Vehicle detection method based on hybrid image template

Patent metadata

Field	Value
Publication number	US-10157320-B2
Application number	US-201314392309-A
Country	US
Kind code	B2
Filing date	Jun 28, 2013
Priority date	Jun 28, 2013
Publication date	Dec 18, 2018
Grant date	Dec 18, 2018

Abstract

The invention discloses a vehicle detection method based on hybrid image template. This method consists of the three steps. Firstly, use no less than one vehicle image for template learning. Secondly, use information projection algorithm to learn a hybrid image template from the training images for vehicle object. The hybrid image template consists of no one less than image patch. Meanwhile, calculate the likelihood probability distribution of this template. Thirdly, use the learned HIT template to detect vehicle objects from testing images. The invention is suitable to detect vehicles with various vehicle shapes, vehicle poses, time-of-day and weather conditions. Besides vehicle localization, this method can also provide the detailed description of vehicle object.

First claim

We claim:

1. A vehicle detection method for detecting a vehicle image in measured images based on hybrid image template HIT, comprising the following steps:
Step S 1 : collecting one or more vehicle images as training images;
Step S 2 : utilizing an information projection algorithm to learn all of image patches for representing vehicle object in the HIT from the training images and to compute image likelihood probability distribution of this hybrid image template, wherein the HIT consists of the multiple patches with different image features which are categorized as sketch patch, texture patch, color patch and flatness patch, wherein the image likelihood probability distribution indicates occurrence probability of the HIT in conditions of the training image data;
Step S 3 : detecting regions in which an image block of the HIT is present in the input testing image to acquire candidate vehicle regions in which the vehicles are located in the testing images;
Step 3 - 1 - 6 : using the patches in the HIT and the image likelihood probability to compute vehicle detection scores of the vehicle candidates region;
Step S 3 - 1 - 7 : selecting the vehicle candidate regions with the maximum vehicle detection scores, comparing the maximum vehicle score with the predefined vehicle detection threshold; if the maximum vehicle detection score is no less that the vehicle detection threshold, determining the corresponding vehicle candidate region as a vehicle object, and determining the location and detail sketch information of the vehicle object;
Step S 32 : removing the determined vehicle object from the testing image and using the remaining image to detect a next vehicle object and location and detail sketch information thereof by performing above steps, repeating above steps to detect all vehicle objects and location and detail sketch information thereof with an iterative method; and
visually indicating in the input testing image the candidate regions determined to be vehicle objects.

2. The method according to claim 1 , wherein the HIT consists of one or more types of image patches, containing sketch patch, texture patch, color patch and flatness patch.

3. The method according to claim 2 , wherein the sketch patch is modeled by a Gabor wavelet with one specific orientation, the texture patch is modeled by gradient histogram of local rectangular image region for the corresponding training image region, the color patch is modeled by color histogram of local rectangular image region for the corresponding training image region, and the flatness patch is modeled by superposition value of the responsive value of the Gabor filter for in one or more orientations in a local rectangular image region in the corresponding training image.

4. The method according to claim 3 , wherein the gradient histogram for the texture patch is acquired by computing statistics of Gabor wavelet filter responses of local rectangular image region in the corresponding training image region; The color histogram for the color patch is acquired by computing statistics of pixel values of the three color channels in color space of local rectangular image region; and the flatness patch is acquired by summing Gabor filter responses of local rectangular image region with one or more orientations.

5. The method according to claim 1 , wherein the likelihood probability of the HIT computed by the information project algorithm is

p
⁡

(

I
❘

H
⁢

⁢
I
⁢

⁢
T

)

=

q
⁡

(
I
)

⁢

∏

i
=
1

N

⁢

exp
⁢

{

λ
i

⁢

r
⁡

(

I

Λ
i

)

}

Z
i

,

wherein I is an image, HIT is learned from training images, q(I) is a reference distribution, N is the number of the image patch in the HIT, λ i is the coefficient of the i-th image patch in the HIT, Z i is a normalization constant, r(I Λ i ) is the similarity measure between an image region I 79 i in I and the i-th image patch in the HIT.

6. The method according to claim 1 , wherein If the HIT contains the sketch patch, the texture patch, the color patch and the flatness patch, the step S 313 further contains the following sub-steps:
Step S 3131 : the color patches in the HIT are used to filter the testing image to detect and obtain candidate color patches. Local image regions in the testing image are detected as color patch candidates and these local image regions have the similar color feature with the color patches in the HIT; and
Step S 3132 : the sketch patches, texture patches, and flatness patches in the HIT are utilized to filter the revised sketch image to detect and obtain sketch patch candidate, texture patch candidate and flatness patch candidates.

7. The method according to claim 1 , wherein the number of Gabor orientation used in step S 3 - 1 - 1 is no less than the number of the Gabor orientation applied to describing sketch patch in step S 2 , the number of filter orientation applied to compute texture patch in step S 2 , the number of Gabor orientation applied to compute flatness patch in step S 2 .

8. The method according to claim 1 , wherein the vehicle detection score of vehicle candidate regions are calculated by using the following formula:

score
⁡

(

I
❘

Assignees

• Inst Automation Cas

Inventors

• Wang Feiyue
• Li Ye

Classifications

• G06V20/54

  of traffic, e.g. cars on the road, trains or boats · CPC title

• G06T7/74 Primary

  involving reference images or patches · CPC title

• G06V10/7515

  Shifting the patterns to accommodate for positional errors · CPC title

• G06T2207/30252

  Vehicle exterior; Vicinity of vehicle · CPC title

• G06K9/6203

  Physics · mapped topic