Object, scene and ego-centric action classification for robotic vision

Social robotics research focuses on developing intelligent machines that can perform difficult tasks especially those beyond human capabilities, involving repetitions or adverse conditions. However, robotic vision for high-level recognition associated with reasoning of uncertainty in unstructured environments is still far-fetched compared to the visual comprehension of humans. This dissertation investigates techniques to exploit the rich information from multi-modality sensors in pursuit of extending the frontiers of robotic vision. Three robot-centric recognition tasks are specifically explored; object, scene and action recognition with particular emphasis on designing highly effective and efficient feature representation and recognition algorithms.