Method for Recognizing Traffic Signs

• US 20110109476A1
• Filed: 03/31/2009
• Published: 05/12/2011
• Est. Priority Date: 03/31/2009
• Status: Active Grant

• Current Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Inventors: Ruta, Andrzej, Porikli, Fatih M.
• Granted Patent: US 8,041,080 B2
• Time in Patent Office: Days
• US Class Current: 340/905
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 10/255 Detecting or recognising potential candidate objects based on visual cues, e.g. shapes
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 20/582 of traffic signs

Abstract

A method recognizes a set of traffic signs in a sequence of images acquired of a vehicle environment by a camera mounted in a moving vehicle by detecting in each image, a region of interest (ROI) using a parameter space transform. The ROI is tracked and classified as a particular one of the signs. The classifier only uses a same class and a different class, and a regression function to update the classifier.

Claims

1. A method for recognizing a particular traffic sign of a set of traffic signs in a sequence of images acquired of a vehicle environment by a camera mounted in a moving vehicle, comprising a processor for performing steps of the method, comprising the steps of:

determining, in each image, a probability density map using a parameter space transform;

detecting, in each image, a region of interest (ROI) using a kernel density estimator on the probability density map;

tracking, in each next image after the ROI is detected, the ROI using an instant-specific regression function; and

recognizing the ROI as contain the particular traffic sign of the set of traffic signs using multiple one-versus-all (OVA) classifiers, wherein each classifier only uses a same class and a different class, and a final decision is taken using a winner-takes-all strategy that associates a real-valued score with each class, and the final decision recognizes the particular traffic sign.

2. The method of claim 1, further comprising:

generating an alert in response to recognizing the particular sign.

3. The method of claim 1, further comprising:

scaling each image, to a fixed size;

filtering, each image, to amplify a set of predetermine colors;

generating a gradient map for a set of predetermined colors;

applying the parameter space transform to the gradient map; and

treating a response of parameter space transform as a probability density map a confidence-weighted mean shift operator.

4. The method of claim 1, wherein the detecting is view-independent.

5. The method of claim 4, wherein the instant-specific regression uses an affine matrix M transform and an inverse M⁻

1 to transform between object coordinates and image coordinates, wherein the matrix M is for a unit square at an origin to an affine region enclosing the ROI.

6. The method of claim 1, wherein the tracking further comprises:

generating random deformations for each next image.

7. The method of claim 1, wherein the classifier is trained using training images of all possible pairs of the set of traffic signs.

8. The method of claim 7, wherein the classifier is trained using a modified AdaBoost process.

9. The method of claim 8, wherein the Adaboost process is trained with a sum of image features ƒ

_j

10. The method of claim 8, wherein the AdaBoost process minimizes a weighted error of the training examples, which is

e_j=W₊⁻

+W₋

⁺,

where W⁻

₊ and W⁺₋

indicate false negatives and false positives, respectively.

11. The method of claim 9, wherein the Adaboost classifier is