prediction of route and destination intent shibumon alampatta

26
Prediction of Route and Destination Intent Shibumon Alampatta (Roll No. 12CS60D02) Guided by: Prof. Arobinda Gupta 1

Upload: shibu-alampatta

Post on 29-Jun-2015

284 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Prediction of route and destination intent  shibumon alampatta

Prediction of Route and Destination Intent

Shibumon Alampatta

(Roll No. 12CS60D02)

Guided by: Prof. Arobinda Gupta

1

Page 2: Prediction of route and destination intent  shibumon alampatta

What is it About?

• Predict driver’s intent – intended route and destination

– Predict the goal and route; given current location

– Predict the route; given a goal(destination) and current location

2Image Courtesy: http://exploringthemind.com

Page 3: Prediction of route and destination intent  shibumon alampatta

Application Area

• Route Guidance in Navigation

• Improving Hybrid Fuel Economy

– 7.8% fuel economy, Research by Nissan

(Froehlich 2008)

• Intelligent Transportation System

• VANET

• Points of interest and Advertisement

3

Page 4: Prediction of route and destination intent  shibumon alampatta

4

Ability to predict something normallycomes from the Experience, Knowledgeand Analytical skill to understandPatterns

Page 5: Prediction of route and destination intent  shibumon alampatta

A Prediction Model

ModelPast

DataFuture

Prediction

Real Time

Input

5

Page 6: Prediction of route and destination intent  shibumon alampatta

Algorithmic Techniques

• Hidden Markov Model [Simmons 2006][Froehlich 2008][Nagaraj 2011]

• Artificial Neural Network [Mikluščák 2012]• Nearest Neighbour algorithm [Yu 2011]• Genetic Algorithm [Kanoh 2008]• Bayesian Classifier [Cook 2004]• Support Vector Machines [Yu 2011]• Choice set Generation Model [Prato 2009] • Pattern Matching [Hattab 2012]• Plan Recognition [Cook 2004]

6

Page 7: Prediction of route and destination intent  shibumon alampatta

Markov Model

• Captures sequential model of behavior

• Markov Property: – Future is independent of past; given present

• <S, A, T>– S : Set of States

– A : Set of Actions

– T: Transition function T: S x A x S R• T(si, a, sj) = P(si| sj, a)

• Probability of transitioning to a state si; given that the system is in state sj and action a is executed

• In some cases explicit actions may not be there

7

S(t-1) S(t) S(t+1)

Page 8: Prediction of route and destination intent  shibumon alampatta

Hidden Markov Model (HMM)

• A Markov Model with Hidden(Unobservable) States

• <S, A, O, T, Z, ∏ >

• O – Finite set of Observations

• Z – Observation function– Z : O x S x A R

– Z(o, s, a) = P(o| s, a)

– Probability of receiving observation o, given system ends up in state s on executing action a

– For many problems Z(o, s, ai) = Z(o, s, aj), so we write Z(o, s)

• ∏ - Initial state distribution

8

Hidden State

Intentions in Drivers Mind

Observation

Current Link

Page 9: Prediction of route and destination intent  shibumon alampatta

HMM

9

S(t-1) S(t) S(t+1)

s1 s2

o1 o2 o3

O(t-1) O(t) O(t+1)

Sequential Representation

State Transition Representation

Page 10: Prediction of route and destination intent  shibumon alampatta

Predicting Driver’s Intent

10

Build the Intent Prediction Model

Train the Model using Collected Trip Data

Use the Model for Prediction

Page 11: Prediction of route and destination intent  shibumon alampatta

Building the Model

• Assumption– Driving is mostly routine– Past performance can be used to predict what the

driver will do in future– Route map and a GPS is available and can compute

segment of the map the vehicle is on

• Routine nature of driving– Tend to go to same destination again and again– Tend to follow same route– Same time– Even when better alternatives exists (shorter or faster)

11

Page 12: Prediction of route and destination intent  shibumon alampatta

Building the Model

• Perfect prediction is not possible

– Example scenarios

• Conclusion: Prediction of driver intent is probabilistic

– So, we can make prediction with certain probability only

– But never be 100% sure about the prediction

• So we can use a probabilistic approach

12

Page 13: Prediction of route and destination intent  shibumon alampatta

Building the Model

• Road Graph Representation

– Model a Graph G(V, E) from the road map

– Vertices (v) for each intersection

– Link (l)– unique labeling for an edge between two intersections

13

Map Courtesy: maps.google.com

Page 14: Prediction of route and destination intent  shibumon alampatta

Building the Model

• We want to predict the intention that driver is going to have in his mind

• Based on intention in his mind he take turns

• State s = <l, g> ; l – link, g - goal

• State Transition Function T(si, sj) = p(si|sj)

14

l

g<l, g>

Image Courtesy: depositphotos.com Map Courtesy: maps.google.com

Page 15: Prediction of route and destination intent  shibumon alampatta

Building the Model

• What we can observe?

– Current link ; ie segment on the road corresponding to current location

– Observation function

Z(ol, s) = p(ol|<l, g>) which is 1 here

– Probability of current link being l given the system is in state s = <l, g>

15

Page 16: Prediction of route and destination intent  shibumon alampatta

Building the Model

• We can write the transition probability p(si|sj) as

p(<li, gi> | sj ) = p(li | sj) p(gi | li)

• Given the current state,we first predict the nextlink that the driver will go and then we predict his goal destination based on that link

16

lj

li

gi

Map Courtesy: maps.google.com

Page 17: Prediction of route and destination intent  shibumon alampatta

Building the Model – Probability Computation

17

To compute p(gi|li)

To compute p(li | sj) = p(li | <lj, gj>)

lj

li

gj

Page 18: Prediction of route and destination intent  shibumon alampatta

Building the Model – Next Link Prediction

• Possible next states

– <l1,gi>

– <l2,gi>

– <l3, gi>

• Link li which

scores maximum

is predicted as

next link

18

Map Courtesy: maps.google.com

c

Page 19: Prediction of route and destination intent  shibumon alampatta

Building the Model – Next Link Prediction

• p(si|sj) = p(<li, gi> | sj)

19

<l3, g1><l3, g2><l3, g3>

<l2, g1><l2, g2><l2, g3>

<l1, g1><l1, g2><l1, g3>

Score for l1

Score for l2

Score for l3

Page 20: Prediction of route and destination intent  shibumon alampatta

Building the Model - Prediction

• Predicting the goal/route given current link

– Using the ability to predict next link continue the prediction until we reach some goal or most probable goal

• Predicting route given goal

– Use this to bias the prediction of next link and continue prediction until we reach g

20

Page 21: Prediction of route and destination intent  shibumon alampatta

Training the Model

• Collect Trips data

• A Trip is an ordered list of links (<l1, t1>, <l2, t2>…..)

• Go through and trip sequence, fill or update the tables.

• This helps in computing the probabilities

• Training data shall be reliable

• More the data; better accuracy

• Once Training is done; Use for prediction with real-time data

21

Page 22: Prediction of route and destination intent  shibumon alampatta

Past Trip Data

22

Ability to predict something normallycomes from the Experience, Knowledgeand Analytical skill to understandPatterns

Probabilistic Model

Page 23: Prediction of route and destination intent  shibumon alampatta

Observations

• Achieves more than 80% accuracy in average– Can be harnessed for , route planning, traffic

prediction, smarter route guidance

– emergency route etc.

• We can include the parameters like time of the day, day of the week etc. to state tuple to enhance the model

• Scenarios where routine nature is not maintained (sales people or delivery boys)

• Ad-hoc predictions are difficult (Like terrorists)

23

Page 24: Prediction of route and destination intent  shibumon alampatta

Summary

• Problem Definition

• Applications and Motivation

• Various Algorithmic Techniques

• HMM based model

• Limitations

• Possible enhancements

• Extension of application domains

24

Page 25: Prediction of route and destination intent  shibumon alampatta

References

[1] Reid Simmons, Brett Browning, Yilu Zhang, Varsha Sadekar, Learning to predict driver route and destination intent, IEEE Intelligent Transportation System Conference, 2006

[2] Hitoshi Kanoh, Kenta Hara, Hybrid genetic Algorithm for Dynamic Multi Objective Route Planning with Predicted Traffic in a Real World Road Network, Proceedings of the 10th Annual Conference on Genetic and Evolutionary Computation, 2008

[3] John Froehlich, John Krumm, Route Prediction from Trip Observation, Society of Automotive Engineers (SAE) World Congress, 2008

[4] Uma Nagaraj, N N Kadam, Study of Statistical Models for Route Prediction Algorithms in VANET, Journal of Information Engineering and Applications, Vol 1, No. 4, 2011

[5] Carlo Giacomo Prato, Route Choice Modeling: Past, Present and Future Research Directions, Journal of Choice Modeling, 2(1), pp. 65-100, 2009

[6] http://en.wikipedia.org/wiki/Hidden_Markov_model[7] http://en.wikipedia.org/wiki/Markov_property[8] Diane J Cook, Prediction Algorithms for Smart Environments, Chapter 8, Smart Environments:

Technology, Protocols and Applications, John Wiley & Sons, 2004[9] M Al-Hattab, M Takruri, J Agbinya, Mobility Prediction using Pattern Matching, International

Journal of Electrical and Computer Sciences, Vol.12 No.3, 2012[10] Tomáš Mikluščák, Michal Gregor, Aleš Janota, Using Neural Networks for Route and Destination

Prediction in Intelligent Transport Systems, 12th International Conference on Transport Systems Telematics, 2012

[11] Bin Yu, William H.K. Lam, Mei Lam Tam, Bus Arrival Time Prediction at Bus Stop with Multiple Routes, Transportation Research Part C: Emerging Technologies, Volume 19 Issue 6 pp. 1157–1170, 2011

25

Page 26: Prediction of route and destination intent  shibumon alampatta

Thank You

26