An Interactive-Voting Based Map Matching Algorithm Jing Yuan 1 , Yu Zheng 2 , Chengyang Zhang 3 , Xing Xie 2 and Guangzhong Sun 1 1 University of Science and Technology of China 2 Microsoft Research Asia 3 University of North Texas
Feb 24, 2016
An Interactive-Voting Based Map Matching Algorithm
Jing Yuan1, Yu Zheng2, Chengyang Zhang3, Xing Xie2 and Guangzhong Sun1
1University of Science and Technology of China2Microsoft Research Asia
3University of North Texas
Outline
• Introduction• Our Contributions• Related Work• Interactive-Voting Algorithm• Evaluation• Conclusion and Future Work
Introduction
• Popular GPS-enabled devices enable us to collect large amount of GPS trajectory data
Introduction
• These data are often not precise–Measurement error: caused by limitation of
devices– Sampling error: uncertainty introduced by
sampling– It is desirable to match GPS points with road
segments on the map
Introduction
• In practice there exists large amount of low-sampling-rate GPS trajectories
Distribution of sampling intervals of Beijing taxi dataset
0~1 minutes34%
1~2 minutes8%
2~6 minutes86%
6~20 minutes14%
2~20 minutes58%
Outline
• Introduction• Our Contributions• Related Work• Interactive-Voting Algorithm• Evaluation• Conclusion and Future Work
Our Contributions
• We study the interactive influence of the GPS points and propose a novel voting-based IVMM algorithm
• Extensive experiments are conducted on real datasets
• The evaluation results demonstrate the effectiveness and efficiency of our approach for map-matching of low-sampling rate GPS trajectories
Outline
• Introduction• Our Contributions• Related Work• Interactive-Voting Algorithm• Evaluation• Conclusion and Future Work
Related Work
• Information utilized in the input data– Geometric, topological, probabilistic, …– Usually performs poor for low-sampling rate
trajectories• Range of sampling points considered– Incremental/Local algorithms– Global algorithms
A screen shot of ST-Matching result (green pushpins are the matched points of the red trace)
Related Work
• Sampling density of the tracking data– Dense-sampling-rate approach– Low-sampling-rate approach
A screen shot of ST-Matching result (green pushpins are the matched points of the red trace)
Related Work
• Problem with ST-Matching– The similarity function only considers two
adjacent candidate points– The influence of points is not weighted– The mutual influence is not considered
Outline
• Introduction• Our Contributions• Related Work• Interactive-Voting Algorithm• Evaluation• Conclusion and Future Work
Problem Definition
• Given a low-sampling rate GPS trajectory T and a road network G(V,E), find the path P from G that matches T with its real path.
Key Insights
• Position context influence
• Mutual influence• Weighted
influence a
b c d e
f
System Overview
Candidate Road Segments / Points
Range Query
Spatial Analysis
Candidate Graph
Static Score Matrix Building Find SequenceRoad Network
I. Candidates Preparation II. Position Context Analysis III. Mutual Influence Modeling IV. Interactive Voting
Raw GPS data
Temporal Analysis Weighted Influence Modeling
Weighted Score Matrix
Parallel Voting
Matched Road Segments
Step 1: Candidate Preparation
• Candidate Road Segments (CRS) • Candidate Points (CP)
• Candidate Graph G’=(V’,E’)
𝑒𝑖3
𝑒𝑖1
𝑒𝑖2
𝑐𝑖3
𝑝𝑖
𝑐𝑖2
𝑐𝑖1
r
11c
21c
31c
12c
22c
13c
23c
14c
24c
34c
1p
2p3p
4p11e
21e
31e
12e
22e
13e
23e 1
4e
24e
34e
p1's candidates p2's candidates p3's candidates p4's candidates
11c
21c
31c
12c
22c
13c
23c
14c
24c
34c
Step 2: Position Context Analysis
• Spatial Analysis– Measure the similarity between the candidate paths
with the shortest path of two adjacent candidate points
11
1, ( , )
.t s i ii i
i t i s
dV c c
w
'1, ( , )
11 2' 2
1, ( , )1 1
( . )
( . )
ki t i sut s u
t i ik k
i t i suu u
e v vF c c
e v v
1 1 1t s t s t si i s i i t i iF c c F c c F c c
p1's candidates p2's candidates p3's candidates p4's candidates
11c
21c
31c
12c
22c
13c
23c
14c
24c
34c
Step 2: Position Context Analysis
• Spatial Analysis
11
1, ( , )
.t s i ii i
i t i s
dV c c
w
1 1 1t s t s t si i s i i t i iF c c F c c F c c
Step 2: Position Context Analysis
• Temporal Analysis– Considers the speed constraints of the road segment
• Spatial Temporal Function
11
1, ( , )
.t s i ii i
i t i s
dV c c
w
1 1 1t s t s t si i s i i t i iF c c F c c F c c
Step 3: Mutual Influence Modeling
• Static Score Matrix– represents the probability of candidate points to be
correct when only considering two consecutive points– e.g.
1 2 1 1
2 31 1 1 1
, , , , , 2,3,...
, , ,
i i ni i i i i
n
diag w w w w w i n
w w w
iW
W ( ( , )) 1,2,...ji i jw f dist p p j n
2 3, , , 1,2,3,...diag i n ni i i i iΦ W M Φ Φ Φ
Step 3: Mutual Influence Modeling
• Distance Weight Matrix– a (n-1) dimensional diagonal matrix for each sampling
point– The value of each element is determined by a distance-
based function f– e.g.
w1=diag{1/2,1/4,1/8} ( ( , )) 1,2,...ji i jw f dist p p j n
2 3, , , 1,2,3,...diag i n ni i i i iΦ W M Φ Φ Φ
Step 3: Mutual Influence Modeling
• Weighted Score Matrix– probability when remote points are also considered– e.g.
1 2 1 1
2 31 1 1 1
, , , , , 2,3,...
, , ,
i i ni i i i i
n
diag w w w w w i n
w w w
iW
W ( ( , )) 1,2,...ji i jw f dist p p j n
2 3, , , 1,2,3,...diag i n ni i i i iΦ W M Φ Φ Φ
Step 4: Interactive Voting
• Interactive Voting Scheme– Each candidate point determines an optimal path
based on weighted score matrix– Each point on the best path gets a vote from that
candidate point– The points with most votes are selected– Can be processed in parallel
Step 4: Interactive Voting
• Find optimal path for one candidate point– The path with largest weighted score summation– Dynamic programming– A value is obtained to break the tie of voting
Step 4: Interactive Voting
• Find Optimal Path
• Voting results
• Matching result
Outline
• Introduction• Our Contributions• Related Work• Interactive-Voting Algorithm• Evaluation• Conclusion and Future Work
Evaluation
• Dataset– Beijing road network– 26 GPS traces from Geolife System
• Evaluation approach (Correct Matching Percentage)
0
2
4
6
8
10
12
0~50 50~100 100~200 200~450
Cou
nts
Number of Sampling Points
012345678
0~10 10~20 20~30 30~40 40~50 50~60 60~
Cou
nts
Average Vehicle Speed (km/h)
CMP = Correct matched pointsNumber of points to be matched× 100%
Evaluation Results
• Visualized results
IVMM
IVMM
ST
ST
Evaluation Results
• Accuracy
50
55
60
65
70
75
80
85
0.5 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5 9.5 10.5
Cor
rect
Mat
chin
g Pe
rcen
tage
(%)
Sampling Interval (minute)
ST-Matching
IVMM(β=7km)
Evaluation Results
• Running time
0 50 100 150 200
0.51.52.53.54.55.56.57.58.59.5
10.5
Running Time(s)
Sam
plin
g In
terv
al (m
inut
e)
IVMM
ST-Matching
Evaluation Results
• Impact of different distance weight functions
60
62
64
66
68
70
72
2.5 4.5 6.5 8.5 10.5
Cor
rect
mat
chin
g pe
rcen
tage
(%)
Sampling Interval (minute)
IVMM(β=10)
IVMM (exponential)
IVMM (none)
IVMM (linear)
Outline
• Introduction• Our Contributions• Related Work• Interactive-Voting Algorithm• Evaluation• Conclusion and Future Work
Conclusion and Future Work
• Conclusion– Modeling the mutual influence of the GPS sampling points – A voting-based approach for map matching low-sampling-rate GPS
traces– Evaluation with real world GPS traces
• Future Work– The mutual influence related with the topology of the road network– Combination with other statistical methods, e.g., HMM and CRF
models
Thank You!