Adaptation for clear path detection using reliable local model updating

• US 8,773,535 B2
• Filed: 12/08/2010
• Issued: 07/08/2014
• Est. Priority Date: 12/08/2010
• Status: Active Grant

• Current Assignee: GM Global Technology Operations LLC (General Motors Company)
• Original Assignee: GM Global Technology Operations LLC (General Motors Company)
• Inventors: Zhang, Wende
• Primary Examiner(s): Czekaj, Dave
• Assistant Examiner(s): ABIMBOLA, KEHINDE
• Application Number:
  US12/963,404
• Publication Number:
  US 20120147189A1
• Time in Patent Office: 1,308 Days
• Field of Search: None
• US Class Current: 348/149
• CPC Class Codes: G06V 20/56 exterior to a vehicle by using sensors mounted on the vehicle

Abstract

A method and system for vehicular clear path detection using adaptive machine learning techniques including reliable local model updating. Digital camera images are segmented into patches, from which characteristic features are extracted representing attributes such as color and texture. The patch features are analyzed by a Support Vector Machine (SVM) or other machine learning classifier, which has been previously trained to recognize clear path image regions. The SVM classifier is adaptively updated using reliable local test samples, such as positive clear path samples which just passed by the vehicle. The resultant classifier, being continuously and adaptively updated with recent, reliable training samples, exhibits improved performance and accuracy in analyzing subsequent image regions or patches for the existence of a clear driving path.

Claims

1. A method for detecting a clear driving path for a vehicle, said method comprising:

processing images from a camera to produce feature data for one or more patches from each of the images;

analyzing the feature data for the one or more patches using a trained classifier to determine a probability value;

selecting patches whose probability value exceeds a first probability threshold value as positive test samples;

selecting patches whose probability value does not exceed a second probability threshold value as negative test samples;

selecting patches whose probability value does not exceed a second probability threshold value as negative test samples;

using locations of the one or more patches to identify which training samples and which of the positive and negative test samples should be used to adaptively train the trained classifier, where the locations are defined in terms of feature vector space; and

making a clear-driving-path or non-clear-path determination for each of the one or more patches based on the probability value.

2. The method of claim 1 wherein processing images from a camera to produce feature data for one or more patches includes segmenting the images to define the one or more patches, and extracting the feature data for the one or more patches.

3. The method of claim 2 wherein segmenting the images to define the one or more patches includes selecting only patches with a size exceeding a predetermined size threshold.

4. The method of claim 2 wherein extracting the feature data for the one or more patches includes extracting one or more of;

geometric features, color features, texture features, and Histogram of Oriented Gradients features.

5. The method of claim 1 wherein analyzing the feature data for the one or more patches using a trained classifier includes using a Support Vector Machine classifier with a Radial Basis Function kernel.

6. The method of claim 5 wherein selecting patches whose probability value exceeds a first probability threshold value as positive test samples includes storing data for the positive test samples as stored elements in a current data frame, a previous data frame, and a cluster data pool.

7. The method of claim 6 wherein the stored elements in the current data frame, the previous data frame, and the cluster data pool are used as inputs to the Radial Basis Function kernel.

8. The method of claim 5 further comprising using patches for regions which were recently passed by the vehicle to further adaptively train the trained classifier.

9. The method of claim 8 wherein using patches for regions which were recently passed by the vehicle to further adaptively train the trained classifier includes using the patches for the regions which were recently passed by the vehicle as input to the Radial Basis Function kernel.

10. The method of claim 1 further comprising using the clear-driving-path or non-clear-path determination in a driving assistance system in the vehicle.

11. A method for detecting a clear driving path for a vehicle, said method comprising:

processing images from a camera to produce feature data for one or more patches from each of the images;

analyzing the feature data for the one or more patches using a trained classifier to determine a probability value, where the trained classifier is a Support Vector Machine classifier with a Radial Basis Function kernel;

selecting patches whose probability value exceeds a probability threshold value as positive test samples;

using locations of the one or more patches to identify which training samples and which of the positive test samples should be used to adaptively train the trained classifier, where the locations are defined in terms of feature vector space;

making a clear-driving-path or non-clear-path determination for each of the one or more patches based on the probability value; and

using the clear-driving-path or non-clear-path determination in a driving assistance system in the vehicle.

12. The method of claim 11 wherein processing images from a camera to produce feature data for one or more patches includes segmenting the images to define the one or more patches, and extracting the feature data for the one or more patches.

13. The method of claim 11 wherein selecting patches whose probability value exceeds a first probability threshold value as positive test samples includes storing data for the positive test samples as stored elements in a current data frame, a previous data frame, and a cluster data pool, and using the stored elements in the current data frame, the previous data frame, and the cluster data pool as inputs to the Radial Basis Function kernel.

14. The method of claim 11 further comprising using patches for regions which were recently passed by the vehicle as input to the Radial Basis Function kernel to further adaptively train the trained classifier.

15. A clear path detection system for a vehicle comprising:

a camera for providing images of a scene in front of the vehicle or behind the vehicle;

a segmentation module for defining one or more patches from each of the images;

a feature extraction module for extracting feature data for the one or more patches from each of the images;

a classifier module for analyzing the feature data to determine a probability value for each of the one or more patches, where the classifier module uses a Support Vector Machine classifier with a Radial Basis Function kernel;

a test sample retrieval module for using the feature data from patches whose probability value exceeds a probability threshold value, and which are local in feature vector space, as inputs to the Radial Basis Function kernel to adaptively train the classifier module; and

a decision module for evaluating the probability value to make a clear-driving-path or non-clear-path determination for each of the one or more patches.

16. The system of claim 15 wherein the test sample retrieval module stores data for the patches whose probability value exceeds the probability threshold value as stored elements in a current data frame, a previous data frame, and a cluster data pool.

17. The system of claim 15 wherein the test sample retrieval module uses locations, in feature vector space, of the one or more patches to identify local training samples and local test samples, and uses the local training samples and the local test samples as input to the Radial Basis Function kernel to adaptively train the classifier module.

18. The system of claim 15 wherein the test sample retrieval module also uses patches for regions which were recently passed by the vehicle as additional input to the Radial Basis Function kernel to adaptively train the classifier module.