1. Danger Prediction by Case-Based Approach on Expressways
C. Y. Fang, P. Y. Wu,
S. L. Chang, and S. W. Chen
National Taiwan Normal University
Department of Computer Science and Information Engineering

2. Outline Introduction System Flowchart and Database
IEEE ITSC 2008
Outline
Introduction
System Flowchart and Database
Weighted Relational Map
The Matching Algorithm
Experimental Results
Conclusions and Future Work

3. Introduction Driving risk reduction approach
IEEE ITSC 2008
Introduction
Driving risk reduction approach
Passive approach
To reduce the degree of injury in case of an accident
Examples: seat belts and airbags
Active approach
To prevent accidents in advance
Example: driver assistance system
The dangerous driving event prediction system
To predict dangerous driving events
Based on the weighted relational map
the driving factors for the host vehicle
the driving factors for nearby vehicles
the driving factors for the roadway conditions

4. System Flowchart Driving Factors
IEEE ITSC 2008
Driving Factors
Relational Map for Driving Event Construction
Relational Map of Dangerous Case Database
Map C
Map D
Relational Map Matching
Degree of Danger
Generation
Dangerous Case Insertion
yes
Map C no
Accident Occurred? no
Dangerous?
The current driving factors can be obtained from different detection system with various kinds of sensors.
Using this driving factors, the system can construct the relational map for driving events. Called map C.
This current relational map (map C) can be used to find a most similar map from the dangerous relational maps stored in the database.
Called Map D.
Compared with map D, the system can compute the degree of danger of map C.
yes
Warning Output
System Flowchart

5. Dividing Database into Sub-Databases
IEEE ITSC 2008
Dividing Database into Sub-Databases
To speed the matching process
Dangerous case database is divided into four sub-databases based on road conditions.
Dangerous Case Database
In this paper the database will be partitioned into sub-database according to road conditions,
because different road conditions will generate different dangerous driving events.
For example, a vehicle changes its lane on an ordinary road is usually safe, but this kind of vehicle behavior is dangerous in a tunnel section.
Interchange
Section
Sub-Database
Tunnel
Section
Sub-Database
Ordinary
Section
Sub-Database
Tollbooth
Section
Sub-Database

6. Dividing Sub-Database into Classes
IEEE ITSC 2008
Dividing Sub-Database into Classes
Each sub-database is divided into six classes based on weather conditions.
Interchange
Section
Sub-Database
So these kinds of partition will speed the matching process.
Cloudy Class
Hazy Class
Snowy Class
Sunny Class
Misty Class
Rainy Class

7. Outline Introduction System Flowchart and Database
IEEE ITSC 2008
Outline
Introduction
System Flowchart and Database
Weighted Relational Map
The Matching Algorithm
Experimental Results
Conclusions and Future Work

8. Driving Factors Lateral distance Longitudinal n3 n1 n2
IEEE ITSC 2008
Driving Factors
Lateral
distance
Longitudinal n3 n1 n2
Host
vehicle
Input data for nearby vehicles
Lateral distance
Longitudinal distance
Relative lateral speed
Relative longitudinal speed
The driving factors for nearby vehicles are:
(1) the left-front vehicle and the host vehicle are close
(2) the preceding vehicle and the host vehicle are close
(3) the right-front vehicle and the host vehicle are close
(4) the left vehicle and the host vehicle are close
(5) the right vehicle and the host vehicle are close
(6) the left-rear vehicle and the host vehicle are close
(7) the following vehicle and the host vehicle are close
(8) the right-rear vehicle and the host vehicle are close
Many kinds of input data can be obtained from different detection systems with various sensors.
The input data can be divided into three classes: for nearby vehicles, for host vehicle and for the roadways.
For example, if we obtain the … and … between the host vehicle and left nearby vehicle,
Then we can check the driving factor number 4 occur or not.
In our paper we define eight driving factors for nearby vehicle list here.

9. Driving Factors Input data for host vehicle
IEEE ITSC 2008
Driving Factors
Input data for host vehicle
Lateral distance to left/right obstacle
Turning angle of front wheel
Turn signal on/off
Speed of host vehicle
Driver’s level of alertness
The driving factors for host vehicle are:
(9) the host vehicle turns left
(10) the host vehicle turns right
(11) the host vehicle speeds up
(12) the host vehicle slows down
(13) driver’s level of alertness increases
(14) driver’s level of alertness decreases
(15) the host vehicle turns on the turn signal
(16) the host vehicle maintains constant speed
(17) the host vehicle goes straight
Similar to the driving factors for nearby vehicle, we define nine driving factors for host vehicle in this paper.
the driving factors for host vehicle can be obtained from the input data for host vehicle.
For example, speed of host vehicle can be used to check the occurrence of driving factors number 11, 12, and 16.
This three driving factors are all relative to the speed of host vehicle.
In this paper we define nine driving factors for host vehicle. List here.

10. Driving Factors Driving factors for roadway:
IEEE ITSC 2008
Driving Factors
Driving factors for roadway:
(18) smooth and straight roadway
(19) smooth and curved left roadway
(20) smooth and curved right roadway
(21) downgrade and straight roadway
(22) downgrade and curved right roadway
(23) downgrade and curved left roadway
(24) upgrade and straight roadway
(25) upgrade and curved left roadway
(26) upgrade and curved right roadway
Similar to the above examples, here list the driving factors for the roadway.
In this paper, we totally list 26 driving factors.

11. A Relational Map Each node represents one driving factor.
IEEE ITSC 2008
A Relational Map
Each node represents one driving factor.
Node 18 : smooth and straight roadway
Node 16 : the host vehicle maintains constant speed
Node 1 : the left-front vehicle and the host vehicle are close
Node 11 : the host vehicle speeds up
Node 9 : the host vehicle turns left
Two requirements to generate new nodes
Fixed sampling interval
Any driving factors occurring between samples 18 1 11 16 9 T
Here we introduce how to construct a relational map.
Here is an example of a relational map.
If each node represents on driving factor introduce above, then
a relational map can be constructed by linking the driving factors based on time order.
Every fixed sampling interval, the system will general new nodes.
And if any driving factors occurring between samples, the system will general new nodes.

12. Weighted Relational Map
IEEE ITSC 2008
Weighted Relational Map 18 1 11 16 9
0.5
0.7
0.9
0.4
0.8 T
Node value
Node number
Link weight
Relational map
Weighted Relational map
We need to extra define two kind of weights in the weighted relational map.
One is the node value, indicate the importance of the driving factor.
The other is the link weight between adjacent nodes, indicate the degree of relationship between two adjacent nodes.

13. The Node Value The node value (the importance of node)
IEEE ITSC 2008
The Node Value
The node value (the importance of node)
Initialized with a constant
Increased or decreased based on the relationships with the previous, present and following nodes
Examples of increasing node values
The left-front vehicle and the host vehicle are close at time t -1, and the host vehicle turns left at time t.
The host vehicle speeds up at time t -1, and slows down at time t.
Examples of decreasing node values
The host vehicle turns on the turn signal at time t -1, and turns left at time t.
The host vehicle turns on the turn signal at time t -1, and turns right at time t.

14. The Weight Between Adjacent Nodes
IEEE ITSC 2008
The Weight Between Adjacent Nodes
The link weight:
: value of driving factor at
: value of driving factor at　
α : a constant
∆t : the time between successive driving factors
The weight is large if and are very different.
Example: vehicle changes its speed
The weight will be large if delta t is small and the difference between l t and l t-1 is large.
For example, when a vehicle change its speed between two successive driving factors,
If the speed change is large, then the weight is high.
The larger the speed change, the higher the weight is.
改變量越大 改變的時間越短 值會越大
說出此式子的優點…. 煞車 方向燈

15. Outline Introduction System Flowchart and Database
IEEE ITSC 2008
Outline
Introduction
System Flowchart and Database
Weighted Relational Map
The Matching Algorithm
Experimental Results
Conclusions and Future Work

16. Defining the Driving Factor Sets for Each Node
IEEE ITSC 2008
Defining the Driving Factor Sets for Each Node
Three driving factor sets for :
previous, present and following sets
: set for node at time 18 1 11 16 9
0.5
0.7
0.9
0.4
0.8 T
Weighted Relational map
Before discussing the matching algorithm, we first define three sets of node ni.
For example, here is a weighted relational map.
Let the node ni be the node 16.
Then the previous set of node 16 is the set including one node, node 18.
Since there is no other node on the same layer with node 16, the present set is empty.
Finally the following set of node 16 is the set including node 1 and node 11.

17. Relationship Between Adjacent Nodes
IEEE ITSC 2008
Relationship Between Adjacent Nodes
Node value
This equation defines the relationship between two adjacent nodes.
For example, the relationship value will be large if …,…, and ,,, these three values are all large.
Link weight T

18. Relationship Between the Nodes on Same Layer
IEEE ITSC 2008
Relationship Between the Nodes on Same Layer
Node value
This equation defines the relationship between the nodes on the same layer.
We find the minimum node value of these nodes as the degree of the relationship between the nodes on same layer. T

19. Table from Weighted Relational Map
IEEE ITSC 2008
Table from Weighted Relational Map
Set
null
0.9/16
0.5/18
0.318/1,0.182/11
0.4/11
0.9/1
0.4/16
0.4/1
0.444/1,0.356/11
0.4/9,0.4/16
end
0.9/9
0.9
0.9
Now we can transform the weighted relational map into a table like this one.
For example, the first column of three row indicates the node ni,t is node 16.
The second column indicates the previous set of node 16 includes a node, node 18.
And the number 0.5 is the degree of relationship between node 16 and node 18.
The three column indicates the present set of the same layer. It is empty here.
The forth column indicates the following set of node 16 is nodes 1 and 11.
And the number is the degree of relationship between node 16 and node 1.
0.5
0.7 1
0.5
0.9
0.8
0.7 9
0.7
Weighted Relational map 18 16
0.4 1
0.9
0.5 11
0.8
0.7 16 T

20. IEEE ITSC 2008
Matching Algorithm
Map C is the current weighted relational map formed in real time, and is the driving factor in C.
Map D is the dangerous weighted relational map in the database, is the driving factor in D.
: the similarity between two maps
N : the number of driving factors in C
: the similarity between two driving factors.
The matching algorithm is used to match the current map C and one dangerous case of the database, map D.
The matching algorithm calculates the similarity between map C and map D to measure the matching degree of these two maps.

21. Matching Algorithm where | | : scalar cardinality : fuzzy intersection
IEEE ITSC 2008
Matching Algorithm
where | | : scalar cardinality
: fuzzy intersection
: fuzzy union
The similarity function consists of the summation of f functions.
And the f function consists of the scalar cardinalities of the fuzzy intersection and fuzzy union.

22. Matching Algorithm and are constants.
IEEE ITSC 2008
Matching Algorithm
　　 : fuzzy driving factor set for node at time in the map C.
　　 : fuzzy driving factor set for node at time in the map D.
T( ) is the weighting function.
Here shows the relative equations of fuzzy intersection.
You can see other equations in the paper.
and are constants.

23. Outline Introduction System Flowchart and Database
IEEE ITSC 2008
Outline
Introduction
System Flowchart and Database
Weighted Relational Map
The Matching Algorithm
Experimental Results
Conclusions and Future Work

24. IEEE ITSC 2008
Experimental Results
0.9
0.5
0.5
0.833
0.182 11 1 11
0.2
0.5
0.9
0.833
0.182
Map C 18 1
0.957
0.6
0.5
0.444 9 14 21
0.6
0.5
0.9
0.5
0.5
Sim(C,D)=1
0.9
0.5
0.5
0.833
0.182 11 1 11
0.2
0.5
0.9
There is no doubt, if the map C is equal to map D then the similarity is 1.
0.833
0.182 18 1
0.957
0.6
0.5
Map D
0.444 9 14 21
0.6
0.5
0.9
0.5
0.5 T
2.0
1.5
1.0
0.5
0.0

25. Example (1) 18 1 11 9 0.5 0.9 0.957 0.2 0.444 Sim(C3,D) =0.66 Map C3
IEEE ITSC 2008 18 1 11 9
0.5
0.9
0.957
0.2
0.444
Sim(C3,D) =0.66
Map C3 14
0.833
0.6
Sim (C4,D) =0.74
Map C4
Example (1)
Sim(C1,D) =0.48 18
0.5 1
0.9
0.957
Sim(C2,D) =0.65
Map C2
Map C1 18 1 14 11 9 21
0.0
0.5
1.0
1.5
2.0 T
0.9
0.957
0.2
0.444
0.833
0.6
0.182
Map D
The first example shows when the map C is growing as time goes on,
The similarity between map C and map D will increase too.

26. IEEE ITSC 2008
Example (1)
0.9
0.5
0.5
0.833
0.182 11 1 11
0.2
0.5
0.9
0.833
0.182
Map C5 18 1
0.957
0.6
0.5
0.444 9 14 21
0.6
0.5
0.9
0.5
0.5
Sim(C5,D)=1 18 1 14 11 9 21
0.0
0.5
1.0
1.5
2.0 T
0.9
0.957
0.2
0.444
0.833
0.6
0.182
Map D

27. IEEE ITSC 2008
Example (2)
Map D1 18 1 11 9
0.5
0.9
0.957
0.2
0.444
0.9
0.5
Sim(C,D1)= 0.508
0.833 11 1
0.5
0.9
0.2
0.833 18 1
Map C
0.957
0.6
0.444 9 14
0.6
0.9
0.5
Sim(C,D2)= 0.743
Map D2 18 1 14 11 9 21
0.5
0.9
0.957
0.2
0.444
0.833
0.6
0.182
This example shows the case that there are two similar maps stored in the database.
One map is a submap of another.
For example, the map D1 is a submap of map D2.
Moreover, the map D1 is the submap of map C, and the map C is the sebmap of map D2.
Since the current map map C dose not occur the dangerous case represented by map D1, then the similarity
between map D1 and map C is smaller than that between map D2 and C. T
1.5
1.0
0.5
0.0

28. Example (3) T Sim(C,D2)= 0.938 Map D2 Map C 18 1 14 11 9 21 0.5 0.9
IEEE ITSC 2008
Example (3) T
Sim(C,D2)= 0.938
Map D2
Map C 18 1 14 11 9 21
0.5
0.9
0.348
0.25
0.222
0.842
0.2
0.4
0.211
0.118
Map D1
0.316
Sim(C,D1)= 0.967
This example shows the case that the nodes are the same but the link weights are different.
From this example, we know that the link weights of map C is more similar to those of maps D1.
Thus the similarity between map D1 and C is larger than that between map D2 and C.
2.0
1.5
1.0
0.5
0.0

29. IEEE ITSC 2008
Example (4) T 18 1 14 11 9 21
0.5
0.9
0.957
0.2
0.444
0.833
0.6
0.182
0.0
1.0
1.5
2.0 22 3 20 10
0.4
0.125
Sim(C,D2)= 0.081
Map C
Hit with right-front vehicle
Map D2
Hit with left-front vehicle 6 12
Map D1
Hit with left-rear vehicle
Sim(C,D1)= 0
When the current map C is very different to the maps stored in the dangerous case database.
Then each similarity is very low.

30. Conclusions and Future Work
IEEE ITSC 2008
Conclusions and Future Work
The proposed system
Predicting dangerous driving events based on the weighted relational map which is constructed by the driving factors
Using fuzzy matching algorithm to get the similarity between two weighted relational maps
Future Work
Improving the method to experimental threshold of level of danger
Hoping test vehicles could equip with the prototype system in the future

31. Thank you for your attention!
IEEE ITSC 2008
Thank you for your attention!