Crash Cube: An application of Map Cube to Hotspot Discovery in Vehicle Crash Data Mark Dietz, Jesse Vig CSCI 8715 Spatial Databases University of Minnesota.

Crash Cube: An application of Map Cube to Hotspot Discovery in Vehicle Crash Data

Mark Dietz, Jesse VigCSCI 8715 Spatial Databases

University of Minnesota

Outline

Motivation Problem definition

Identifying crash “hot spots” through visualization

Key Challenges Related Work Limitations of Related Work Contribution – Crash Cube Validation

Case study: Crash Cube applied to crashes in Houston 1999

Conclusions and Future Work

Motivation

Automobile crashes kill 1.2 million annually 50 million more injured

City and road design – major factors in crashes Identifying “hot spots” benefits:

City planners and traffic engineers Insurance companies Drivers

Problem Definition

Given: Locations and times of vehicle crashes Other dimensions also possible

Find: Visualization of the data across spatial and temporal dimensions

Objective: Make hot spots easy to recognize by manual inspection

Key Challenges

What is the appropriate visualization? Shape and scale of hot spots unknown

How to filter? Some hot spots only visible with certain filter

Visualization two dimensions But the crash data is higher dimension!

Statistical detection – out of scope

Types of visualization Purely spatial

Thematic mapping Color-coded continous surface Statistically-based elliptical objects

Spatio-temporal Animation Album of maps

Related Work – Hot spot detection

Related Work – Data cube and Map Cube

Data cube – aggregate operator on N dimensions

Creates 2^N aggregations for each possible combination of dimensions

Map cube – extension of data cube to spatial data

Has been applied to: Census data Traffic data

Limitations of Related Work• Purely spatial visualizations – not temporal

• Animation – only shows most prominent hot spots

• Album of maps – limit number of visualizations – Creator must know desired visualizations

• Map Cube – not applied to crashes

Spatial onlySpatio-

temporal

Thematic mapping

Color-coded continuous

surface

Statistically-based

ellipses

AnimationAlbum of

maps

AnimationAlbum

of maps

Contribution – Crash Cube

Album of maps – more visualizations than related work

2^N visualizations instead of 2 or 3

Aggregates data on 2^N combinations

Spatial Time

Time of Day (TD) Day of Week (DW) Month of Year (MY)

Drilldown and rollup Find desired visualization easily

Key Concepts

Dimension Visualization

One non-spatial dimension Bar chart

One spatial dimension Geographic plot

One spatial and one non-spatial dimension

Album of geographic plots – one for each value of non-spatial dimension

Two non-spatial dimensions 2-D plot with each axis representing a non-spatial dimension

One spatial and two non-spatial dimensions

Album of geographic plots – one for each combination of values of the non-spatial dimension

Three non-spatial dimensions Album of matrices – one for each value of the third dimension

One spatial and three non-spatial dimensions

Album of geographic plots – one for each combination of values of the non-spatial dimension

Approach

For each combination of dimensions Create a table with

Fields for non-spatial dimensions OGIS Point representing the crash location

Value of the aggregation Create visualization of table as graph or map

Answer following types of queries What are the temporal hot spots? What are the spatial hot spots? How do crashes vary throughout the days of the

week?

Case study: crashes from Houston 1999

Validation – Spatial dimension

Plot of all crashes

Plot of all crashes aggregated in 16x16 grid

Spatial plots show hot spots in different ways

Validation – Single time dimension?

?

?

Time of Day shows rush hours and bar closing hot spots

Validation – Two time dimensions

TD and DW shows rush hours and bar closings hot spots confined to certain days

Validation – Spatial and time dimensions

Plot of all crashes on Sunday

Plot of all crashes on Friday

TD and Spatial shows different hot spots on Sunday vs. Friday

Conclusions

Crash cube aggregations can reveal hotspots in both space and time

Different cube dimensions reveal different types of hotspots

Rollup and drilldown allow user to explore dataset without prior knowledge

Album allows careful inspection of data Example: Sunday vs. Friday visualization

Future Work

Alternate aggregation functions Sum of fatalities Sum of vehicles involved in crashes Average number of fatalities per 1000

crashes Additional aggregation levels

By county By street By year By month over several years

Incorporate spatial visualizations from related work into Crash Cube framework