A relational learning approach to activity recognition from sensor readings

The ability of understanding humanpsilas behavior is a required component for many applications. This understanding includes, among other tasks, automatically generating and maintaining models of human actions, goals and plans. This paper presents a system to infer the actions that people perform in order to accomplish activities of daily living starting from sensory inputs. Our approach is based on using relational learning to infer predictions about which action has just been executed. We learn a model for recognizing executed actions based on the state changes detected from sensor readings. Each change has been produced by a performed action, while a sequence of these actions forms a plan to accomplish a high-level action or to achieve a goal. Using a relational learning tool, Tilde, we obtain classifiers to map changes in the states to actions performed by a user. We have performed some experiments using an environment simulator feeded by data gathered from real human behaviour. The results show that we can obtain a good accuracy even in presence of noise.