Crash prediction network with graded warning for vehicle

• US 20070073463A1
• Filed: 03/22/2006
• Published: 03/29/2007
• Est. Priority Date: 03/23/2005
• Status: Active Grant

• Current Assignee: Toyota Motor Corporation, Board of Regents of the University of Texas System (University of Texas System)
• Original Assignee: Toyota Motor Corporation
• Inventors: Sherony, Rini, Miikkulainen, Risto, Stanley, Kenneth, Koh, Nathaniel
• Granted Patent: US 7,565,231 B2
• Time in Patent Office: Days
• US Class Current: 701/45
• CPC Class Codes: B60R 21/013 including means for detecting collisions, impending collisions or roll-over
• CPC Class Codes: B60W 2050/0057 Frequency analysis, spectral techniques or transforms
• CPC Class Codes: B60W 40/09 Driving style or behaviour
• CPC Class Codes: B60W 50/14 Means for informing the driver, warning the driver or prompting a driver intervention
• CPC Class Codes: G08G 1/165 for passive traffic, e.g. including static obstacles, trees

Abstract

A method for facilitating the avoidance of a vehicle collision with an object includes the following steps: a) providing a neural network, b) evolving a good driver, c) evolving a crash predictor, and d) outputting a graded warning signal.

Claims

1. A method for facilitating the avoidance of a vehicle collision with an object comprising the steps of:

a) providing a neural network;

b) evolving a good driver;

c) evolving a crash predictor;

d) outputting a graded warning signal.

2. The method of claim 1 wherein the step of evolving a good driver includes the steps of:

e) providing a biasing input;

f) generating sensory signals;

g) inputting the sensory signals and biasing input to the neural network;

h) calculating a reward value;

i) calculating a fitness value determining the retention of the neural network;

j) repeating steps e-i a plurality of times.

3. The method of claim 1 wherein the step of evolving a crash predictor includes the steps of:

j) calculating a biasing input

k) generating sensory signals;

l) inputting the sensory signals and biasing input to the neural network;

m) calculating a temporal value;

n) inputting the calculated temporal value into a queue;

o) providing an ideal temporal queue;

p) determining if a crash occurs;

q) calculating a reward value

r) calculating a fitness value determining the retention of the neural network;

s) repeating steps k-r a plurality of times.

4. The method of claim 2 wherein the biasing input is 1.

5. The method of claim 2 wherein the sensory signals are rangefinder data signals.

6. The method of claim 5 wherein the rangefinder data signals are converted from simulated driving data.

7. The method of claim 2 wherein the fitness is calculated in proportion to a distance traveled and a damage incurred over time spent off the road.

8. The method of claim 2 wherein the steps e-h are repeated 1000 times for a single trial.

10. The method of claim 3 wherein the biasing input is 1.

11. The method of claim 3 wherein the sensory signals are rangefinder data signals.

12. The method of claim 11 wherein the rangefinder data signals are converted from simulated driving data.

13. The method of claim 3 wherein the temporal value is calculated from an output of the neural network.

14. The method of claim 3 wherein the reward value is calculated according to the formula:

R=Σ

_t²⁵₌₁25−

|q_t−

I_L|/25

Wherein q_t is a prediction in the queue and I_L is an ideal prediction where 1≦

L≦

25.

15. The method of claim 3 wherein the reward is assigned a relative value such that desirable characteristics are favored in the neural network.

16. The method of claim 3 wherein the fitness is computed by accumulating a total reward R_tot equaling the fitness of the entire evaluation including all crashes and prediction sand wherein R_{tot , is modified at each time step in one of two cases:}

1) If the car crashes, R_tot increases according to how close the actual prediction queue is to the ideal prediction queue, and

2) If the car does not crash, a small bonus is added to R_tot if the oldest prediction in the queue was that no crash would occur.

17. The method of claim 3 wherein the fitness is calculated according to the formula:

E=Σ

_t=1(O_t−

I_t)²/n

wherein E is an average error over n timesteps, I_t is an ideal prediction at time step t and O_t is a prediction output.

9. The method of claim 9 including a plurality of trials.

18. A method for facilitating the avoidance of a vehicle collision with an object comprising the steps of:

a) providing a neural network;

b) evolving a good driver including the steps of;

i) providing a biasing input;

ii) generating sensory signals;

iii) inputting the sensory signals and biasing input to the neural network;

iv) calculating a reward value;

v) calculating a fitness value determining the retention of the neural network;

vi) repeating steps e-i a plurality of times;

c) evolving a crash predictor including the steps of;

vii) calculating a biasing input

viii) generating sensory signals;

ix) inputting the sensory signals and biasing input to the neural network;

x) calculating a temporal value;

xi) inputting the calculated temporal value into a queue;

xii) providing an ideal temporal queue;

xiii) determining if a crash occurs;

xiv) calculating a reward value

xv) calculating a fitness value determining the retention of the neural network;

xvi) repeating steps k-r a plurality of times; and

d) outputting a graded warning signal.