Boosted Markov Networks for Activity Recognition

Recognising human activities using sensors is currently a major challenge in research. Typically, the information extracted directly from sensors is either not discriminative enough or is too noisy to infer activities occurring in the scene. Human activities are complex and evolve dynamically over time. Temporal probabilistic models such as hidden Markov models (HMMs) and dynamic Bayesian networks (DBNs) have been the dominant models used to solve the problem [1, 4, 19]. However, these models make a strong assumption in the generative process by which the data is generated in the model. This makes the representation of complex sensor data very difficult, and possibly results in large models. Markov networks (MNs) (also known as Markov random fields) offer an alternative approach, especially in form of conditional random fields (CRFs) [10]. In CRFs, the observation is not modelled, and so we have the freedom to incorporate overlapping features, multiple sensor fusion, and long-range dependencies into the model.