Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.
Learning and Inferring Transportation Routines
Outline <ul><li>Introduction </li></ul><ul><li>Hierarchical Activity Model </li></ul><ul><li>Inference, Learning, and Erro...
Introduction <ul><li>A hierarchical Markov model that can learn and infer a user’s daily movements through the community <...
Hierarchical Activity Model <ul><li>Locations and transportation modes </li></ul><ul><li>Trip segments </li></ul><ul><li>G...
 
Inference <ul><li>GPS-based tracking on street maps </li></ul><ul><li>Goal and trip segment estimation </li></ul>
<ul><li>The factorization (1) separates the state space of our estimation problem into its continuous and discrete parts <...
 
Learning <ul><li>Finding goals </li></ul><ul><li>Finding mode transfer locations </li></ul><ul><li>Estimating transition m...
Finding mode transfer locations
Detection of User Errors <ul><li>The hierarchical tracker has “expensive” particles, each containing much state informatio...
Experimental Results   <ul><li>Activity model learning </li></ul><ul><li>Detection of user errors </li></ul>
<ul><li>(a) Probability of the true goal (work place) during an episode from home to work, estimated using the flat and th...
 
 
 
Thank you
Upcoming SlideShare
Loading in …5
×

071203

586 views

Published on

Published in: Economy & Finance, Education
  • Be the first to comment

  • Be the first to like this

071203

  1. 1. Learning and Inferring Transportation Routines
  2. 2. Outline <ul><li>Introduction </li></ul><ul><li>Hierarchical Activity Model </li></ul><ul><li>Inference, Learning, and Error Detection </li></ul><ul><ul><li>Inference </li></ul></ul><ul><ul><li>Learning </li></ul></ul><ul><ul><li>Detection of User Errors </li></ul></ul><ul><li>Experimental Results </li></ul>
  3. 3. Introduction <ul><li>A hierarchical Markov model that can learn and infer a user’s daily movements through the community </li></ul><ul><li>The model allows : </li></ul><ul><ul><li>Infer the locations of usual goals, such as home or workplace </li></ul></ul><ul><ul><li>Infer a user’s mode of transportation, such as foot, car, or bus, and predict when and where she will change modes </li></ul></ul><ul><ul><li>Predict her future movements, both in the short term and in terms of distant goals </li></ul></ul><ul><ul><li>Infer when a user has broken his ordinary routine in a way that may indicate that he has made an error, such as failing to get off his bus at his usual stop on the way home </li></ul></ul><ul><ul><li>Robustly track and predict locations even in the presence of total loss of GPS signals and other sources of noise </li></ul></ul>
  4. 4. Hierarchical Activity Model <ul><li>Locations and transportation modes </li></ul><ul><li>Trip segments </li></ul><ul><li>Goals </li></ul>
  5. 6. Inference <ul><li>GPS-based tracking on street maps </li></ul><ul><li>Goal and trip segment estimation </li></ul>
  6. 7. <ul><li>The factorization (1) separates the state space of our estimation problem into its continuous and discrete parts </li></ul><ul><li>The left term on the right hand side of (1) , (2) follows by applying Bayes rule and the independences in our estimation problem </li></ul>
  7. 9. Learning <ul><li>Finding goals </li></ul><ul><li>Finding mode transfer locations </li></ul><ul><li>Estimating transition matrices </li></ul>
  8. 10. Finding mode transfer locations
  9. 11. Detection of User Errors <ul><li>The hierarchical tracker has “expensive” particles, each containing much state information, but requires few particles for accurate tracking </li></ul><ul><li>The flat, untrained tracker needs more particles to maintain tracking, but each particle is cheaper to compute </li></ul>
  10. 12. Experimental Results <ul><li>Activity model learning </li></ul><ul><li>Detection of user errors </li></ul>
  11. 13. <ul><li>(a) Probability of the true goal (work place) during an episode from home to work, estimated using the flat and the hierarchical model </li></ul><ul><li>(b,c) Probability of an activity being normal or abnormal, estimated by the concurrent trackers </li></ul><ul><li>(b) Normal activity: driving from home to work </li></ul><ul><li>(c) Abnormal activity: missing to get off the bus at time t2 </li></ul>
  12. 17. Thank you

×