Download presentation
Presentation is loading. Please wait.
Published byBenedict Simmons Modified over 9 years ago
1
© 2005 Ritsumeikan Univ. All Rights Reserved. Embedded Action Detector to Enhance Freedom from Care Ritsumeikan University Graduate School of Computer Science Data Engineering Laboratory Kyohei Koyama
2
© 2005 Ritsumeikan Univ. All Rights Reserved. Tagged World Pocket Assistant Access Log Detect you going out RFID Tag Leaving Without locking Leaving something behind Coordination Alert ! Leaving the stove on Ubiquitous Facility Service !
3
© 2005 Ritsumeikan Univ. All Rights Reserved. Main Subject of This Presentation The Pocket Assistant is an embedded computer, thus it only has limited power resources The load can be huge, because the Pocket Assistant inspects all logs for each every access to the objects The new way to reduce the load, keeping the accuracy of human activity recognition
4
© 2005 Ritsumeikan Univ. All Rights Reserved. Definitions of Human Activity The human activity is composed of three elements Act : A Minimum unit of human activity i.e. an access to an object Action : A sequence of acts Behavior : A set of actions
5
© 2005 Ritsumeikan Univ. All Rights Reserved. Definitions of Human Activity Turning the knob Undoing the door chain Pushing the power button Unlocking the door Opening the door Putting on shoes Turning off TV Going out Behavior Action Act Having a bag Having baggage Taking the remote control Putting on shoes Taking a shoehorn Taking shoes
6
© 2005 Ritsumeikan Univ. All Rights Reserved. Bayesian Network Bayesian Network methodology is applied for inspecting the access logs Shoes Chain Result (Going outside) Probability Propagation Observed Value is Assigned Probability Variable is Changed Look Knob Shoehorn Door Key The probability of user going outside is 78%!
7
© 2005 Ritsumeikan Univ. All Rights Reserved. Initial Approach Term Sequence Bayesian Network Access Log Candidates time Second Stage First Stage Act Detect a Behavior!!
8
© 2005 Ritsumeikan Univ. All Rights Reserved. Experiment “Going outside” behavior Two kinds of cases are prepared True case : When the user go outside False case : When it looks like the user is going outside, but actually staying home 324 cases have been sampled in total
9
© 2005 Ritsumeikan Univ. All Rights Reserved. Ideas from Experiment ( Threshold Value ) False Cases True Cases Threshold Value Threshold Value
10
© 2005 Ritsumeikan Univ. All Rights Reserved. Ideas from Experiment ( Key Event ) BN1 BN2 BN3 BN4 Shoes Shoehor n Shoe s Loc k Key Shoe s Lock Graph 1 Graph 2 Graph 3 Graph 4
11
© 2005 Ritsumeikan Univ. All Rights Reserved. Ideas from Experiment ( Key Event ) The occurrence probability does not change dramatically when accesses other than the key event occur It is reasonable to calculate the probability only when the Key Event occurs The Key Event is effective to reduce the number of calculation for the probability of the Bayesian Network
12
© 2005 Ritsumeikan Univ. All Rights Reserved. Revised Approach time Access Log Bayesian Network Trigger Layoff (0.5 ~ 1.0sec) Layoff ( 0.5 ~ 1.0sec ) Detection of Key Event Term Sequence Initial Approach
13
© 2005 Ritsumeikan Univ. All Rights Reserved. Evaluation Case ID Initial ApproachRevised Approach Number of Times (%) True TC1 TC2 TC3 TC4 TC5 TC6 114 78 53 66 59 37 27 21 15 18 16 4 (23.68) (26.92) (28.30) (27.27) (27.12) (10.81) False FC1 FC2 FC3 18 67 76 222222 (11.11) (2.99) (2.63) Total-14.80% The revised approach reduces the number of calculation by 14.8% compared with the initial approach
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.