Gal A. Kaminka: Publications

Sorted by DateClassified by Publication TypeClassified by TopicGrouped by Student (current)Grouped by Former Students

Non-stationary Hidden Semi Markov Models in Activity Recognition

Einat Marhasev, Meirav Hadad, and Gal A. Kaminka. Non-stationary Hidden Semi Markov Models in Activity Recognition. In Proceedings of the AAAI Workshop on Modeling Others from Observations (MOO-06), 2006.

Download

[PDF]187.5kB  

Abstract

Activity recognition is a process by which the ongoing observed behavior of an agent is tracked and mapped to a given model, explaining the behavior and accounting for hidden or unobservable state (e.g., goals or beliefs of the obser ved agents). Various methods for activity recognition exist. A popular family of such methods rely on Hidden Markov Models HMMs and variants for recognition. These models, however, do not account for changes in transition probabilities base d on the duration an agent has spent in a given state. This paper investigates Markov models that go beyond existing models, to explicitly model the dependency of transition probabilities on state duration. In particular, we propose the use of Non-stationary Hidden Semi Markov Models (NHSMMs) in activity recognition. We present the NHSMM model, and compare its performance in recognizing normal and abnormal behavior, using synthetic da ta from an industry simulator. We show that for relatively simple activity recognition tasks, both HSMMs and NHSMMs easily and significantly outperform HMMs. In more complex tasks, the NHSMMs also outperf orm the HSMMs, and allow significantly more accurately recognition.

Additional Information

BibTeX

@InProceedings{moo06einat, 
  author = 	 {Einat Marhasev and Meirav Hadad and Gal A. Kaminka}, 
  title = 	 {Non-stationary Hidden Semi Markov Models in Activity Recognition}, 
  OPTcrossref =  {}, 
  OPTkey = 	 {}, 
  booktitle = MOO-06, 
  OPTpages = 	 {}, 
  year = 	 {2006}, 
  abstract = { 
      Activity recognition is a process by which the ongoing observed 
      behavior of an agent is tracked and mapped to a given model, 
      explaining the behavior and accounting for hidden or 
      unobservable state (e.g., goals or beliefs of the obser ved 
      agents). Various methods for activity recognition exist. A 
      popular family of such methods rely on Hidden Markov Models HMMs 
      and variants for recognition. These models, however, do not 
      account for changes in transition probabilities base d on 
      the duration an agent has spent in a given state. This paper 
      investigates Markov models that go beyond existing models, to 
      explicitly model the dependency of transition probabilities on 
      state duration. In particular, we propose the use of  
      Non-stationary Hidden Semi Markov Models (NHSMMs) in activity 
      recognition.  We present the NHSMM model, and compare its 
      performance in recognizing normal and abnormal behavior, using 
      synthetic da ta from an industry simulator.  We show that for 
      relatively simple activity recognition tasks, both HSMMs and 
      NHSMMs easily and significantly outperform HMMs. In more complex 
      tasks, the NHSMMs also outperf orm the HSMMs, and allow 
      significantly more accurately recognition. 
 }, 
  wwwnote = {}, 
  OPTeditor = 	 {}, 
  OPTvolume = 	 {}, 
  OPTnumber = 	 {}, 
  OPTseries = 	 {}, 
  OPTaddress = 	 {}, 
  OPTmonth = 	 {}, 
  OPTorganization = {}, 
  OPTpublisher = {}, 
  OPTnote = 	 {}, 
  OPTannote = 	 {} 
} 

Generated by bib2html.pl (written by Patrick Riley ) on Mon Nov 16, 2020 22:25:46