Learning Activity-dependent Dynamic Bayesian Networks with Qualitative Constraints for Activity Recognition

In this paper, we propose solutions for learning activitydependent dynamic Bayesian network (DBN) for human activity recognition. As our model is designed to capture the underlying state dependencies among multiple features, a DBN with unique structure and parametrization is learned for each activity to encode its specific state dependencies. To alleviate the common problem of lack of sufficient training data, we propose a new structure learning algorithm, constrained structural EM (CSEM), to learn the activitydependent model structures combining the training data with qualitative and generic domain knowledge. Expressed as the qualitative constraints on the DBN model structure and parameters, the domain knowledge is extracted from generic physical and dynamic rules that govern the behavior of the human activities. The experimental results demonstrate the effectiveness of the activity-dependent DBN model for recognizing basic human activities. They further demonstrate that the simple and generic qualitative constraints can compensate quite effectively for the shortage of the training data and therefore reduce our dependencies on training data.