Plan and goal recognition is the task of inferring the plan and goal of an agent through the observation of its actions and its environment and has a number of applications on computer-human interaction, assistive technologies and surveillance. Although such techniques using planning domain theories have developed a number of very accurate and effective techniques, they often rely on assumptions of full observability and noise-free observations. These assumptions are not necessarily true in the real world, regardless of the technique used to translate sensor data into symbolic logic-based observations. In this work, we develop plan recognition techniques, based on classical planning domain theories, that can cope with observations that are both incomplete and noisy and show how they can be applied to sensor data processed through deep learning techniques. We evaluate such techniques on a kitchen video dataset, bridging the gap between symbolic goal recognition and real-world data.