Computational Transportation Science

Computational Transportation Science

Ouri Wolfson

Computer Science

Vision

• Take advantage of advances in – Wireless communication (communicate)– Mobile/static Sensor technologies (integrate)– Geospatial-temporal information management (analyze)

• To address transportation problems– Congestion– Safety – Mobility– Energy – Environmental

• Funded by the National Science Foundation ($3M+)• Train about 20 Scientists

– Will develop novel classes of applications• Colleges: engineering, business, urban planning• $30K/year stipend, international internships

TransportationInformation Technology

IGERT Ph.D. program in Computational Transportation Science

Outline• Abstraction of concepts from sensor data:

extracting semantic locations from GPS traces. • Coping with imprecision and uncertainty: map matching. • Mixed environments: information in vehicular and other

peer-to-peer networks. • Managing spatial-temporal data: compression.• Software tools: Databases with

– spatial, – temporal, – uncertainty

capabilities for – Tracking,– analysis, – routing;

Introduction – location information

• Location information– Physical location

• Provided by positioning systems– GPS: (122.39, 239.11, 11:20am)

• Unreadable by users

– Semantic location• Not directly provided by positioning systems

– Dominick’s grocery store, 1340 S. Canal St.– Dermatologist’s office– Home

• Useful to users

4030selpc2

There are two kinds of location information: physical lcoation and semantic location.Physical location is provided by positioning sytems. For example, GPS provide the physical location and time information with the format: (122.39, 239.11, 11:20am). It is unreadable to users.Semantic location gives useful information to user's, such as .... However, it is not provided by positioning systems.

Introduction – problem statement

• Physical location -> semantic location • Devices

– Outdoor positioning systems– Internet access

• Application examples: – context awareness of mobile devices

(autocomplete)– Reminder applications– “Total Recall” by Gordon Bell

Ouri Wolfson

This paper extracts the semantic locations from physical locations for the places that the user stays.Our method can be used in devices with an outdoor positioning system and internet access. Some cell phones and PDAs can be such devices.

Main Input and Output• Input: Trajectory: T =(x1, y1, t1), (x2, y2, t2), …,

(xn, yn, tn)• Output 1: Semantic location

– Location name (BestBuy)– Semantic category

• Business type (electronics store), • office • home

– Street address

• Output 2: Semantic location log file– (date, begin_time, end_time, semantic location)

Trajectory is the data we get from outdoor positioning systems. It is a sequence of the triple (x, y, t). (x, y) is the postion, and t is the correspondent time.Our semantic location has three attributes: location name, semantic category and street address. Semantic category is business types (eg. groecry stores), office, home, relative or friend's home, etc. Semantic log file is the log file for semantic locations. The format is (date, begin_time, end_time and the correspondent semantic location). With semantic log file, a user can know where he was. It is a good memory aid.

Online and offline versions

• Online: determine the current location– On mobile device– Based on incomplete trip trajectory

• Offline: Determine multiple past locations – Based on complete trip trajectory

Auxiliary inputs

• Profile– Calendar – (event date, semantic location)– Address Book – (phone number, semantic location)– Phone Call List – (calling date, semantic location)– Web Page List - (visiting date, semantic location)– Destination List – (searching date, address)– User’s Feedback

• Confirmed list• Denied list

In profile, we record some static context of the user. Our profile include these items. Calendar records the event and the correspondent semantic location. Our calendar is similar with the calendar in a PDA, or PC. In the address book, we have phone numbers and the correspondent semantic location.In phone call list, we record the semantic location related to the telephone numbers of the user's outgoing and incoming calls, and the calling dates.Web Page List, recording the semantic locations about links that were visited on the Internet, and the visiting dates.Destination List, recording the semantic locations about the destinations whose driving directions are searched on the Internet, and the searching dates.User's Feedback, indicating whether the predicted semantic locations are correct (Confirmed List) or not (Denied List).The profile can be build automatically. With a phone number, location name, or street address, we can get the semantic location from the Address Book or yellow pages.

Algorithm

GPS dataStep1:Extract stays

Step5: Decide thesemantic location

Yellow pages

Step2: Get streetaddress candidates

Map

Step3: Get semanticlocation candidates

Profile

Step4.3Calculate

profile utility

Step4.2Calculate SA

utility

Step4.1CalculateSC utility

User confirmation

SemanticLog file

First, let me introduce the architecture of our method.In the first step, we extract the stay from user's trajectory. Then we get the street address candidates for that stay by map. In step 3, from street addresses, we get the semantic lcoation candidates with the help of profile and yellow pages.To determine the sematic location, we calculate three utilities for each semantic location in the semantic location candidates. In step5, we use the utilities to decide the semantic location for the stay, and write it to the semantic log file. With user's confirmation, we update the semantic log file and the profile.

Juhong Liu, Ouri Wolfson, Huabei Yin, UIC

11 04/21/23

Step1 - Stay extraction • Stay

– Loss of GPS signal– To spend at least min_time in an area with the

diameter no larger than d.

• (stay_position, date, stay_start, stay_end)

This is the definition of Stay: to spend at least min_time in an area with the diameter no larger than d.There are four attributes for a stay: stay_postion, date, the time when the stay starts and ends.

Juhong Liu, Ouri Wolfson, Huabei Yin, UIC

12 04/21/23

Step2 – Street address candidates

• Reverse Geocoding– Physical location

(stay_position) -> street address

• Traditional geocoding method– Nearest street address– Incorrect result

Street address candidates: the street addresses within k meters (graph distance) from stay_position.

Bui l di ng B

850 S. Hal sted St

E

The process to retrieve the street address from a physical location is called reverse geocoding. A reverse geocoding method returns the nearest street address for the physical location.However, it often generates the incorrect result. For example, this figure illustrates a street block in a map, a building B and its entrance E. The traditional reverse geocoding algorithm returns “722 W. Taylor St” as the street address of E because E is closest to W. Taylor street. However, actually, the address of Building B is “850 S. Halsted St”.

Step3-semantic location candidates

• Street address candidates ->

semantic location candidates– Yellow pages

• Such as switchboard.com

– Profile• Calendar, Address Book, Phone Call List, Web

Page List, Destination List, User's Feedback

With a street address, yellow pages usually provide the correspondent business name(s), business type(s) (so-called semantic category). Thus for a street address, we get the semantic location triple (loc_name, semantic_category, street_address).In addition, we extract the semantic locations, which match the street address, from the profile, and add them to SLC as candidates.

At end of step 3: A set of Semantic Location candidates

• Semantic location– Location name (BestBuy)– Semantic category

• Business type (electronics store; theater), • office • home

– Street address

Step4- three utilities calculation

• For each semantic location SL in set of candidates compute:– Semantic category (SC) utility: likelihood of

semantic category, given semantic log (history)– Street address (SA) utility: likelihood the street

address, given the stay location– Profile (P) utility: Likelihood of SL, given profile P

To decide the semantic location from the sematic location candidates, we calculate three unitlities for each semantic location SL in semantic location candidates SLC.They are Semantic category utility or SC utility, Street address utility or SA utility, and Profile utility



peer-to-peer networks. • Spatial-temporal data: compression.• Software tools: Databases with



Problem

• Most information systems are client/server

• Nearby mobile devices are inaccessible– Parking slot info– Video of road construction– Malfunctioning brakelight– Taxi cab– Ride-share opportunity

Environment

A central server does not necessarily exist

Local database

Local query

“Floating database”Resources of interest in a limited geographic area possibly for short time durationApplications coexist

resource-query C

resource-query Bresource 4resource 5

resource 8

resource-query Aresource 1resource 2resource 3

Pda’s, cell-phones, sensors, hotspots, vehicles, with short-rangewireless

Short-range wireless networks wi-fi (100-200 meters) bluetooth (2-10, popular) zigbee

Unlicensed spectrum (free)

High bandwidth

Bandwidth-Power/search tradeoff

Ouri Wolfson

So the environment consists of wireless mobile devices, cell phones, pda's, laptops, sensors. some of which have a local database of local information, i.e. information of interest in a local area, possibly for a short time durationand some may search for local information

Mobile Local Search: applications

• social networking (wearable website) – Personal profile of interest at a convention– Singles matchmaking– Games – Reminder

• mobile advertising (coupons, rfid-tag info)– Sale on an item of interest at mall– Music-file exchange

• Transportation • emergency response

– Search for victims in a rubble • military

– Sighting of insurgent in downtown Mosul in last hour• asset management and tracking

– Sensors on containers exchange security information => remote checkpoints

• mobile collaborative work• tourist and location-based-services

– Closest ATM

How to enable Mobile P2P applications?

• Develop a platform for building them

Problems in data management

• Query processing

• Dissemination analysis

• Participation incentives

Floating (Probe) car data

A Segment of the road network

・・・

Periodically the ITA on a vehicle generates a velocity report:

Vehicle id IL391645 Average speed 45mph Time 3:49:45pm Location (12345.25, 4321.52) Travel direction east

P2P method

1

4

25

36

BA

C

1

4

1

2

B

A

C

(a) (b)

4

5

1

3

4

6

1

4

25

36

BA

C

1

4

1

2

B

A

C

(a) (b)

4

5

1

3

4

6

Each vehicle communicates reports to other vehiclesusing short-range (e.g. 300 meters), unlicensed, wireless spectrum, e.g. 802.11

Travel-time map

Multimedia info: view/hear traffic conditions 1 mile ahead by a click

on your smartphone.

Query Processing StrategiesWiMaC paradigm: WiFi-disseminate,

Match

Wifi/cellular-respond

media media Q

(b) Z sends Q to M-producer via cellular

Z

Z

(c) M-producer sends media to Q-producer via cellular

(a) media and Q are initially disseminated. They collocate at Z.

Z

Q

media

M-producer Q-producer



WiMaC Design Space

Evaluation criteria:• Throughput• Response time• Wi-Fi communication volume• Cellular communication volume

Comparison Results

3a (query)-WiFi

2a (meta)-WiFi

1 (media)

5b (media,query)-cell 3b (query)-cell

7b (media,meta,query)-cell

WiFi-onlystrategies

WiFi-cellularstrategies

6b (media,query)-cell

4b (media,meta)-cell 2b (meta)-cell

X Y: Strategy X dominates strategy Y

4a (media,meta)-WiFi

5a (media,query)-WiFi

7a (media,meta,query)-WiFi

6a (meta,query)-WiFi

X Y: Strategy X weakly dominates strategy Y

push-media

pull

hy-MuM-cell

hy-meta-cell

1 (media)

3a (query)-WiFi

6b (meta,query)-cell


0

2

4

6

8

10

1% 12.5% 25% 37.5% 50%

penetration ratio

an

sw

er

thro

ug

hp

ut

1 (query) 3a (query)-WiFI


dominance analysis

simulations



peer-to-peer networks. • Spatial-temporal data: compression• Software tools: Databases with



Data Compression -- Motivation

– Tracking the movements of all vehicles in the USA needs approximately 4TB/day (GPS receivers sample a point every two seconds).

Trajectory Lossy-Compression

• approximate a trajectory by another which is not farther than ε.

Desiderata for Trajectory Compression

• bounded error when answering queries on compressed trajectories.

Relational-Oriented Queries• Point queries:

– Where (T,t): where is the moving object with trajectory T at time t

– When (T,x,y): when is the moving object with trajectory T at location (x,y)

• Range queries (R,t1,t2,O): retrieve the moving objects (i.e.

trajectories) of O that are in region R between times t1 and t2.

• Nearest neighbor (t,T,O): retrieve the object of O that is closest

to trajectory T at time t

• Join queries (O,d): Retrieve the pairs of objects of O that are

within distance d.

Distance Functions• The distance functions considered

are:– E3: 3D Euclidean distance.

– E2: Euclidean distance on 2D projection of a trajectory

– Eu: the Euclidean distance of two trajectory points with same time.

– Et: It is the time distance of two trajectory points with same location or closest Euclidean distance.

• #(T'2) ≤ #(T'3) ≤ #(T'u), which is also verified by experimental saving comparison.

Soundness of Distance Functions • Soundness: bound on the error when answering spatio-temporal queries on compressed trajectories.

• The appropriate distance function depends on the type of queries expected on the database of compressed trajectories. – If all spatio-temporal queries are expected, then Eu and Et should be used.

– If only where_at, intersect, and nearest_neighbor queries are expected, then the Eu distance should be used.

Where_at

When_at

Intersect Nearest_Neighbor

Spatial Join

E2 No No No No Sound when (a) the distance

function D of join is metric

(b) E is weaker than D.

E3 No No No No

Eu Yes No Yes Yes

Et No Yes No No

Aging of Trajectories

• Increase the tolerance ε as time progresses

• Aging friendliness property: If ε1ε2 then

T’ =Comp(Comp(T, ε1 ), ε2) = Comp(T, ε2)

(associative)

Theorem: The DP algorithm is aging-friendly, whereas the optimal algorithm is not.






Matching Methods Matching Methods ---- Straightforward ---- Straightforward

SnappingSnapping

a

A

B

a b

B

A

• A, B: road segments• a, b: GPS points

• A, B: road segments• a, b: GPS points

Sony Customer

applications:insurance, road pricing, car navigation, speed limit alerts to drivers

• Compute the weight of each

road segment (block)

• Compute the shortest weight path between the start and the end GPS points as the route of the moving object

Weight-based MatchingWeight-based Matching

x

y

t

t2

t6

trajectroy

arcpolyline

p8, t8

p6, t6

p7, t7

p5, t5p4, t4

p3, t3

p2, t2p1, t1

b1, t'1

b2, t'2

b3, t'3

b4, t'4

b5, t'5

||

))()((

ij

t

t arctraj

tt

dttgtgW

j

i

Matching VariantsMatching Variants

• Offline– Find the overall route of a vehicle after the

trip is over

• Online Snapping– Real time, i.e. every 2 minutes (online

frequency)– Determine the road segment on which the

vehicle is currently located

Experiments ---- OfflineExperiments ---- Offline

• Evaluation method– Edit Distance

The smallest number of insertions, deletions, and substitutions required to change the snapped route to the correct route

– Correct matching percentage (OFFcorrect)

OFFcorrect = 100(1 – ed/n)

ResultsResults– On average, weight-based alg. is correct

up to 94% of the time, depending on the GPS sampling interval.

– It is always superior to the straightforward closest-block snapping.

– Correct matching decreases significantly when GPS sampling intervals are larger than 120 seconds






Basic element of a moving objects database: a trajectory

Y

X

Time

Present time

2d-ROUTE

3d-TRAJECTORY

Future Trajectory: Motion planPast trajectory: GPS trace

Why are traditional databases inappropriate to manage trajectories?

SELECT o

FROM MOVING-OBJECTS

WHERE Sometime/Always(10,11)

inside (o, R)

Retrieve the objects that are in R sometime/always between 10 and 11am

R

10 1110

11

sometime always

Why are traditional databases inappropriate to manage trajectories?

• Discrete vs. Continuous data

• Operators of the language that are natural in the domain

• Uncertainty

Uncertainty operators in spatial range queries

possibly and definitely semantics based onbranching time

SELECT oFROM MOVING-OBJECTS

WHERE Possibly/Definitely Inside (o, R)

Rdefinitely

possibly

uncertainty interval

Uncertain trajectory model

Possible Motion Curve (PMC) and Trajectory Volume (TV)

• PMC is a continuous function from Time to 2D

• TV is the boundary of the set of all the

PMCs (resembles a slanted cylinder)

Predicates in spatial range queries

Possibly – there exists a possible motion curve

Definitely -- for all possible motion curves

• possibly-sometime = sometime-possibly• possibly-always• always-possibly• definitely-always = always-definitely• definitely-sometime • sometime-definitely

Uncertainty in Language - Quantitative Approach

Uncertainty interval

database location

probability density function

Probabilistic Range Queries

SELECT o

FROM MOVING-OBJECTS

WHERE Inside(o, R)

R

Answer: (RWW850, 0.58)

(ACW930, 0.75)

Outline• Databases with



• compression of spatial-temporal data; • query and dissemination of (possibly multimedia)

information in vehicular and other peer-to-peer networks;

• extracting semantic locations and activity knowledge from GPS traces;

• map matching.

Adapt Uncertainty to Update frequency

• Tradeoff :

precision vs. resource-consumption• Cost based approach

(1 update = 2 units of imprecision)• Dynamic cost minimization

Information-Cost of a tripComponents:• Cost-of-location-update• Cost-of-imprecision

• Cost-of-deviation

• Cost-of-uncertainty

Current location = 15 + 5

proportional to length of period of time for which persist

14

actual location

database locationdeviation = 1

1510 20Uncertainty = 10

Ouri Wolfson

for a fixed uncertainty, the higher the deviation, the higher the error

Outline• Databases with



• compression of spatial-temporal data; • Databases in vehicular and other peer-to-peer networks; • extracting semantic locations from GPS traces; • map matching.

Example queries

• Find a multimodal route that will get me home by 7pm with 90% certainty.

• Find a route that will get me home by 7pm with 90% certainty, and

lets me stop at a grocery store for 30 minutes

Example Graph

ALL_TRIPS

ALL_TRIPS( origin-vertex, destination-vertex)

Returns a non-materialized relation of all trips (sequences of vertices) between the origin and destination

General Query Structure

SELECT *FROM ALL_TRIPS(origin, destination)WHERE

<WITH STOP VERTICES> (florist, grocery)

<WITH MODES> (Bus, boat)

<WITH CERTAINTY> (0.8)

<OPTIMIZE>) (time, distance, cost, #transfers),…)

Example Query

SELECT * FROM ALL_TRIPS(work, home) AS t WITH STOP_VERTICES v1, v2 WITH CERTAINTY .75 WHERE "pharmacy" IN v1.facilities AND "florist" IN v2.facilities AND DURATION(v1) > 10min AND DURATION(v2) > 10minAND MODES(t)contained-in {pedestrian, rail, bus} MINIMIZE number-of-transfers

With a certainty greater than or equal to .75, find a trip home from work that uses public transportation and visits a pharmacy and then a florist (spending at least 10 minutes at each) and has minimum number of transfers

Query Semantics

From the set of trips that satisfy:

– the non-temporal constraints, and – the temporal constraints with the required

certainty (remember probabilistic travel times)

Select the optimal (according to single criteria)

Semantics

Select *From All_Trips (work, home) as tWITH STOP-VERTICES v1WHERE pharmacy in v1.facilities, and modes(t) contained-in {train, bus}, and begin(t) > 8pm, and arrive(t) <10pm, and duration(v1) > 10minsWITH CERTAINTY 0.9 MINIMIZE NUMBER-OF-TRANSFERS

For each trip from work to home create a mapping from v1 to vertices of t:t1…. (t1,map1) map1: v1 -> UnionStationt1…. (t1,map2) map2: v1 -> CentralStationt2…. (t2,map1) map1: …....

For each (ti, mapj) evaluate WHERE condition and if satisfied with CERTAINTY > 0.9 put pair in RESULT.

From RESULT return the pair that MINIMIZES the number of transfers.

Evaluation of WHERE condition W on (ti,mapj)

• Evaluate non-temporal conditions and if W = ‘true’ or ‘false’ , then done.

• Otherwise split trip into legs: L1, v1, L2• L1 has departure y1 and duration z1

• L2 has departure y2 and duration z2

• y1>8pm, y2+z2<10pm, y2-y1-z1>10mins defines a region S in R4.

• Assume that we know the joint density function f(y1,z1,y2,z2).

• Then we compute the probability of W as the integral ∫S f(y1,z1,y2,z2)dy1dz1dy2dz2

Plug-and-play Query Processing

• Based on a framework– Algorithms are chosen based on the structure of the

query

SELECT *

FROM ALL_TRIPS(source, dest) AS t

WITH STOP VERTICES is empty

WHERE number-of-transfers (t) < k

OPTIMIZE is the minimization of the sum of some numeric edge attribute (e.g., length, duration)

Can be solved with

A. Lozano and G. Storchi. Shortest viable path algorithm in multimodal networks. In Transportation Research Part A: Policy and Practice, volume 35, pages 225–241, March 2001.

Conclusion• Abstraction of concepts from sensor data:


peer-to-peer networks. • Managing spatial-temporal data: compression.• Software tools: Databases with



Ongoing work

• Autonomous driving – Grand Cooperative-Driving Challenge– high precision maps

• Database platform for intellidrive applications (nsf grant)

• Competitive routing

Documents

Computational Transportation Science