1

2

Outline 1. Introduction 2. Hierarchical activity model 3. Inference 4. Learning 5. Experimental results 6. Application: Opportunity Knocks 7. Conclusions

3

1. Introduction

Track and predict a user’s location Infer a user’s mode and predict when and where Infer the locations of transportation destinations. Indicate he or she has made an error.

4

2. Hierarchical activity model

Locations and transportation modes Trip segments Goals Novelty

5

two time slices k − 1 and k

novelty mode

next goal andcurrent trip segment

GPS sensor measurement

6

Locations and transportation modes xk =〈 lk , vk , ck〉 zk ： GPS sensor measurements

generated by the person carrying a GPS sensor

mk ： transportation mode BUS, FOOT, CAR, and BUILDING.

The domain of τk is the set of outgoing neighbors of the current edge;

θk is picked from the edges close to zk

7

Trip segments A trip segment is defined by its start location, tk.start, end

location, tk.end, and the mode of transportation, tk.mode.

： trip switching

： a counter that measures the time steps until the next transportation mode is entered.

8

Goals & Novelty A goal represents the current target location of the person.

gk ： goal

： goal switching node

nk ： indicating whether a user’s behavior is consistent with historical patterns.

9

3. Inference Flat model

Rao–Blackwellized particle filter for estimation in the flat model

RBPF algorithm for the flat model Hierarchical model

RBPF algorithm for the hierarchical model Detecting novel behavior

10

Rao–Blackwellized particle filter for estimation in the flat model

Sampling step Kalman filter step Importance weights

11

histories

location of the person

car location

12

RBPF algorithm for the flat model

13

Sampling step

Kalman filter step

Importance weights

RBPF algorithm for the hierarchical model

14

Detecting novel behavior

15

4. Learning Finding mode transfer locations

estimate the mode transition probabilities using the expectation maximization (EM) algorithm

Finding goals a person typically spends extended periods of time whether

indoors or outdoors. Estimating transition matrices

use EM to estimate the transition matrices

16

5. Experimental results

Activity model learning Empirical comparison to other models Error detection

17

Activity model learning

six most common transportation goals frequently used bus stops and parking lots, the common routes using different modes of

transportation

18

Empirical comparison to other models

19

Error detection

20

6. Application: Opportunity Knocks The name of our system is derived from the desire to

provide our users with a source of computer generated opportunities

it plays a sound like a door knocking to get the user’s attention.

21

Opportunity Knocks(1)

22

Opportunity Knocks(2)

23

Opportunity Knocks(3)

24

Opportunity Knocks(4)

25

Opportunity Knocks(5)

26

Opportunity Knocks(6)

27

7. Conclusions provide predictions of movements to distant goals, and

support a simple and effective strategy for detecting novel events that may indicate user errors

system limitation ： it uses f ixed thresholds to extract goals and mode transfer locations.

28