TRB 2014 - Automatic Spatial-temporal Identification of Points of Interest in Global Navigation...

Automatic Spatial-Temporal Identification of Points of Interest (ASTIPI) in Global Navigation Satellite

System Data

Khoa Tran1, Sean Barbeau, Ph.D.2, Miguel Labrador, Ph.D.1

Computer Science & Engineering1, Center for Urban Transportation Research 2

Slide # 2

Agenda

Introduction

ASTIPI Algorithm

Evaluation

Conclusions

Introduction Background ASTIPI Evaluation ConclusionsRelated Works

3Slide #

Introduction

Slide # 4


Point of Interest Specific geographical locations that users may find

useful or interesting Example: Home, work, restaurants, supermarket, schools, etc. Many transportation applications

School

Market

Restaurant

Slide # 5


GPS Dataset Properties Property #1: Spatial temporal data.

Each coordinate is associated with a recorded time Example: (-82.520950, 28.033525, 15m) at 09:14:54.777am

Property #2: Switching between stop and movement. Trip re-construction can be accomplished after successfully

identifying POIs.

Slide # 6


GPS Dataset Properties Property #3: Noise (i.e. outliers) and the loss of data

Without clear sky view (e.g. indoors, tree cover, mountains). POI may be represented by one or two GPS fixes, if building

blocks signal

Indoor Stationary Location

Slide # 7


GPS Dataset Properties Property #4: Full dataset and reduced datasets – due

to energy-saving algorithms Sample and send less GPS fixes to conserve battery life

0

2

4

6

8

10

12

14

16

14

4.21

RequirementSanyo Pro 200

Batt

ery

Life

(hou

rs)

Impact of GPS and Wireless Tx on Battery Life for a 4-second interval

Full Dataset Reduced Dataset

GPS Auto-Sleep – reduces density of clusters

GPS Auto-Sleep & Critical Point Algorithm – reduces density of clusters AND decimates path

Slide # 8


GPS Dataset Properties Property #5: Duplicated trips

Users often travel back and forth from one place to another using the same routes.

9Slide #


Automatic Spatial Temporal Identification of Points of Interest:

The Algorithm

Slide # 10


ASTIPI An extension of DBSCAN

Use MinTime (instead of MinPoints) to determine Core Point Keep track of the current index to improve running time Consider both spatial and temporal properties

Two main steps The Main ASTIPI Algorithm The Eps-K-Neighborhood Search

Slide # 11


Main ASTIPI Algorithm Determine POIs given a time-ordered list of GPS fixes

Form POI by finding Core Point P and all coordinates that are density-reachable from P or density-connected to P

Use the Eps-K-Neighborhood Search

Maintain the index of the last coordinate in Eps-K-Neighborhood of the most recent Core Point

Perform search only on necessary location data

Slide # 12


Eps-K-Neighborhood Search Find coordinates in the trajectory that are:

Spatially close to point P Temporally close in time to point P

Return neighbors if point P is a Core Point

Start the search from a given startIndex

Stop the search when the number of coordinates that are outside Eps distance exceeds K

Slide # 13


Sample Execution Parameter Values

MinTime = 300 seconds Eps = 100 meters K = 3

C1

C2

C3

C4

C5

C6

C7

C8C9 C10

C11

C12C13

C14

C15

Core Point

Coordinate inside POI

Coordinate outside POI

Noise

User stops for 15min(but single GPS point)

Slide # 14




C1

C2

C3

C4

C5

C6

C7

C8C9 C10

C11

C12C13

C14

C15

Core Point



Noise

User stops for 15min(but single GPS point)

Slide # 15




C1

C2

C3

C4

C5

C6

C7

C8C9 C10

C11

C12C13

C14

C15

Core Point



Noise

POI1

POI2

Slide # 16


Sample Input & Output Input

Output

17Slide #


Evaluation

Slide # 18


Experimental Design Use TRAC-IT Java ME app for data collection,

Sanyo Pro 200 mobile phone w/ assisted GPS

Ground truth: user carried the phone throughout a day, reporting visited places at end of day (via web interface)

Walking, driving, riding bus

Slide # 19


Experimental Design Two different datasets:

Use GPS Auto-Sleep (Full Dataset) Use both GPS Auto-Sleep and Critical Points

Algorithm (Reduced Dataset ~1/8th size of Full Dataset)

Full Dataset Reduced Dataset

GPS Auto-Sleep – reduces density of clusters

GPS Auto-Sleep & Critical Point Algorithm – reduces density of clusters AND decimates path

Slide # 20


Experimental Design Accuracy = TP / (TP + FP + FN) Evaluated against other known algorithms:

Density-Based Spatial Clustering of Applications with Noise (DBSCAN)

Spatial Temporal (ST) DBSCAN Clustering-Based Stops and Moves of Trajectories (CB-SMoT) Stay Point Detection (SPD) Fast Clustering (FC) of GNSS Data

Slide # 21


Full Dataset

ASTIPI CB-SMoT SPD FC0%

10%20%30%40%50%60%70%80%90%

100%88%

28%

65%

18%

85%

7%

26%

7%

Average Ac-curacy

Percent of Ac-curacy > 75%

Mode # of Tests ASTIPI(%)

CB-SMoT (%)

SPD(%)

FC(%)

Stationary 2 100 100 0 100Walking 3 69 39 46 37

Walking + Bus 3 67 23 60 22Walking + Bus + Driving 1 100 13 63 30

Walking + Driving 14 93 22 75 5Driving 4 94 15 82 3

Slide # 22



10%20%30%40%50%60%70%80%90%

100%

45%35% 31% 31%

19%11%

4%11%

Average Accuracy


Mode # of Tests ASTIPI(%)

CB-SMoT (%)

SPD(%)

FC(%)

Stationary 2 75 100 0 100Walking 3 34 24 31 29

Walking + Bus 3 35 30 32 22Walking + Bus + Driving 1 71 43 63 31

Walking + Driving 14 47 31 40 27Driving 4 34 24 36 21

Reduced Dataset

Slide # 23


ASTIPI-With-Speed Attempt to address low accuracy on reduced dataset Reduces the number of False Positives Core Point with Speed

A point p =(xp,yp,tp) of a trajectory is called core point with speed with respect to Eps, MinTime, K, and MaxSpeed if

(1) |tsm − tp|≥ MinTime where sm is the last point of N(Eps,K)(p) OR (2) |tsm − tsm+1 |≥ MinTime where sm is the last point of N(Eps,K)(p)

and sm+1 is the first next point not in N(Eps,K)(p) (3) The speed at point p is less than MaxSpeed

Slide # 24


ASTIPI-With-Speed


10%20%30%40%50%60%70%80%90%

100%

59%

35% 31% 31%

52%

11%4%

11%

Average Accuracy



20%

40%

60%

80%

100%87%

28%

65%

18%

81%

7%

26%

7%

Average Accuracy

Percent of Accuracy > 75%

On dataset with a GPS Auto-Sleep Module

On dataset with a GPS Auto Sleep Module and Critical Point Algorithm Strategy

Previously 45%

Previously 88%

Slide # 25


Execution Time Average execution time of twenty runs on each of the test

ASTIPI CB-SMoT SPD FC0.00

50.00

100.00

150.00

200.00

250.00

300.00

350.00

400.00

2.66

357.05

0.36

356.38

ms

26Slide #


Conclusions

Slide # 27


Summary - ASTIPI 88% of accuracy on datasets using GPS Auto-Sleep

module 59% of accuracy on reduced datasets using a

combination of GPS Auto-Sleep module and the CP algorithm

Linear running time – O(n) Maintains a temporal order of GPS coordinates for

fast trip segmentation

Slide # 28


Future Work Need to improve the accuracy on reduced datasets using

the CP algorithm 45% and 59% are low and cannot be used in practice Increase the number of True Positives

Perform tests in different areas Different Eps and K values More users and more data over more modes

Replace the absolute distance Eps with a relative parameter related to the dataset

On-the-fly processing

Slide # 29


Questions?

Sean J. Barbeau, [email protected]

mailto:[email protected]

Slide #ASTIPI - IWGS 2013 30


November 5th, 2013

ASTIPI - IWGS 2013 31Slide #


Extra Slides

November 5th, 2013



Motivation

November 5th, 2013

The rise of mobile devices and mobile apps In 2011, 87% world population used mobile devices and 50%

American mobile users have apps

The rise of location-based apps in smartphone 74% smartphone owners use LBS GPS and LBS devices will reach ~1,015 millions units by 2015 LBS that can personalize search and suggest places

Extract POI from mobile users



Definitions

November 5th, 2013

Trajectory Sample

Eps-K-Neighborhood

Core Point Uses MinTime (duration) instead of MinPoints

Directly Density-Reachable

Density-Reachable

Density-Connected

Point of Interest



Definitions - Example

November 5th, 2013

Trajectory Sample = {C1, C2, …, C15} Eps-K-Neighborhood of C3 = {C4, C5} Core Point = {C3, C8, C11, C13} C12 is directly density-reachable from C11 C11 is directly density-reachable from C8 C12 is density-reachable from C8 C12 and C14 are density-connected

Both are density-reachable from C8 Two points of interest

POI1 POI2



Summary

November 5th, 2013

Problem #1: Lack of POI identification algorithms that have a high accuracy.

Problem #2: Lack of POI identification algorithms that have a sufficiently high accuracy on a reduced GPS dataset for addressing the energy consumption problem on mobile devices.

Problem #3: Slow running time of O(n2) in worst case scenario.

Problem #4: Unable to maintain a temporal order of GPS fixes to support fast trip segmentation.



With GPS Auto-Sleep Module

November 5th, 2013

Test Mode ASTIPI CB-SMoT SPD FC

TP FP FN Acc (%) TP FP FN Acc (%) TP FP FN Acc (%) TP FP FN Acc (%)1 Stationary 1 0 0 100 1 0 0 100 0 1 1 0 1 0 0 1002 Stationary 1 0 0 100 1 0 0 100 0 1 1 0 1 0 0 1003 Walking 3 0 2 60 3 1 2 50 2 1 3 33 4 5 1 404 Walking 4 0 1 80 1 1 4 17 4 1 1 67 2 4 3 225 Walking 4 1 1 67 3 1 2 50 3 3 2 38 3 1 2 506 Walking + Bus 2 0 1 67 1 1 2 25 2 1 1 50 2 5 1 257 Walking + Bus 3 2 4 33 1 2 6 11 3 3 4 30 6 12 1 328 Walking + Bus 4 0 0 100 2 2 2 33 4 0 0 100 4 46 0 89 Walking + Bus + Driving 6 0 0 100 1 2 5 13 5 2 1 63 6 14 0 30

10 Walking + Driving 4 0 0 100 3 3 1 43 3 0 1 75 3 28 1 911 Walking + Driving 3 0 0 100 1 1 2 25 2 0 1 67 2 146 1 112 Walking + Driving 4 0 1 80 1 2 4 14 5 0 0 100 5 96 0 513 Walking + Driving 6 1 0 86 2 2 4 25 5 1 1 71 6 46 0 1214 Walking + Driving 5 0 1 83 2 5 4 18 5 1 1 71 6 95 0 615 Walking + Driving 4 0 0 100 2 5 2 22 3 0 1 75 3 192 1 216 Walking + Driving 5 0 0 100 2 1 3 33 4 1 1 67 5 114 0 417 Walking + Driving 3 0 0 100 0 1 3 0 3 2 0 60 3 150 0 218 Walking + Driving 7 1 0 88 1 5 6 8 6 2 1 67 7 118 0 619 Walking + Driving 6 0 1 86 1 6 6 8 5 1 2 63 7 163 0 420 Walking + Driving 6 0 0 100 4 8 2 29 6 1 0 86 6 130 0 421 Walking + Driving 12 1 0 92 3 4 9 19 12 1 0 92 12 190 0 622 Walking + Driving 6 0 0 100 3 2 3 38 6 2 0 75 6 92 0 623 Walking + Driving 9 0 2 82 3 4 8 20 9 1 2 75 9 288 2 324 Driving 5 1 0 83 3 56 2 5 4 1 1 67 5 335 0 125 Driving 5 0 0 100 2 3 3 25 4 0 1 80 5 182 0 326 Driving 8 0 0 100 8 88 0 8 8 0 0 100 8 142 0 527 Driving 10 0 1 91 3 3 8 21 9 0 2 82 10 278 1 3

Average Accuracy 88% 28% 65% 18%

Percent of Accuracy>75% 85% 7% 26% 7%



With GPS Auto-Sleep Module and CP Algorithm

November 5th, 2013

Test Mode ASTIPI CB-SMoT SPD FC

TP FP FN Acc (%) TP FP FN Acc (%) TP FP FN Acc (%) TP FP FN Acc (%)1 Stationary 1 0 0 100 1 0 0 100 0 0 1 0 1 0 0 1002 Stationary 1 1 0 50 1 0 0 100 0 0 1 0 1 0 0 1003 Walking 1 0 4 20 1 0 4 20 1 1 4 17 1 1 4 174 Walking 3 1 2 50 2 1 3 33 2 3 3 25 2 2 3 295 Walking 2 1 3 33 1 0 4 20 3 1 2 50 2 0 3 406 Walking + Bus 2 1 1 50 2 1 1 50 1 2 2 20 2 5 1 257 Walking + Bus 2 0 5 29 3 5 4 25 3 1 4 38 6 10 4 188 Walking + Bus 2 4 2 25 1 3 3 14 3 4 1 38 4 9 1 239 Walking + Bus + Driving 5 1 1 71 3 1 3 43 5 2 1 63 6 10 1 31

10 Walking + Driving 3 1 1 60 2 2 2 33 2 2 1 40 3 7 1 2711 Walking + Driving 2 1 1 50 2 1 1 50 2 2 1 40 2 4 1 2912 Walking + Driving 2 1 3 33 1 2 4 14 3 2 2 43 5 2 3 2913 Walking + Driving 5 3 1 56 1 1 5 14 5 4 1 50 6 6 3 2514 Walking + Driving 3 1 3 43 2 3 4 22 4 2 2 50 6 5 2 3615 Walking + Driving 3 2 1 50 2 1 2 40 3 3 1 43 3 8 1 2516 Walking + Driving 4 4 1 44 3 6 2 27 3 6 2 27 5 15 2 1517 Walking + Driving 2 2 1 40 2 1 1 50 2 2 1 40 3 9 0 2518 Walking + Driving 5 4 2 45 2 6 5 15 5 4 2 45 7 7 4 2119 Walking + Driving 6 4 1 55 3 4 4 27 5 4 2 45 7 5 3 3320 Walking + Driving 5 7 1 38 4 8 2 29 5 7 1 38 6 15 2 1921 Walking + Driving 10 2 2 71 8 1 4 62 9 5 3 53 12 6 2 5622 Walking + Driving 4 3 2 44 4 9 2 27 3 7 3 23 6 12 3 1723 Walking + Driving 6 11 5 27 5 7 6 28 6 13 5 25 9 20 6 1624 Driving 4 6 1 36 5 14 0 26 5 6 0 45 5 34 1 1025 Driving 3 3 2 38 1 3 4 13 3 3 2 38 5 2 2 4326 Driving 5 9 3 29 5 4 3 42 5 8 3 31 8 20 2 2127 Driving 6 8 5 32 5 18 6 17 6 9 5 30 10 44 5 11

Average Accuracy 45% 35% 31% 31%

Percent of Accuracy>55% 19% 11% 4% 11%



ASTIPI-With-Speed

November 5th, 2013

TestOn dataset with a GPS Auto-Sleep Module On dataset with a GPS Auto Sleep Module and Critical

Point Algorithm Strategy

TP FP FN Acc (%) TP FP FN Acc (%)1 1 0 0 100 1 0 0 1002 1 0 0 100 1 0 0 1003 3 0 2 60 1 0 4 204 4 0 1 80 3 1 2 505 4 1 1 67 2 1 3 336 2 0 1 67 2 1 1 507 3 2 4 33 2 0 5 298 4 0 0 100 2 2 2 339 6 0 0 100 5 1 1 71

10 4 0 0 100 3 0 1 7511 3 0 0 100 2 0 1 6712 4 0 1 80 2 0 3 4013 6 1 0 86 5 1 1 7114 5 0 1 83 3 1 3 4315 3 0 1 75 3 0 1 7516 5 0 0 100 4 1 1 6717 3 0 0 100 2 1 1 5018 7 1 0 88 4 1 3 5019 6 0 1 86 4 2 3 4420 6 0 0 100 6 0 0 10021 12 1 0 92 9 1 3 6922 6 0 0 100 4 1 2 5723 9 0 2 82 7 2 4 5424 5 1 0 83 4 0 1 8025 5 0 0 100 3 0 2 6026 8 0 0 100 6 1 2 6727 10 0 1 91 6 1 5 50

Average Accuracy 87% 59%

Percent of Accuracy>75% 81% Percent of Accuracy>55% 52%

Previously 45%Previously 88%



Execution Time

November 5th, 2013

Average execution time of twenty runs on each of the datasetTest Size (Points) ASTIPI (ms) CB-SMoT (ms) SPD (ms) FC (ms)

1 75 0.05 0.75 0.05 0.702 79 0.05 0.75 0.05 0.903 76 0.05 0.75 0.05 1.904 332 1.35 0.75 0.15 16.705 122 0.05 0.75 0.05 3.056 430 1.55 21.95 0.10 21.607 615 2.00 44.10 0.20 44.708 1489 5.00 258.70 0.40 263.559 1000 2.35 113.90 0.15 119.65

10 730 2.55 63.50 0.35 77.7011 791 1.40 73.10 0.20 77.9012 988 1.60 114.20 0.30 117.4513 1086 1.70 136.40 0.25 140.9014 1137 2.05 151.00 0.25 155.3015 1186 2.05 162.90 0.35 168.6516 1201 2.30 166.90 0.30 177.7017 1222 2.25 171.95 0.35 179.2018 1372 2.60 217.35 0.35 230.0519 1848 3.80 409.80 0.50 418.9020 1898 3.25 429.95 0.55 443.8021 1964 3.40 450.00 0.50 471.3522 2492 4.40 726.85 0.70 738.2523 3122 6.40 1119.85 0.85 1154.6524 2625 4.45 830.90 0.60 817.4025 1199 1.75 167.25 0.25 174.0526 3498 5.75 1508.35 0.95 1411.1027 4331 7.55 2297.60 1.05 2195.05

Average 2.66 357.05 0.36 356.38



November 5th, 2013

Slide # 41


Problem Statement Problem #1: Lack of POI identification algorithms that

have a high accuracy. Problem #2: Lack of POI identification algorithms that

have a sufficiently high accuracy on a reduced GPS dataset to take into consideration the dynamic sampling and sending rates of GPS fixes, solving the limited energy resource problem on mobile devices.

Problem #4: Unable to maintain a temporal order of GPS fixes to support fast trip segmentation.

Problem #3: Slow running time of O(n2) in worst case scenario.

42Slide #


Background

Slide # 43


GPS Auto-Sleep

U.S. Patent # 8,036,679

Goal: Save battery energy while allowing real-time tracking through dynamic GPS sampling rates

Adjust the GPS sampling rate based on user movement Actively moving: High sampling rate Stationary: Low sampling rate

Slide #U.S. Patent # 8,249,807 44


Critical Points Algorithm Goal: Save battery energy and user’s budget by

sending fewer GPS fixes wirelessly Send less data over the network by filtering out GPS fixes

• Change in direction of the path• Speed at a point

Slide # 45


Sample Datasets

Dataset Using GPS Auto-Sleep Reduced Dataset Using Both GPS Auto-Sleep and Critical Points Algorithm

ASTIPI - IWGS 2013 46Slide #


Related Works

November 5th, 2013



DBSCAN

November 5th, 2013

Density-Based Spatial Clustering of Applications with Noise

Eps-Neighborhood of a point P – points within Eps distance from P Core Point if number of neighbors > MinPoints Density-Reachable and Density-Connected from P



DBSCAN – Limitations

November 5th, 2013

Running time: O(n2)

Do not use temporal properties

Loss of location data and returning trips problem Uses MinPoints to identify Core Point A POI may have only one or two GPS fixes



ST-DBSCAN

November 5th, 2013

Spatial Temporal DBSCAN

Extension of DBSCAN Discover clusters based on both spatial and temporal values Neighbors must meet both spatial condition (Eps1) and

temporal condition (Eps2)

Limitation: Treat spatial and temporal values separately

• Each point in the dataset is strongly correlated by space and time• Difficult to find Eps1 and Eps2



CB-SMoT

November 5th, 2013

Clustering-Based Stops and Moves of Trajectories Computes stops and moves Finds interesting places that are missing from the given

geographical locations A two-step algorithm:

Step 1: Identifies potential stops Step 2: Identifies unknown stops Focuses on Step 1

Extension of DBSCAN which determines Core Point by duration

Uses quantile function to obtain a relative parameter of Eps distance



CB-SMoT – Limitations

November 5th, 2013

Running time: O(n2)

Quantile function Requires a priori knowledge of the proportion between points

inside potential stops and total points in the dataset• This proportion varies based on user’s activities• Mean and standard deviation of distance among coordinates vary

when a dynamic GPS sampling rate is involved

Returning trips problem



Stay Point Detection

November 5th, 2013

Stay Point Detection Looks for a region where users spend a long period of time

Uses spatial temporal properties

Limitations: Running time: O(n2) Double error when loss of location data and GPS outliers

occur consecutively



Fast Clustering

November 5th, 2013

Fast Clustering of Global Navigation Satellite System Data

Agglomerative hierarchical clustering approach• Clusters are combined if distance is close (AVL-tree-merge)

Represents clusters by AVL trees to maintain temporal order Memory storage: O(n)

Limitations: Running time: O(n2logn) No noise detection and reduction

• Introduces "pseudo-POIs”

Slide # 54


Experimental Design GPS Auto-Sleep state values

CP Algorithm thresholds values min speed threshold = 0.1 meters per second max walk speed = 0.6 meters per second angle threshold = 4.5 degrees for walk trips and 3 degrees for

car trips.

state[0] = 4 seconds state[1] = 8 seconds state[2] = 16 seconds

state[3] = 64 seconds state[4] = 150 seconds state[5] = 256 seconds

Accuracy = TP/(TP + FP + FN)

Slide # 55


Experimental Design Use TRAC-IT Java ME app for data

collection, Sanyo Pro 200 mobile phone w/ assisted GPS

Ground truth: user carries the phone throughout a day, reporting visited places

Walking, driving, riding bus

Compare ASTIPI with CB-SMoT, SPD, and FCParameters ASTIPI CB-SMoT SPD FC

MinTime 5 minutes 5 minutes 5 minutes N/A

Eps 100 meters 100 meters 180 meters 100 meters

K 3 N/A N/A N/A

Slide # 56


Running Time Eps-K-Neighborhood Search

One for-loop: start from a given startIndex and halted after a constant K times failing to meet the close proximity condition between two coordinates

• Worst case scenario: O(n) – The main ASTIPI algorithm also stops– ASTIPI algorithm has a linear running time

• Other scenario: – Stop after a constant number of comparisons– If Core Point, ASTIPI algorithm continues with next index

Number of comparisons in ASTIPI = F(n) + C ASTIPI runs in linear time

TRB 2014 - Automatic Spatial-temporal Identification of Points of Interest in Global Navigation...

Technology

Transcript of TRB 2014 - Automatic Spatial-temporal Identification of Points of Interest in Global Navigation...