Real-time 3D point cloud obstacle discriminator apparatus and associated methodology for training a classifier via bootstrapping

• US 8,605,998 B2
• Filed: 05/06/2011
• Issued: 12/10/2013
• Est. Priority Date: 05/06/2011
• Status: Active Grant

• Current Assignee: Toyota Jidosha Kabushiki Kaisha (Toyota Motor Corporation)
• Original Assignee: Toyota Motor Engineering & Manufacturing North America Incorporated (Toyota Motor Corporation)
• Inventors: Samples, Michael Edward, James, Michael Robert
• Primary Examiner(s): Bali, Vikkram
• Application Number:
  US13/102,801
• Publication Number:
  US 20120281907A1
• Time in Patent Office: 949 Days
• Field of Search: None
• US Class Current: 382/159
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 20/58 Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
• CPC Class Codes: G06V 20/64 Three-dimensional objects

Abstract

Training a strong classifier by classifying point cloud data with a first classifier, inferring a first terrain map from the classified point cloud data, reclassifying the point cloud data with the first classifier based on the first terrain map, and training a second classifier based on the point cloud data reclassified with the first classifier based on the terrain map. The point cloud data is then classified with the second classifier, and the procedure followed with the first classifier is iteratively repeated until a strong classifier is determined. A strong classifier is determined when a probability of a terrain map matching a given terrain for the strong classifier is approximately equal to a probability of a terrain map matching the given terrain for a prior trained classifier.

Claims

1. A controller for training a classifier, the controller comprising a processor configured to:

classify point cloud data with a first classifier;

infer a first terrain map from the classified point cloud data;

reclassify the point cloud data with the first classifier based on the first terrain map;

train a second classifier based on the point cloud data reclassified with the first classifier based on the first terrain map;

compare a first classification result of the first classifier with a second classification result of the second classifier;

determine whether the first result is sufficiently close to the second result; and

in response to determining the first and second results are not sufficiently close, classify the point cloud data with the second classifier and infer a second terrain map from the point cloud data classified by the second classifier.

2. The controller according to claim 1, the processor further configured to:

reclassify the point could data with the second classifier based on the second terrain map.

3. The controller according to claim 2, the processor further configured to:

train a third classifier based on the point cloud data reclassified with the second classifier based on the second terrain map.

4. The controller according to claim 3, the processor further configured to:

compare the second classification result of the second classifier with a third classification result of the third classifier;

determine whether the second result is sufficiently close to the third result; and

in response to determining the second and third results are not sufficiently close, classify the point cloud data with the third classifier and infer a third terrain map from the point cloud data classified by the third classifier.

5. The controller according to claim 4, the processor further configured to:

in response to determining the second and third results are sufficiently close, utilize the third classier to classify newly received point cloud data.

6. The controller according to claim 4, the processor further configured to:

iteratively train classifiers until a classification result of an Nth classifier is sufficiently close to a classification result of an Nth−

1 classifier; and

utilize the Nth classier to classify newly received point cloud data.

7. The controller according to claim 1, the processor further configured to:

in response to determining the first and second results are sufficiently close, utilize the second classier to classify newly received point cloud data.

8. The controller according to claim 1, the processor further configured to:

infer the first and second terrain maps using a conditional random field (CRF) model.

9. The controller according to claim 8, the processor further configured to:

determine first and second probabilities of the first and second terrain maps, respectively, matching a given terrain, wherein the processor compares the first and second results based on the first and second probabilities.

10. The controller according to claim 9, wherein the processor determines the first and second results are not sufficiently close when the first and second probabilities are not approximately equal.

11. The controller according to claim 9, wherein the processor determines the first and second results are not sufficiently close when the first and second probabilities are not within a predetermine range of each other.

12. The controller according to claim 8, wherein the CRF model is configured to smooth terrain and match labeled data.

13. The controller according to claim 1, the controller further comprising:

a first network interface controller configured to establish a communication link between the processor and a sensor which acquires the point cloud data, the first network controller configured to receive the point cloud data from the sensor.

14. The controller according to claim 13, further comprising:

the sensor, wherein the sensor includes a light detection and ranging (LIDAR) scanning device and a second network controller configured to communicate with the first network controller.

15. The controller according to claim 1, wherein the first and second classifiers are configured to classify data points of the point cloud data as an obstacle or a non-obstacle.

16. The controller according to claim 1, wherein the first classifier is manually constructed and the second classifier is constructed by training performed by the processor.

17. A vehicle including the controller defined in claim 1.

18. A method of training a classifier with a processor, the method comprising:

classifying point cloud data with a first classifier;

inferring a first terrain map from the classified point cloud data;

reclassifying the point cloud data with the first classifier based on the first terrain map;

training a second classifier based on the point cloud data reclassified with the first classifier based on the first terrain map;

comparing a first classification result of the first classifier with a second classification result of the second classifier;

determining whether the first result is sufficiently close to the second result; and

in response to determining the first and second results are not sufficiently close, classifying the point cloud data with the second classifier and inferring a second terrain map from the point cloud data classified by the second classifier.

19. A non-transitory recording medium storing instructions that when executed by a processor cause the processor to perform a method of training a classifier, the method comprising:

classifying point cloud data with a first classifier;

inferring a first terrain map from the classified point cloud data;

reclassifying the point cloud data with the first classifier based on the first terrain map;

training a second classifier based on the point cloud data reclassified with the first classifier based on the first terrain map;

comparing a first classification result of the first classifier with a second classification result of the second classifier;

determining whether the first result is sufficiently close to the second result; and

in response to determining the first and second results are not sufficiently close, classifying the point cloud data with the second classifier and inferring a second terrain map from the point cloud data classified by the second classifier.