more precisely. The last module of the architecture translates motion from the human motion space to the robot one, using a combined strategy. It is employed only when imitation of perceived of learnt gestures is required. experimental results presented in this thesis include a quantitative evaluation of the human motion capture and translation modules. Besides, representation and recognition methods have been compared against other alternatives and evaluated independently. The last experiments involved the complete system working in real environments. These tests validate the proposed system as an interesting element to be integrated in a social robot.
Datos académicos de la tesis doctoral «Vision-based gesture recognition in a robot learning by imitation framwork«
- Título de la tesis: Vision-based gesture recognition in a robot learning by imitation framwork
- Autor: Juan Pedro Bandera Rubio
- Universidad: Málaga
- Fecha de lectura de la tesis: 22/01/2010
Dirección y tribunal
- Director de la tesis
- Juan Antonio Rodriguez Fernandez
- Tribunal
- Presidente del tribunal: Francisco Sandoval hernandez
- adrian Hilton (vocal)
- angel l Pasqual del pobil y ferre (vocal)
- lakmal Seneviratne (vocal)