CN103473567A - Vehicle detection method based on partial models - Google Patents

Abstract

The invention relates to the technical field of vehicle detection, and particularly relates to a vehicle detection method based on partial models. The method comprises the steps of taking two parts of a vehicle object according to the easily covering degree so as to form vehicle models; respectively setting a model for the two parts by hybrid image templates with multiple characteristics; studying the hybrid image templates corresponding to the two parts and the image likelihood probability of the templates by virtue of a training image, and simultaneously studying the probability distribution of the position and scale relationship of the two parts; detecting the candidate of the two parts in a tested traffic image, and combining the vehicle candidate by virtue of the position and scale relationship of the two parts so as to realize vehicle detection. The vehicle detection method has the advantages of being adaptable to moderate vehicle transformation and multiple weather conditions, and can be used for processing vehicle covering in a complex traffic scene.

Description

A kind of vehicle checking method based on department pattern

Technical field

The present invention relates to the vehicle detection technical field, particularly a kind of vehicle checking method based on department pattern.

Background technology

Vehicle detection technology based on video is a part important in intelligent transportation system, for many application provide information of vehicles; As drive assist system, traffic video monitoring system etc.Vehicle checking method commonly used is to utilize the movable information of vehicle to detect vehicle at present, as the people such as Wu Ping Fei were published in the paper " Adaptive Vehicle Detector Approach for Complex Environments " (adaptive vehicle detecting device under complex scene) on " IEEE Transactions on Intelligent Transportation Systems (IEEE intelligent transportation transactions) " in 2012, utilize exactly movable information to detect vehicle.But the method is unsuitable for processing between vehicle, the situation of blocking is arranged, be difficult to separate exactly the adjacent vehicle mutually blocked.And because the information that lacks exercise of vehicle in traffic at a slow speed is also inapplicable.In addition, vehicle checking method based on characteristics of image often utilizes the information such as vehicle edge, texture to detect vehicle, as the paper " Learning Active Basis Model for Object Detection and Recognition " (learning activities basic mode type is for Target detection and identification) that the people such as Wu Yingnian delivered at " International Journal of Computer Vision (computer vision International Periodicals) " in 2010, it utilizes marginal information to detect vehicle.But many methods based on characteristics of image are not considered occlusion, be unsuitable for processing with the traffic scene blocked.

Summary of the invention

The technical matters that the present invention solves is to provide a kind of vehicle checking method based on department pattern, realizes the vehicle detection with the vehicles in complex traffic scene blocked.

The technical scheme that the present invention solves the problems of the technologies described above is:

Mainly comprise part selection, part modeling, model learning and vehicle detection;

Described part is selected, according to the degree of easily blocking Extraction parts (a) and part (b) two parts from Vehicle Object, for forming auto model; Partly (a) is difficult for the zone be blocked near vehicle window, and partly (b) is the zone easily be blocked around car plate;

Described part modeling, utilize and merge described two parts of manifold vision-mix template modeling;

Described model learning, utilize training image to learn the vision-mix template of described two parts and position and the scaling relation between two parts;

Described vehicle detection, utilize iterative process to detect the one or more vehicles in the test traffic image, in each iterative step, at first utilize template matches to detect two-part candidate, then utilize position and scaling relation built-up section candidate between two parts, realize vehicle detection.

Described model learning comprises the following steps:

Step S3-1, from the actual traffic image, the intercepting vehicle image is as training image, and training image quantity is no less than 1 width;

Step S3-2, utilize all image blocks of message maps method from the vision-mix template of described training image learning part (a) and part (b) correspondence and the image likelihood probability of part (a) and the vision-mix template that partly (b) is corresponding;

Step S3-3, utilize described training image study part (a) and the partly position between (b) and the probability distribution of scaling relation.

Described vehicle detection, utilize an iterative process to detect one or more vehicles from the test traffic image, comprises the following steps:

Step S4-1, utilize the vision-mix template of described part (a) and part (b) to carry out template matches to the test traffic image, extracts the candidate of described part (a) and part (b);

Step S4-2, utilize position and the scaling relation of described part (a) and part (b) that the candidate of part (a) and part (b) is combined out to one or more vehicle candidates, and calculate the vehicle detection probability that these vehicle candidates are corresponding;

Step S4-3, extract the vehicle candidate with maximum vehicle detection probability;

Step S4-4, compare maximum vehicle detection probability and vehicle detection threshold value in described step S4-3, if this vehicle detection probability is more than or equal to the vehicle detection threshold value, corresponding vehicle candidate is a detected Vehicle Object; Then utilize the test traffic image of erasing this detected Vehicle Object to carry out next iterative step, repeating step S4-1, S4-2, S4-3, S4-4; If this vehicle detection probability is less than the vehicle detection threshold value, whole iterative process stops, and the vehicle detection process finishes.

Described vision-mix template is comprised of one or more image blocks, and image block is divided into the types such as edge block, texture block, color block and flatness piece, and the image block that described vision-mix template comprises can be one or more in these types;

Described edge block is that the Gabor small echo primitive by specific direction means; Described texture block is that the histogram of gradients by image-region means; Described color block is meaned by the color histogram of image-region; The value representation that described flatness piece is obtained by the response of the Gabor wave filter of one or more directions in the superimposed image zone.

The image likelihood probability of the vision-mix template of described part (a) or part (b) is:

$p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)=q\left(I\right){\mathrm{\Π}}_{j=1}^{N_i}\frac{\exp \left\{{\mathrm{\λ}}_{\mathrm{ij}}f\left(I_{{\mathrm{\Λ}}_{\mathrm{ij}}}\right)\right\}}{Z_{\mathrm{ij}}},$

P wherein ₁and p ₂be respectively the vision-mix template of described part (a) and part (b), N _ip _ithe quantity of middle image block, q (I) is a reference distribution, λ _ijp _iin the coefficient of j image block, p _iin j image block and image

between distance, Z _ijit is normaliztion constant.

Described vehicle detection probability is:

$p\left(I\right|\{p_1,p_2\})=p(I_{{\mathrm{\Λ}}_1},I_{{\mathrm{\Λ}}_2}){\mathrm{\Π}}_{i=1}^{2}p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)$

Wherein I is a width test traffic image,

with

be respectively the image-region at the candidate place of the described part (a) that is detected and part (b) in the vehicle detection process,

meaned a pair of part (a) and the part (b) candidate between position and the probability of scaling relation.

The calculation procedure of described vehicle detection threshold value comprises:

At first, utilize step S4-1, S4-2, S4-3 detects vehicle from all described training images, and calculates corresponding vehicle detection probability;

Then, utilize the vehicle detection probability estimate vehicle detection threshold value of all described training images.

Vehicle checking method of the present invention has the following advantages:

(1) in part is selected, according to the degree of easily blocking, from Vehicle Object, selected two parts (a) and part (b) to form auto model, partly (a) is the zone around the vehicle window that is difficult for being blocked, and partly (b) is the zone around the car plate easily be blocked.This part has selected to reduce the impact of circumstance of occlusion on part (a).

(2) in part modeling, utilize vision-mix template modeling part (a) and part (b), the set of described vision-mix template multiple vehicle characteristics, comprise edge, texture, color and flatness etc., improved the detection accuracy of part (a) and part (b), realize the detailed description to information such as vehicle ' s contours, made the present invention adapt to multiple weather condition simultaneously.

(3) in vehicle detection, partly the testing process of (a) and part (b) is independently, candidate by built-up section (a) and part (b) is realized vehicle detection, make described auto model adapt to the appropriateness distortion of actual vehicle, reduced to block the impact on vehicle detection simultaneously.

The accompanying drawing explanation

Below in conjunction with accompanying drawing, the present invention is further described:

Fig. 1 is with the vehicles in complex traffic scene of occlusion in the embodiment of the present invention;

Fig. 2 is partly (a) and (b) image-region of correspondence in the embodiment of the present invention;

The schematic diagram that Fig. 3 is the training image of a part in the embodiment of the present invention;

Fig. 4 is part (a) and vision-mix template corresponding to part (b) in the embodiment of the present invention;

Fig. 5 is the vehicle detection result on the test traffic image in the embodiment of the present invention.

Embodiment

As Figure 1-5, implementation of the present invention is divided into four key steps: part is selected, part modeling, model learning and vehicle detection.Below introduce in detail this four steps:

Step S1: in part is selected, consider occlusion in vehicles in complex traffic scene (as Fig. 1), select two parts according to the degree of easily blocking from Vehicle Object, for forming auto model, these two parts are respectively part (a) and part (b), and wherein part (a) is difficult for the zone be blocked near vehicle window; Partly (b) is the zone easily be blocked around car plate.Part (a) comprises a part of roof of whole vehicle window, close vehicle window and parts place image-region and these parts image-regions on every side such as a part of hood of close vehicle window in embodiments of the present invention; Partly (b) comprised parts and these parts image-regions on every side such as a part of hood of car light, car plate and close car light.

Step S2: in part modeling, the present invention utilizes the described part of vision-mix template modeling (a) and part (b).A vision-mix template comprises one or more image blocks, and image block is divided into: edge block, and texture block, the types such as color block and flatness piece, the image block in the vision-mix template can be one or more in these types.

Described edge block is that the Gabor small echo primitive by specific direction means.Embodiments of the invention are used the Gabor small echo primitive of 16 directions to mean different edge block for example, certainly, herein as long as select to be no less than the Gabor small echo primitive of 1 direction, are not limited to 16 directions.

Described texture block is meaned by the histogram of gradients of image-region.Described histogram of gradients obtains by the Gabor filter response value of 16 directions in statistical picture zone in an embodiment of the present invention, certainly, herein as long as statistics is no less than the Gabor filter response value of 1 direction, is not limited to 16 directions.

Described color block is meaned by the color histogram of image-region.Described color histogram obtains by the pixel value of three Color Channels in the hsv color space in statistical picture zone in an embodiment of the present invention, certainly be not limited to the hsv color space herein, and be not limited to three Color Channels, as long as be no less than 1 Color Channel.

The value representation that described flatness piece is obtained by the response of the Gabor wave filter of one or more directions in the superimposed image zone.The described flatness piece of value representation that the Gabor filter response value of embodiments of the invention by 16 directions in the superimposed image zone obtains, herein as long as stack is no less than the Gabor filter response value of 1 direction, be not limited to 16 directions certainly.

Step S3: model learning comprises following three steps:

Step S3-1, from the actual traffic image, the intercepting vehicle image is as training image, the embodiment of the present invention has been used 20 width training images, certainly herein as long as use be no less than 1 width training image can, be not limited to 20 width training images, and The more the better.Fig. 3 has showed a part of training image schematic diagram.

Step S3-2, adopt the image block of message maps method (Information Projection Principle) from the vision-mix template of the described part of described training image learning (a) and part (b), obtain the probability distribution of the vision-mix template of described part (a) or part (b) simultaneously

$p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)=q\left(I\right){\mathrm{\Π}}_{j=1}^{N_i}\frac{\exp \left\{{\mathrm{\λ}}_{\mathrm{ij}}f\left(I_{{\mathrm{\Λ}}_{\mathrm{ij}}}\right)\right\}}{Z_{\mathrm{ij}}}---\left(1\right)$

P wherein ₁and p ₂respectively the vision-mix template of described part (a) and part (b), N _ip _ithe quantity of middle image block, q (I) is a reference distribution, λ _ijp _iin the coefficient of j image block,

p _iin j image block and image

between distance, Z _ijit is normaliztion constant.Fig. 4 has showed the vision-mix template of described part (a) that embodiment of the present invention learning goes out and part (b).

Step S3-3, from the position between the described part of described training image learning (a) and part (b) and the probability distribution of scaling relation.Position between embodiment of the present invention setting section (a) and part (b) and the probability distribution Gaussian distributed of scaling relation, from the parameter of described training image study Gaussian distribution.Certainly according to actual conditions herein part (a) and partly the position between (b) and the probability distribution of scaling relation also can obey the distribution of other type, be not limited to Gaussian distribution.

Step S4: the vehicle detection process is an iterative process, comprising:

Step S4-1, utilize the vision-mix template of described part (a) and part (b) to carry out template matches to the test traffic image, extracts the candidate of described part (a) and part (b).The embodiment of the present invention can be carried out convergent-divergent to the test traffic image when the test traffic image is carried out to template matches, to adapt to partly (a) and the partly size of the vision-mix template of (b), reaches the purpose of change of scale.

Step S4-2, utilize described part (a) and the part (b) position and scaling relation built-up section candidate, produce one or more vehicle candidates, and calculate the vehicle detection probability that these vehicle candidates are corresponding.Described vehicle detection probability is:

$p\left(I\right|\{p_1,p_2\})=p(I_{{\mathrm{\Λ}}_1},I_{{\mathrm{\Λ}}_2}){\mathrm{\Π}}_{i=1}^{2}p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)---\left(2\right)$

Wherein, I is a width test traffic image,

with

for the described part (a) that is detected in the vehicle detection process and the image-region at candidate place (b), position between a pair of part (a) and candidate (b) and the probability of scaling relation have been meaned.

Step S4-3, extract the vehicle candidate with maximum vehicle detection probability.

Step S4-4, compare maximum vehicle detection probability and vehicle detection threshold value in described step S4-3.If this maximum vehicle detection probability is more than or equal to the vehicle detection threshold value, corresponding vehicle candidate is a detected Vehicle Object, then will be detected Vehicle Object erases from the test traffic image, and to carry out next iterative step on the test traffic image of erasing detected Vehicle Object (be repeating step S4-1, S4-2, S4-3, S4-4).If this maximum vehicle detection probability is less than the vehicle detection threshold value, whole iterative process stops, and testing process finishes.Fig. 5 has showed the vehicle detection result of the present invention in vehicles in complex traffic scene.

In addition, the computation process of described vehicle detection threshold value comprises,

At first, utilize the vehicle checking method in step S4 to detect vehicle from all described training images, and calculate corresponding vehicle detection probability;

Then, utilize the vehicle detection probability estimate vehicle detection threshold value of all described training images.

Being more than that specific description of embodiments of the present invention, is not limiting the scope of the invention; All equivalents that can obtain according to aforementioned description, within all should being included in protection scope of the present invention.

Claims (15)

1. the vehicle checking method based on department pattern, is characterized in that: mainly comprise part selection, part modeling, model learning and vehicle detection;

Described part is selected, according to the degree of easily blocking Extraction parts (a) and part (b) two parts from Vehicle Object, for forming auto model; Partly (a) is difficult for the zone be blocked near vehicle window, and partly (b) is the zone easily be blocked around car plate;

Described part modeling, utilize and merge described two parts of manifold vision-mix template modeling;

Described model learning, utilize training image to learn the vision-mix template of described two parts and position and the scaling relation between two parts;

Described vehicle detection, utilize iterative process to detect the one or more vehicles in the test traffic image, in each iterative step, at first utilize template matches to detect two-part candidate, then utilize position and scaling relation built-up section candidate between two parts, realize vehicle detection.

2. vehicle checking method according to claim 1, it is characterized in that: described model learning comprises the following steps:

Step S3-1, from the actual traffic image, the intercepting vehicle image is as training image, and training image quantity is no less than 1 width;

Step S3-2, utilize all image blocks of message maps method from the vision-mix template of described training image learning part (a) and part (b) correspondence and the image likelihood probability of part (a) and the vision-mix template that partly (b) is corresponding;

Step S3-3, utilize described training image study part (a) and the partly position between (b) and the probability distribution of scaling relation.

3. vehicle checking method according to claim 1 is characterized in that: described vehicle detection, and utilize an iterative process to detect one or more vehicles from the test traffic image, comprise the following steps:

Step S4-1, utilize the vision-mix template of described part (a) and part (b) to carry out template matches to the test traffic image, extracts the candidate of described part (a) and part (b);

Step S4-2, utilize position and the scaling relation of described part (a) and part (b) that the candidate of part (a) and part (b) is combined out to one or more vehicle candidates, and calculate the vehicle detection probability that these vehicle candidates are corresponding;

Step S4-3, extract the vehicle candidate with maximum vehicle detection probability;

Step S4-4, compare maximum vehicle detection probability and vehicle detection threshold value in described step S4-3, if this vehicle detection probability is more than or equal to the vehicle detection threshold value, corresponding vehicle candidate is a detected Vehicle Object; Then utilize the test traffic image of erasing this detected Vehicle Object to carry out next iterative step, repeating step S4-1, S4-2, S4-3, S4-4; If this vehicle detection probability is less than the vehicle detection threshold value, whole iterative process stops, and the vehicle detection process finishes.

4. vehicle checking method according to claim 2 is characterized in that: described vehicle detection, and utilize an iterative process to detect one or more vehicles from the test traffic image, comprise the following steps:

Step S4-1, utilize the vision-mix template of described part (a) and part (b) to carry out template matches to the test traffic image, extracts the candidate of described part (a) and part (b);

Step S4-2, utilize position and the scaling relation of described part (a) and part (b) that the candidate of part (a) and part (b) is combined out to one or more vehicle candidates, and calculate the vehicle detection probability that these vehicle candidates are corresponding;

Step S4-3, extract the vehicle candidate with maximum vehicle detection probability;

Step S4-4, compare maximum vehicle detection probability and vehicle detection threshold value in described step S4-3, if this vehicle detection probability is more than or equal to the vehicle detection threshold value, corresponding vehicle candidate is a detected Vehicle Object; Then utilize the test traffic image of erasing this detected Vehicle Object to carry out next iterative step, repeating step S4-1, S4-2, S4-3, S4-4; If this vehicle detection probability is less than the vehicle detection threshold value, whole iterative process stops, and the vehicle detection process finishes.

5. according to the described vehicle checking method of claim 1 to 4 any one, it is characterized in that: described vision-mix template is comprised of one or more image blocks, image block is divided into the types such as edge block, texture block, color block and flatness piece, and the image block that described vision-mix template comprises can be one or more in these types;

Described edge block is that the Gabor small echo primitive by specific direction means; Described texture block is that the histogram of gradients by image-region means; Described color block is meaned by the color histogram of image-region; The value representation that described flatness piece is obtained by the response of the Gabor wave filter of one or more directions in the superimposed image zone.

6. according to claim 2,3 or 4 described vehicle checking methods, it is characterized in that: the image likelihood probability of the vision-mix template of described part (a) or part (b) is:

$p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)=q\left(I\right){\mathrm{\Π}}_{j=1}^{N_i}\frac{\exp \left\{{\mathrm{\λ}}_{\mathrm{ij}}f\left(I_{{\mathrm{\Λ}}_{\mathrm{ij}}}\right)\right\}}{Z_{\mathrm{ij}}},$

P wherein ₁and p ₂be respectively the vision-mix template of described part (a) and part (b), N _ip _ithe quantity of middle image block, q (I) is a reference distribution, λ _ijp _iin the coefficient of j image block,

p _iin j image block and image

between distance, Z _ijit is normaliztion constant.

7. vehicle checking method according to claim 5 is characterized in that: the image likelihood probability of the vision-mix template of described part (a) or part (b) is:

$p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)=q\left(I\right){\mathrm{\Π}}_{j=1}^{N_i}\frac{\exp \left\{{\mathrm{\λ}}_{\mathrm{ij}}f\left(I_{{\mathrm{\Λ}}_{\mathrm{ij}}}\right)\right\}}{Z_{\mathrm{ij}}},$

P wherein ₁and p ₂be respectively the vision-mix template of described part (a) and part (b), N _ip _ithe quantity of middle image block, q (I) is a reference distribution, λ _ijp _iin the coefficient of j image block,

p _iin j image block and image between distance, Z _ijit is normaliztion constant.

8. according to the described vehicle checking method of claim 3 or 4, it is characterized in that: described vehicle detection probability is:

$p\left(I\right|\{p_1,p_2\})=p(I_{{\mathrm{\Λ}}_1},I_{{\mathrm{\Λ}}_2}){\mathrm{\Π}}_{i=1}^{2}p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)$

Wherein I is a width test traffic image, with

be respectively the image-region at the candidate place of the described part (a) that is detected and part (b) in the vehicle detection process,

meaned a pair of part (a) and the part (b) candidate between position and the probability of scaling relation.

9. vehicle checking method according to claim 5, it is characterized in that: described vehicle detection probability is:

$p\left(I\right|\{p_1,p_2\})=p(I_{{\mathrm{\Λ}}_1},I_{{\mathrm{\Λ}}_2}){\mathrm{\Π}}_{i=1}^{2}p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)$

Wherein I is a width test traffic image,

with

be respectively the image-region at the candidate place of the described part (a) that is detected and part (b) in the vehicle detection process,

meaned a pair of part (a) and the part (b) candidate between position and the probability of scaling relation.

10. vehicle checking method according to claim 6, it is characterized in that: described vehicle detection probability is:

$p\left(I\right|\{p_1,p_2\})=p(I_{{\mathrm{\Λ}}_1},I_{{\mathrm{\Λ}}_2}){\mathrm{\Π}}_{i=1}^{2}p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)$

Wherein I is a width test traffic image, with

be respectively the image-region at the candidate place of the described part (a) that is detected and part (b) in the vehicle detection process,

meaned a pair of part (a) and the part (b) candidate between position and the probability of scaling relation.

11. vehicle checking method according to claim 7 is characterized in that: described vehicle detection probability is:

$p\left(I\right|\{p_1,p_2\})=p(I_{{\mathrm{\Λ}}_1},I_{{\mathrm{\Λ}}_2}){\mathrm{\Π}}_{i=1}^{2}p\left(I_{{\mathrm{\Λ}}_i}\right|p_i)$

Wherein I is a width test traffic image,

with be respectively the image-region at the candidate place of the described part (a) that is detected and part (b) in the vehicle detection process,

meaned a pair of part (a) and the part (b) candidate between position and the probability of scaling relation.

12. according to the described vehicle checking method of claim 3 or 4, it is characterized in that: the calculation procedure of described vehicle detection threshold value comprises:

At first, utilize step S4-1, S4-2, S4-3 detects vehicle from all described training images, and calculates corresponding vehicle detection probability;

Then, utilize the vehicle detection probability estimate vehicle detection threshold value of all described training images.

13. vehicle checking method according to claim 5 is characterized in that: the calculation procedure of described vehicle detection threshold value comprises:

At first, utilize step S4-1, S4-2, S4-3 detects vehicle from all described training images, and calculates corresponding vehicle detection probability;

Then, utilize the vehicle detection probability estimate vehicle detection threshold value of all described training images.

14. vehicle checking method according to claim 7 is characterized in that: the calculation procedure of described vehicle detection threshold value comprises:

At first, utilize step S4-1, S4-2, S4-3 detects vehicle from all described training images, and calculates corresponding vehicle detection probability;

Then, utilize the vehicle detection probability estimate vehicle detection threshold value of all described training images.

15. vehicle checking method according to claim 11 is characterized in that: the calculation procedure of described vehicle detection threshold value comprises:

At first, utilize step S4-1, S4-2, S4-3 detects vehicle from all described training images, and calculates corresponding vehicle detection probability;

Then, utilize the vehicle detection probability estimate vehicle detection threshold value of all described training images.

Images