Innovative analysis methods applied to data extracted by off-the-shelf peripherals can provide useful results in activity recognition without requiring large computational resources. We propose a framework for automated posture and gesture detection, exploiting depth data from Microsoft Kinect. Novel features are:

1. the adoption of Semantic Web technologies for posture and gesture annotation;
2. the exploitation of non-standard inference services provided by an embedded matchmaker [1] to automatically detect postures and gestures.

Particularly, the recognition problem is handled as a resource discovery, grounded on a semantic-based matchmaking [2]. An ontology for geometry-based semantic description of postures has been developed and encapsulated in a Knowledge Base (KB), also including several instances representing pose templates to be detected. 3D body model data detected by Kinect are pre-processed on-the-fly to identify key postures, i.e. unambiguous and not transient body positions. They typically correspond to the initial or final state of a gesture. Each key posture is then annotated adopting standard Semantic Web languages grounded on Description Logics (DL). Hence, non-standard inferences allows to compare the retrieved annotations with templates populating the Knowledge Base and a similarity-based ranking supports the discovery of the best matching posture. The ontology further allows to annotate a gesture from its component key postures, in order to enable recognition of gestures in a similar way.

The framework has been implemented in a prototype and experimental tests have been carried out on a reference dataset. Results indicate good posture/gesture identification performance with respect to approaches based on machine learning.