Hierarchical Markov Models have been recently applied to recognize human behavior at different levels of abstraction. The term behavior recognition is used to refer to the task of determining a high-level activity that a person is performing (e.g., cooking) based on a sequence of low-level observations (e.g., the ___location of the person in a room) often captured by devices such as video cameras and motion sensors. This article briefly introduces two kinds of Hierarchical Markov Models--Hierarchical Hidden Markov Models and Abstract Hidden Markov Models--and then discusses how they have been used for behavior recognition.
Examples of Hierarchical Markov Models
Hierarchical Hidden Markov Model
A Hierarchical Hidden Markov Model (HHMM) [1] is an extension of a Hidden Markov Model that contains states at different levels of abstraction. Thus, the structure of an HHMM is more suitable for representing the hierarchy in human activities. For instance, the activity of eating dinner could be decomposed in the activities of preparing ingredients, cooking, cleaning, and washing dishes. A corresponding HHMM would have a high-level state corresponding to the activity of eating dinner and low-level states corresponding to the sub-activities. The reader is referred to the article on Hierarchical Hidden Markov Models for more detailed information.
Abstract Hidden Markov Model
An Abstract Hidden Markov Model (AHMM) [2] is an extension of a Hierarchical Hidden Markov Model that allows the choice of how a high-level activity (policy) will be decomposed into a sequence of lower-level activities (policies) to be dependent on the current state of the environment. Thus, simple types of context-sensitive behaviors can be captured by an AHMM [3].
Applications to Human Behavior Recognition
Hierarchical Hidden Markov Models (HHMMs) have been used in the context of eldercare aiming to identify behaviors such as eating dinner, watching television, and cooking [3]. The goal is to learn the pattern of executing such activities normally, so that when an abnormal execution occurs (e.g., an elder person falls down while cooking), it can be automatically and instantaneously detected. Then, the appropriate action (e.g., provide audio/video cues to help the elderly or notify the appropriate caretaker that the elderly needs help) can then be taken before negative consequences for the elder person's health occur.
Supervised learning (a class of machine learning techniques) was used in [3] to learn an HHMM for four simplistic ways for dinner preparation ("spending some time preparing the food and rummaging through the fridge before the meal is cooked", "washing dishes prior to cooking on the stove", "going to the sink to wash the dishes then spending time at the food preparation area and the fridge before finally cooking the meal", "transition between each area of interest in a round robin fashion, starting and ending at the stove, before leaving the room") and five typical living room activities ("watch television", "read a book on the couch", "eat dinner", "eat dinner while watching television" and "there is nothing good on TV, read a book instead"). The person's position in the room was captured at short intervals of time and a sequence of observations was thus created. Observation sequences were labelled with the corresponding behavior and were used to learn HHMMs for the high-level dinner preparation and living room behaviors.
The HHMMs are then used to classify new sequences.
References
- ^ S. Fine and Y. Singer The hierarchical hidden markov model: Analysis and applications. Journal of Machine Learning, vol. 32, 1998.
- ^ H. H. Bui, S. Venkatesh, and G. West Policy recognition in the abstract hidden markov model. Journal of Artificial Intelligence Research, vol. 17, p. 451–499, 2002.
- ^ a b c S. Lu ̈hr, H. H. Bui, S. Venkatesh, and G. A. W. West Recognition of Human Activity through Hierarchical Stochastic Learning. PERCOM '03 Proceedings of the First IEEE International Conference on Pervasive Computing and Communications, 2003.