Body posture graph: a new graph‐based posture descriptor for human behaviour recognition

Abstract

Behaviour recognition and analysis is one of the most challenging problems in video processing. In this study, the authors propose a new graph‐based representation of body silhouette by means of non‐linear mixture modelling. Firstly, the authors suppose the body silhouette as an objective function and try to approximate it by a set of elliptical basis functions (EBFs). By using parameters of these learned EBF kernels, vertices and edges of a graph are created. Since this graph is highly matched with the real skeleton of the body silhouette and represents the posture, they name it body posture graph (BPG). Then a posture descriptor is constructed by sorting the BPG's vertices according to a position‐dependent ordering algorithm. Thus, the descriptor contains not only body limbs connectivity information but also the spatial information. Therefore the descriptor is very effective for body posture description and is accurate for behaviour recognition purposes. They use simple fully‐connected hidden Mrakov model to learn and classify the sequences of postures. Good performance of the proposed features was proved by results of various and numerous experiments which were implemented on three different datasets: Sinica Academia, KTH (Kungliga Tekniska Högskolan) and UCF (University of Central Florida) Sports Action.

Keywords

• KTH dataset
• UCF sport action datasets
• Sinica Academia datasets
• posture sequence classification
• fully-connected hidden Mrakov model
• body posture description