METHOD AND DEVICE FOR TRAFFIC SIGN RECOGNITION

• US 20170323452A1
• Filed: 11/18/2015
• Published: 11/09/2017
• Est. Priority Date: 11/18/2014
• Status: Active Grant

• Current Assignee: Agency For Science Technology and Research (Government of Singapore)
• Original Assignee: Agency For Science Technology and Research (Government of Singapore)
• Inventors: CHEN, Tao, LU, Shijian
• Granted Patent: US 10,332,266 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 16/434 using image data, e.g. images, photos, pictures taken by a user
• CPC Class Codes: G06F 17/18 for evaluating statistical data , e.g. average values, frequency distributions, probability functions, regression analysis forecasting specially adapted for a specific administrative, business or logistic context G06Q10/04
• CPC Class Codes: G06T 7/246 using feature-based methods, e.g. the tracking of corners or segments
• CPC Class Codes: G06V 20/582 of traffic signs

Abstract

Embodiments provide a method for recognizing a traffic sign from an input image. The method may include extracting image features from the input image; modifying the image features based on a predetermined image saliency map to determine modified image features; determining a plurality of traffic sign candidates by applying an adaptive boosting algorithm on the modified image features; determining a confidence score for each traffic sign candidate by applying a support vector regression algorithm; and recognizing a traffic sign based on the confidence score for each traffic sign candidate.

Claims

1. A method for recognizing a traffic sign from an input image, the method comprising:

extracting image features from the input image;

modifying the image features based on a predetermined image saliency map to determine modified image features;

determining a plurality of traffic sign candidates by applying an adaptive boosting algorithm on the modified image features;

determining a confidence score for each traffic sign candidate by applying a support vector regression algorithm; and

recognizing a traffic sign based on the confidence score for each traffic sign candidate.

2. The method of claim 1, further comprising:

determining the predetermined image saliency map based on at least one of color saliency or spatial location information of traffic signs.

3. The method of claim 2, wherein the color saliency is determined dependent on categories of traffic signs.

4. The method of claim 1, wherein the image features comprise at least one of color, gradient, orientation or shape features.

5. The method of claim 1, wherein the adaptive boosting algorithm comprises:

a traffic sign training process using training images to learn a set of decision trees, the training images comprising traffic sign images and non-traffic sign images.

6. The method of claim 5, wherein applying the adaptive boosting algorithm on the modified image features comprises:

applying the set of decision trees on the modified image features of the input image to determine the plurality of traffic sign candidates.

7. The method of claim 1, wherein determining the confidence score for each traffic sign candidate comprises:

quantizing each traffic sign candidate into a bag-of-words histogram using a predefined codebook; and

applying the support vector regression algorithm on the bag-of-words histogram to determine the confidence score for each traffic sign candidate.

8. The method of claim 1, wherein the support vector regression algorithm is trained using training samples, wherein the training samples comprises traffic sign training candidates determined by applying the adaptive boosting algorithm to training images, wherein the training samples comprise non-traffic sign regions randomly sampled from the training images.

9. The method of claim 8, wherein the support vector regression algorithm is trained based on a similarity coefficient of each training sample, wherein the similarity coefficient represents an intersection ratio between each training sample and a corresponding true traffic sign.

10. The method of claim 1, wherein recognizing the traffic sign comprises:

recognizing that the traffic sign is present in the traffic sign candidate if the confidence score for the traffic sign candidate is higher than a predetermined score.

11. The method of claim 1, wherein recognizing the traffic sign comprises:

recognizing that the traffic sign of a predetermined category is present in the traffic sign candidate if the confidence score for the traffic sign candidate matches a predetermined score corresponding to the predetermined category.

12. A method for recognizing a traffic sign from a video having a sequence of input images, the method comprising:

recognizing a first traffic sign from a first input image of the video according to claim 1;

determining a first size of the first traffic sign in the first input image;

determining one or more sliding windows with one or more sizes, the one or more sizes are in a range from the first size to a second size larger than the first size;

scanning a subsequent input image of the video using the one or more sliding windows to recognize a second traffic sign in the subsequent input image.

13. The method of claim 12, further comprising:

determining a position of the first traffic sign in the first input image;

determining a search window for the subsequent input image, the search window for the subsequent input image having a position substantially similar to the position of the first traffic sign; and

scanning the subsequent input image using the one or more sliding windows in the search window to recognize the second traffic sign.

14. The method of claim 12, further comprising:

if no traffic sign is recognized from the first input image, randomly generating a search window for the subsequent input image with a position different from a position of a search window used for the first input image; and

scanning the subsequent input image in the randomly generated search window to recognize the second traffic sign in the subsequent input image.

15. A device for recognizing a traffic sign from an input image, the device comprising:

an image feature extracting circuit for extracting image features from the input image;

an image feature modifying circuit for modifying the image features based on a predetermined image saliency map to determine modified image features;

a traffic sign candidate determining circuit for determining a plurality of traffic sign candidates by applying an adaptive boosting algorithm on the modified image features;

a confidence score determining circuit for determining a confidence score for each traffic sign candidate by applying a support vector regression algorithm; and

a traffic sign recognizing circuit for recognizing a traffic sign based on the confidence score for each traffic sign candidate.

16. The device of claim 15, wherein the predetermined image saliency map is determined based on at least one of color saliency or spatial location information of traffic signs.

17. The device of claim 15, wherein the traffic sign candidate determining circuit is configured to train the adaptive boosting algorithm using training images to learn a set of decision trees, the training images comprising traffic sign images and non-traffic sign images.

18. The device of claim 17, wherein the traffic sign candidate determining circuit is configured to apply the set of decision trees on the modified image features of the input image to determine the plurality of traffic sign candidates.

19. The device of claim 15, wherein the confidence score determining circuit is configured to quantize each traffic sign candidate into a bag-of-words histogram using a predefined codebook;

and apply the support vector regression algorithm on the bag-of-words histogram to determine the confidence score for each traffic sign candidate.

20. The device of claim 15, wherein the confidence score determining circuit is configured to train the support vector regression algorithm using training samples, wherein the training samples comprises traffic sign training candidates determined by applying the adaptive boosting algorithm to training images, wherein the training samples comprises non-traffic sign regions randomly sampled from the training images.