Context-Aware Similarity of Trajectories Maike Buchin Somayeh Dodge Bettina Speckmann.

Context-Aware Similarity of Trajectories

Maike BuchinSomayeh Dodge Bettina Speckmann

Location sampled over time• noisy sampling• geographic context

2:402:50

3:00

3:20 3:30

3:40

3:503:554:10

4:154:20

4:25

4:304:35

Trajectory Data

3:10

Trajectory SimilarityHow similar are two trajectories?

Geographic context needs to be taken into account!

Geographic Context

• Landcover• Network • Terrain • Ambient attributes

Model: labeled polygonal subdivision

Context-Aware Similarity

How does geographic context influence similarity?

Context-Aware Similarity

Geographic context distinguishes trajectories

Context-Aware Similarity

Integrate context & spatial distance: • use time to match points • for matched points add spatial and context distance

(p,t,c)(p‘,t‘,c‘)

for matched points (p,t,c) and (p’,t’,c’) dist(p,p’) + α dist(c,c’)

2. spatial distance(e.g., Euclidean dist)

3. context distance(based on cells)

4. context scale(depends on application)

1. matching based on time (e.g., Fréchet dist)

Ingredients:

Context Distance

Compare two points based on cells they lie in

• Labels• Subdivision distance

Example: dist(p,p’) = dist(sand,grass)

Context Distance

Compare two points based on cells they lie in

• Labels• Subdivision distance

Example: dist(p,p’’) =

dist(sand,sand) = 0 or dist(p,p’’) =

min ( 2*dist(sand,water),

2*dist(sand,grass) )

Computing Similarity

Preprocessing• Compute context distance• Locate trajectory points in

subdivision• Possibly split trajectories

at subdivision boundariesAlgorithm• Adapt known algorithms

using as distance dist(pt,pt’) + α dist(ct,ct’)

2:402:50

3:00

3:20 3:30

3:40

3:503:554:104:15

4:204:25

4:304:35

C1

C2

C3

Equal time distance: straightforward

Fréchet distance: add context distance per free space cell

Experiments: Hurricane Data

Geographic context• Hurricanes are strongly

influenced by land/sea

Similarity• Prediction• Classification

Experiments: Hurricane Data

Data• Hurricanes

– North Atlantic Basin – years 1995, 2004, 2005– sampled every 6 hours– 48 in total

• Coastline data

Results of Experiments

interesting triple among the 10 most similar hurricane pairs:

• Erin & Katrina become most similar than the other pairs

• order of similarity changes: Rita & Katrina become more similar than Rita & Erin

Rita Katrina Erin2004 2005 1995

Conclusion

We can and should integrate geographic context into the analysis of trajectories!

Future Work:• more data sets• more analysis tasks