CN105160362A - Runway FOD (Foreign Object Debris) image detection method and device - Google Patents

Abstract

The invention relates the field of 2D image-based object detection and object identification and particularly relates to a runway FOD (Foreign Object Debris) image detection method and device. The invention provides a runway FOD (Foreign Object Debris) image detection method and device for solving problems in the prior art. Imaging characteristic distribution rules of the inherent structure of a runway in various weather and illumination conditions can be obtained through training, and an elimination model is established for non-FOD objects; and in the FOD detection process on a real runway, edge detection of local appearance differences is firstly carried out to obtain a suspected object, then rapid non-FOD object elimination is carried out through the elimination model obtained in the previous training, and a FOD object is finally obtained and an alarm is given out. The invention establishes the elimination model for the non-FOD objects through four positive sample image sets, and calculates and counts identification parameters of four types of runway surface objects different from a negative sample image set, so as to achieve the complete FOD detection of a runway monitor area.

Description

A kind of runway FOD image detection method and device

Technical field

The present invention relates to the object detection based on 2D image and object identification field, especially relate to a kind of runway FOD image detection method and device.

Background technology

Runway invasion foreign matter (hereinafter referred FOD) has significant threat safely to flight.After French Concorde air crash, research institution and company are all being devoted to the exploitation of investigation and application system of FOD Detection Techniques in many ways.The technology that the application system of current main-stream uses is the radar exploration technique and image detection technology.Image detection technology is because of its advantage of lower cost, and night vision technology and image detection development, just more and more accepted and approve.

The technological difficulties that image detection faces are mainly from the diversity of runway appearance details and FOD.First, runway itself is not simple gray plane, but has various structures.Runway has the structure such as runway light of guide identifier line, embedding, there is gap between runway plate, the runway of concrete stone block material is visible finger stone material in face, road.Secondly, FOD's is of a great variety, and common just relates to more than ten kinds such as spanner, iron chains, sheet metal, rubber for tire, and the difference in appearance brought because losing attitude difference more cannot be added up, and some FOD that even airport occurs are unpredictable.3rd, runway and FOD appearance in the environment of work may occur multiple changeable.In outdoor environment, the change of illumination, rainfall, snowfall, frost all can cause the change of runway and FOD appearance.In the process used, tire friction can be left a trace on runway, and weathering also can cause runway crack, even local damage.

Above difficult point has larger restriction to more existing Detection Techniques performances.The method (as rim detection, color distortion etc.) detected is carried out in change based on local appearance, easily detects runway self structure and sends false-alarm.Take pictures as original image to runway, take pictures in each detection and to compare with original image, detection difference thus the method for detection FOD, be difficult to the variability adapting to working environment.And detector is trained respectively to common FOD kind, and carry out the method for certain objects detection, limited in one's ability to non-common FOD species detection.If also train detector to non-common FOD kind, then need the of a great variety of identification in detection process, speed of detection is slow.

Summary of the invention

Technical matters to be solved by this invention is: for prior art Problems existing, a kind of runway FOD image detection method and device are provided, this method obtains the Characteristic Distribution of runway self inherent structure imaging under multiple meteorology and illumination condition in early stage by training, sets up the troubleshooting model of non-FOD target; In FOD detection process, first rim detection is carried out to local appearance difference and obtain suspected target, then carry out quick non-FOD target by the troubleshooting model that early stage, training obtained and get rid of, finally obtain FOD target and report to the police.

The technical solution used in the present invention is as follows:

A kind of runway FOD image detection method comprises:

Step 1: pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway are taken pictures, sets up four positive sample image set respectively; Common FOD is taken pictures, sets up a negative sample image collection;

Step 2: the troubleshooting model setting up non-FOD target from four positive sample image set, calculates respectively and adds up the identification parameter that four class runway surface objects are different from negative sample image collection;

Step 3: image acquisition is carried out to actual runway, suspected target detection is carried out to often opening image edge detection method, then get rid of runway surface object according to the identification parameter in the troubleshooting model of non-FOD target, realize detecting the complete FOD of runway monitor area.

Further, described step 1 specifically comprises:

Step 11: take pictures to the whole and part of the pilot lamp of runway, tag line, plate gap, tire trace four type objects, every type objects should take at least 10 samples, each sample should at least be taken from 10 different angles; All images are cut off redundant area, only retains the smallest square region comprising subject; Rear image down to 32 × 32 pixel is cut out by all;

Step 12: under fine day, cloudy day, rainy day, haze weather, and morning, noon, under dusk illumination condition, repeat step 11, generate positive sample image set;

Step 13: put common FOD on runway, the diameter of each object is all not less than minimum detection of a target diameter, repeat step 11,12 shooting and image processing method, generate negative sample image collection.

Further, described minimum detection of a target diameter is for being less than 3cm.

Further, described step 2 concrete steps comprise:

Step 21: coloured images all in sample image set are converted into gray level image, use character description method that each sub-picture is converted into a proper vector respectively, wherein the proper vector of four positive sample image set generations is called four positive proper vector set, and the proper vector that negative sample image collection generates is called negative proper vector set;

Step 22: by each positive proper vector set, clustering method is used proper vector to be divided into K cluster respectively, setting pilot lamp cluster is a1, tag line cluster is a2, plate gap cluster is a3, tire trace cluster is a4, then K=a1+a2+a3+a4; Wherein the value of a1, a2, a3, a4 is the integer between 1 to 5;

Step 23: to each cluster, calculates the approximate Gaussian model parameter (μ of its proper vector distribution density _k, Φ _k), parameter μ _kfor the mean value of proper vectors all in cluster, parameter Φ _kfor the covariance matrix of proper vectors all in cluster, k is the sequence number of cluster and 1≤k≤K; Arbitrfary point X and cluster k similarity P in feature space _k(X) computing formula is:

P _k(X)＝exp(-(X-μ _k) ^TΦ _k ^-1(X-μ _k))

Step 24: to each cluster, if it comprises N ₁individual proper vector, each proper vector X _irepresent, wherein 1≤i≤N ₁, calculate the minimum similarity of all proper vectors and this cluster in this cluster if negative proper vector set comprises N2 proper vector, each proper vector Y _j, represent, wherein 1≤j≤N ₂, calculate the highest similarity of all proper vectors and this cluster in negative proper vector set cluster k similarity decision threshold T _kcomputing formula be:

$T_k=\left\{\begin{array}{cc}{L}_2& L_1\≤L_2\\ \frac{L_1+L_2}{2}& L_1>L_2\end{array}\right..$

Further, described character description method is SIFT feature extracting method, HOG feature extracting method or SURF algorithm; Clustering method is K-Mean algorithm, BIRCH algorithm or DBSCAN algorithm.

Further, described step 3 specifically comprises:

Step 31: to sub regions each in runway color image shot successively, change the coloured image of shooting into gray level image, uses smoothing and noise-reducing process to gray level image;

Step 32: use Canny edge detection algorithm to generate binary edge image to often opening gray level image, to obtain in image the coordinate set of promising 1 pixel, the pixel being set to 1 has N ₃individual, then set expression is C={ (x ₁, y ₁), (x ₂, y ₂) ... (x _n3, y _n3);

Step 33: newly-built null set D, moves to set D by first coordinate in set C; Repeat traversal set C, if there is the distance of arbitrary coordinate in coordinate to set D to be less than or equal to minW in set C, then the coordinate in set C is transferred to set D, until the arbitrary coordinate in set C is all greater than minW to the arbitrary coordinate distance in set D, or set C is null set; Wherein minW unit is pixel, and minW value is 10 to 30;

Step 34: the institute in statistics set D a little in X-axis maximum coordinates be maxX, min coordinates is minX, institute a little maximum coordinates maxY in Y-axis, min coordinates minY in statistics set D, get with centered by, the square gray level image region that max (maxX-minX, maxY-minY) is the length of side, uses the character description method of step 21 that this image-region is changed into proper vector V;

Step 35: the similarity P calculating V and K cluster respectively _k(V), if all P _k(V) all P is met _k(V) <T _k, then think and find FOD object, send FOD and report to the police, and target respective coordinates is provided $(\frac{\min X+\max X}{2},\frac{\min Y+\max Y}{2});$

Step 36: repeat step 32 to 35 successively, until set C is null set, completes the detection to FOD in an image;

Step 37: repeat step 32 to 36 successively, detect until complete to the FOD of all images, realizes detecting the complete FOD of runway monitor area.

A kind of runway FOD imaging detection device comprises

Non-FOD target troubleshooting model module, for taking pictures to pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway, sets up four positive sample image set respectively; Common FOD is taken pictures, sets up a negative sample image collection; Set up the troubleshooting model of non-FOD target from four positive sample image set, calculate respectively and add up the identification parameter that four class runway surface objects are different from negative sample image collection;

FOD image detection module, image acquisition is carried out to actual runway, carry out suspected target detection to often opening image edge detection method, then according in the troubleshooting model of non-FOD target identification parameter get rid of runway surface object, realize detecting the complete FOD of runway monitor area.

Further, the troubleshooting model module of described non-FOD target is taken pictures to pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway, sets up four positive sample image set respectively; Common FOD is taken pictures, sets up a negative sample image collection detailed process and comprise:

Step 11: take pictures to the whole and part of the pilot lamp of runway, tag line, plate gap, tire trace four type objects, every type objects should take at least 10 samples, each sample should at least be taken from 10 different angles; All images are cut off redundant area, only retains the smallest square region comprising subject; Rear image down to 32 × 32 pixel is cut out by all;

Step 12: under fine day, cloudy day, rainy day, haze weather, and morning, noon, under dusk illumination condition, repeat step 11, generate positive sample image set;

Step 13: put common FOD on runway, the diameter of each object is all not less than minimum detection of a target diameter, repeat step 11,12 shooting and image processing method, generate negative sample image collection.

Further, the troubleshooting model module of described non-FOD target calculates respectively from four positive sample image set and adds up the decision threshold T that four class runway surface objects are different from negative sample image collection _kdetailed process is:

Step 21: coloured images all in sample image set are converted into gray level image, use character description method that each sub-picture is converted into a proper vector respectively, wherein the proper vector of four positive sample image set generations is called four positive proper vector set, and the proper vector that negative sample image collection generates is called negative proper vector set;

Step 22: by each positive proper vector set, clustering method is used proper vector to be divided into K cluster respectively, setting pilot lamp cluster is a1, tag line cluster is a2, plate gap cluster is a3, tire trace cluster is a4, then K=a1+a2+a3+a4; Wherein the value of a1, a2, a3, a4 is the integer between 1 to 5;

Step 23: to each cluster, calculates the approximate Gaussian model parameter (μ of its proper vector distribution density _k, Φ _k), parameter μ _kfor the mean value of proper vectors all in cluster, parameter Φ _kfor the covariance matrix of proper vectors all in cluster, k is the sequence number of cluster and 1≤k≤K; Arbitrfary point X and cluster k similarity P in feature space _k(X) computing formula is:

P _k(X)＝exp(-(X-μ _k) ^TΦ _k ^-1(X-μ _k))

Step 24: to each cluster, if it comprises N ₁individual proper vector, each proper vector X _irepresent, wherein 1≤i≤N ₁, calculate the minimum similarity of all proper vectors and this cluster in this cluster if negative proper vector set comprises N2 proper vector, each proper vector Y _j, represent, wherein 1≤j≤N ₂, calculate the highest similarity of all proper vectors and this cluster in negative proper vector set cluster k similarity decision threshold T _kcomputing formula be:

$T_k=\left\{\begin{array}{cc}{L}_2& L_1\&le;L_2\\ \frac{L_1+L_2}{2}& L_1>L_2\end{array}\right..$

Further, described FOD image detection module, carries out image acquisition to actual runway, according to the judgment threshold T in the troubleshooting model module of non-FOD target _kthe actual runway image collected is detected, realizes detecting the complete FOD of runway monitor area;

Step 31: to sub regions each in runway color image shot successively, change the coloured image of shooting into gray level image, uses smoothing and noise-reducing process to gray level image;

Step 32: use Canny edge detection algorithm to generate binary edge image to often opening gray level image, to obtain in image the coordinate set of promising 1 pixel, the pixel being set to 1 has N ₃individual, then set expression is C={ (x ₁, y ₁), (x ₂, y ₂) ... (x _n3, y _n3);

Step 33: newly-built null set D, moves to set D by first coordinate in set C; Repeat traversal set C, if there is the distance of arbitrary coordinate in coordinate to set D to be less than or equal to minW in set C, then the coordinate in set C is transferred to set D, until the arbitrary coordinate in set C is all greater than minW to the arbitrary coordinate distance in set D, or set C is null set; Wherein minW unit is pixel, and minW value is 10 to 30;

Step 34: the institute in statistics set D a little in X-axis maximum coordinates be maxX, min coordinates is minX, institute a little maximum coordinates maxY in Y-axis, min coordinates minY in statistics set D, get with centered by, the square gray level image region that max (maxX-minX, maxY-minY) is the length of side, uses the character description method of step 21 that this image-region is changed into proper vector V;

Step 35: the similarity P calculating V and K cluster respectively _k(V), if all P _k(V) all P is met _k(V) <T _k, then think and find FOD object, send FOD and report to the police, and target respective coordinates is provided $(\frac{\min X+\max X}{2},\frac{\min Y+\max Y}{2});$

Step 36: repeat step 32 to 35 successively, until set C is null set, completes the detection to FOD in an image;

Step 37: repeat step 32 to 36 successively, detect until complete to the FOD of all images, realizes detecting the complete FOD of runway monitor area.

In sum, owing to have employed technique scheme, the invention has the beneficial effects as follows:

1, propose one FOD image detection method fast, method, by detecting local appearance difference, can realize quick doubtful FOD detection;

2, the method testing result proposed is accurate, and method carries out non-FOD object identification to doubtful FOD, effectively can get rid of false-alarm;

3, the method detectivity proposed and adaptive capacity to environment by force, as long as FOD appearance can be different from runway, runway inherent structure or texture, also have detectivity to multiple FOD (comprising non-common FOD).

Embodiment

All features disclosed in this instructions, or the step in disclosed all methods or process, except mutually exclusive feature and/or step, all can combine by any way.

Arbitrary feature disclosed in this instructions (comprising any accessory claim, summary), unless specifically stated otherwise, all can be replaced by other equivalences or the alternative features with similar object.That is, unless specifically stated otherwise, each feature is an example in a series of equivalence or similar characteristics.

Principle of work: the present invention is directed to runway invasion foreign matter (hereinafter referred FOD) detection problem, propose a kind of image FOD detection method.Method obtains the Characteristic Distribution of runway self inherent structure imaging under multiple meteorology and illumination condition in early stage by training, sets up the troubleshooting model of non-FOD target.In FOD detection process, first rim detection is carried out to local appearance difference and obtain suspected target, then carry out quick non-FOD target by the troubleshooting model that early stage, training obtained and get rid of, finally obtain FOD target and report to the police.Specifically comprise:

One, training in early stage obtains runway self the inherent structure regularity of distribution, and set up the troubleshooting model of non-FOD target, concrete steps are:

Step 1: pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway are taken pictures, sets up four positive sample image set respectively; Common FOD (at least comprising runway concrete block, spanner, tire debris, fuel tank cap, plastic tube, metallic element six type objects) is taken pictures, sets up a negative sample image collection;

Step 2: the troubleshooting model setting up non-FOD target from four positive sample image set, calculates respectively and adds up the identification parameter that four class runway surface objects are different from negative sample image collection;

Further, wherein step 1 is specially:

Step 11: take pictures to the whole and part of the pilot lamp of runway, tag line, plate gap, tire trace four type objects, every type objects should take at least 10 samples, each sample should at least be taken from 10 different angles; All images are cut off redundant area, only retains the smallest square region comprising subject; Rear image down to 32 × 32 pixel is cut out by all;

Step 12: under fine day, cloudy day, rainy day, haze weather, and morning, noon, under dusk illumination condition, repeat step 11;

Step 13: put common FOD (at least comprising runway concrete block, spanner, tire debris, fuel tank cap, plastic tube, metallic element six type objects) on runway, the diameter of each object is all not less than minimum detection of a target diameter (being generally 3cm), repeat step 11,12 shooting and image processing method.

Further, wherein step 2 is specially:

Step 21: all coloured images are converted into gray level image, use character description method that each sub-picture is converted into a proper vector respectively, wherein the proper vector of four positive sample image set generations is called four positive proper vector set, and the proper vector that negative sample image collection generates is called negative proper vector set; The character description method used can be SIFT, HOG, SURF;

Step 22: by each positive proper vector set, clustering method is used proper vector to be divided into K cluster respectively, setting pilot lamp cluster is a1, tag line cluster is a2, plate gap cluster is a3, tire trace cluster is a4, then K=a1+a2+a3+a4;

Step 23: to each cluster, calculates the approximate Gaussian model parameter (μ of its proper vector distribution density _k, Φ _k), parameter μ _kfor the mean value of proper vectors all in cluster, parameter Φ _kfor the covariance matrix of proper vectors all in cluster, k is the sequence number of cluster and 1≤k≤K; In feature space, arbitrfary point X and cluster k calculating formula of similarity are:

P _k(X)＝exp(-(X-μ _k) ^TΦ _k ^-1(X-μ _k))

Step 24: to each cluster, if it comprises N ₁individual proper vector, each proper vector X _i(1≤i≤N ₁) represent, calculate the minimum similarity of all proper vectors and this cluster in this cluster if negative proper vector set comprises N2 proper vector, each proper vector Y _j(1≤j≤N ₂) represent, calculate the highest similarity of all proper vectors and this cluster in negative proper vector set cluster k similarity decision threshold T _kcomputing formula be

$T_k=\left\{\begin{array}{cc}{L}_2& L_1\&le;L_2\\ \frac{L_1+L_2}{2}& L_1>L_2\end{array}\right..$

Two, in detection, image detection is carried out by step 3 couple runway FOD.

Image acquisition is carried out to actual runway, suspected target detection is carried out to often opening image edge detection method, then get rid of runway surface object according to the identification parameter in the troubleshooting model of non-FOD target, realize detecting the complete FOD of runway monitor area, concrete steps are:

Step 111: to sub regions each in runway color image shot successively, change the coloured image of shooting into gray level image, uses smoothing and noise-reducing process to gray level image;

Step 112: use Canny edge detection algorithm to generate binary edge image to often opening gray level image, to obtain in image the coordinate set of promising 1 pixel, the pixel being set to 1 has N ₃individual, then set expression is C={ (x ₁, y ₁), (x ₂, y ₂) ... (x _n3, y _n3);

Step 113: newly-built null set D, moves to set D by first coordinate in set C; Repeat traversal set C, if there is the distance of arbitrary coordinate in coordinate to set D to be less than or equal to minW in set C, then the coordinate in set C is transferred to set D, until the arbitrary coordinate in set C is all greater than minW to the arbitrary coordinate distance in set D, or set C is null set; Wherein the value of minW is generally 10 to 30;

Step 114: the institute in statistics set D a little in X-axis maximum coordinates be maxX, min coordinates is minX, institute a little maximum coordinates maxY in Y-axis, min coordinates minY in statistics set D, get with centered by, the square gray level image region that max (maxX-minX, maxY-minY) is the length of side, uses the character description method of step 21 that this image-region is changed into proper vector V; ;

Step 115: the similarity P calculating V and K cluster respectively _k(V), if all P _k(V) all P is met _k(V) <T _k, then think and find FOD object, send FOD and report to the police, and target respective coordinates is provided $(\frac{\min X+\max X}{2},\frac{\min Y+\max Y}{2});$

Step 116: repeat step 112 to 115 successively, until set C is null set, completes the detection to FOD in an image;

Step 117: repeat step 112 to 116 successively, detect until complete to the FOD of all images, realizes detecting the complete FOD of runway monitor area.

The present invention is not limited to aforesaid embodiment.The present invention expands to any new feature of disclosing in this manual or any combination newly, and the step of the arbitrary new method disclosed or process or any combination newly.

Claims (10)

1. a runway FOD image detection method, is characterized in that comprising:

Step 1: pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway are taken pictures, sets up four positive sample image set respectively; Common FOD is taken pictures, sets up a negative sample image collection;

Step 2: the troubleshooting model setting up non-FOD target from four positive sample image set, calculates respectively and adds up the identification parameter that four class runway surface objects are different from negative sample image collection;

Step 3: image acquisition is carried out to actual runway, suspected target detection is carried out to often opening image edge detection method, then according in the troubleshooting model of non-FOD target identification parameter get rid of runway surface object, realize detecting the complete FOD of runway monitor area.

2. a kind of runway FOD image detection method according to claim 1, is characterized in that described step 1 specifically comprises:

Step 11: take pictures to the whole and part of the pilot lamp of runway, tag line, plate gap, tire trace four type objects, every type objects should take at least 10 samples, each sample should at least be taken from 10 different angles; All images are cut off redundant area, only retains the smallest square region comprising subject; Rear image down to 32 × 32 pixel is cut out by all;

Step 12: under fine day, cloudy day, rainy day, haze weather, and morning, noon, under dusk illumination condition, repeat step 11, generate positive sample image set;

Step 13: put common FOD on runway, the diameter of each object is all not less than minimum detection of a target diameter, repeat step 11,12 shooting and image processing method, generate negative sample image collection.

3. a kind of runway FOD image detection method according to claim 2, is characterized in that described minimum detection of a target diameter is for being less than 3cm.

4. a kind of runway FOD image detection method according to claim 1, is characterized in that described step 2 concrete steps comprise:

Step 21: coloured images all in sample image set are converted into gray level image, use character description method that each sub-picture is converted into a proper vector respectively, wherein the proper vector of four positive sample image set generations is called four positive proper vector set, and the proper vector that negative sample image collection generates is called negative proper vector set;

Step 22: by each positive proper vector set, clustering method is used proper vector to be divided into K cluster respectively, setting pilot lamp cluster is a1, tag line cluster is a2, plate gap cluster is a3, tire trace cluster is a4, then K=a1+a2+a3+a4; Wherein the value of a1, a2, a3, a4 is the integer between 1 to 5;

Step 23: to each cluster, calculates the approximate Gaussian model parameter (μ of its proper vector distribution density _k, Φ _k), parameter μ _kfor the mean value of proper vectors all in cluster, parameter Φ _kfor the covariance matrix of proper vectors all in cluster, k is the sequence number of cluster and 1≤k≤K; Arbitrfary point X and cluster k similarity P in feature space _k(X) computing formula is:

P _k(X)＝exp(-(X-μ _k) ^TΦ _k ^-1(X-μ _k))

Step 24: to each cluster, if it comprises N ₁individual proper vector, each proper vector X _irepresent, wherein 1≤i≤N ₁, calculate the minimum similarity of all proper vectors and this cluster in this cluster if negative proper vector set comprises N2 proper vector, each proper vector Y _j, represent, wherein 1≤j≤N ₂, calculate the highest similarity of all proper vectors and this cluster in negative proper vector set cluster k similarity decision threshold T _kcomputing formula be:

$T_k=\{\begin{array}{cc}{L}_2& L_1\&le;L_2\\ \frac{L_1+L_2}{2}& L_1>L_2\end{array}.$

5. a kind of runway FOD image detection method according to claim 4, is characterized in that described character description method is SIFT feature extracting method, HOG feature extracting method or SURF algorithm; Clustering method is K-Mean algorithm, BIRCH algorithm or DBSCAN algorithm.

6. a kind of runway FOD image detection method according to claim 1, is characterized in that described step 3 specifically comprises:

Step 31: to sub regions each in runway color image shot successively, change the coloured image of shooting into gray level image, uses smoothing and noise-reducing process to gray level image;

Step 32: use Canny edge detection algorithm to generate binary edge image to often opening gray level image, to obtain in image the coordinate set of promising 1 pixel, the pixel being set to 1 has N ₃individual, then set expression is C={ (x ₁, y ₁), (x ₂, y ₂) ... (x _n3, y _n3);

Step 33: newly-built null set D, moves to set D by first coordinate in set C; Repeat traversal set C, if there is the distance of arbitrary coordinate in coordinate to set D to be less than or equal to minW in set C, then the coordinate in set C is transferred to set D, until the arbitrary coordinate in set C is all greater than minW to the arbitrary coordinate distance in set D, or set C is null set; Wherein minW unit is pixel, and minW value is 10 to 30;

Step 34: the institute in statistics set D a little in X-axis maximum coordinates be maxX, min coordinates is minX, institute a little maximum coordinates maxY in Y-axis, min coordinates minY in statistics set D, get with centered by, the square gray level image region that max (maxX-minX, maxY-minY) is the length of side, uses the character description method of step 21 that this image-region is changed into proper vector V;

Step 35: the similarity P calculating V and K cluster respectively _k(V), if all P _k(V) all P is met _k(V) <T _k, then think and find FOD object, send FOD and report to the police, and target respective coordinates is provided

Step 36: repeat step 32 to 35 successively, until set C is null set, completes the detection to FOD in an image;

Step 37: repeat step 32 to 36 successively, detect until complete to the FOD of all images, realizes detecting the complete FOD of runway monitor area.

7. a runway FOD imaging detection device, is characterized in that comprising

Non-FOD target troubleshooting model module, for taking pictures to pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway, sets up four positive sample image set respectively; Common FOD is taken pictures, sets up a negative sample image collection; Set up the troubleshooting model of non-FOD target from four positive sample image set, calculate respectively and add up the identification parameter that four class runway surface objects are different from negative sample image collection;

FOD image detection module, image acquisition is carried out to actual runway, carry out suspected target detection to often opening image edge detection method, then according in the troubleshooting model of non-FOD target identification parameter get rid of runway surface object, realize detecting the complete FOD of runway monitor area.

8. a kind of runway FOD imaging detection device according to claim 7, it is characterized in that the troubleshooting model module of described non-FOD target is taken pictures to pilot lamp, tag line, plate gap, tire trace four class runway surface object in desirable runway, set up four positive sample image set respectively; Common FOD is taken pictures, sets up a negative sample image collection detailed process and comprise:

Step 11: take pictures to the whole and part of the pilot lamp of runway, tag line, plate gap, tire trace four type objects, every type objects should take at least 10 samples, each sample should at least be taken from 10 different angles; All images are cut off redundant area, only retains the smallest square region comprising subject; Rear image down to 32 × 32 pixel is cut out by all;

Step 12: under fine day, cloudy day, rainy day, haze weather, and morning, noon, under dusk illumination condition, repeat step 11, generate positive sample image set;

Step 13: put common FOD on runway, the diameter of each object is all not less than minimum detection of a target diameter, repeat step 11,12 shooting and image processing method, generate negative sample image collection.

9. a kind of runway FOD imaging detection device according to claim 7, is characterized in that the troubleshooting model module of described non-FOD target calculates respectively from four positive sample image set and adds up the decision threshold T that four class runway surface objects are different from negative sample image collection _kdetailed process is:

Step 21: coloured images all in sample image set are converted into gray level image, use character description method that each sub-picture is converted into a proper vector respectively, wherein the proper vector of four positive sample image set generations is called four positive proper vector set, and the proper vector that negative sample image collection generates is called negative proper vector set;

Step 22: by each positive proper vector set, clustering method is used proper vector to be divided into K cluster respectively, setting pilot lamp cluster is a1, tag line cluster is a2, plate gap cluster is a3, tire trace cluster is a4, then K=a1+a2+a3+a4, wherein the value of a1, a2, a3, a4 is the integer between 1 to 5;

Step 23: to each cluster, calculates the approximate Gaussian model parameter (μ of its proper vector distribution density _k, Φ _k), parameter μ _kfor the mean value of proper vectors all in cluster, parameter Φ _kfor the covariance matrix of proper vectors all in cluster, k is the sequence number of cluster and 1≤k≤K; Arbitrfary point X and cluster k similarity P in feature space _k(X) computing formula is:

P _k(X)＝exp(-(X-μ _k) ^TΦ _k ^-1(X-μ _k))

Step 24: to each cluster, if it comprises N ₁individual proper vector, each proper vector X _irepresent, wherein 1≤i≤N ₁, calculate the minimum similarity of all proper vectors and this cluster in this cluster if negative proper vector set comprises N2 proper vector, each proper vector Y _j, represent, wherein 1≤j≤N ₂, calculate the highest similarity of all proper vectors and this cluster in negative proper vector set cluster k similarity decision threshold T _kcomputing formula be:

$T_k=\{\begin{array}{cc}{L}_2& L_1\&le;L_2\\ \frac{L_1+L_2}{2}& L_1>L_2\end{array}.$

10. a kind of runway FOD of one according to claim 7 imaging detection device, is characterized in that described FOD image detection module, carries out image acquisition to actual runway, according to the judgment threshold T in the troubleshooting model module of non-FOD target _kthe actual runway image collected is detected, realizes detecting the complete FOD of runway monitor area;

Step 31: to sub regions each in runway color image shot successively, change the coloured image of shooting into gray level image, uses smoothing and noise-reducing process to gray level image;

Step 32: use Canny edge detection algorithm to generate binary edge image to often opening gray level image, to obtain in image the coordinate set of promising 1 pixel, the pixel being set to 1 has N ₃individual, then set expression is C={ (x ₁, y ₁), (x ₂, y ₂) ... (x _n3, y _n3);

Step 33: newly-built null set D, moves to set D by first coordinate in set C; Repeat traversal set C, if there is the distance of arbitrary coordinate in coordinate to set D to be less than or equal to minW in set C, then the coordinate in set C is transferred to set D, until the arbitrary coordinate in set C is all greater than minW to the arbitrary coordinate distance in set D, or set C is null set; Wherein minW unit is pixel, and minW value is 10 to 30;

Step 34: the institute in statistics set D a little in X-axis maximum coordinates be maxX, min coordinates is minX, institute a little maximum coordinates maxY in Y-axis, min coordinates minY in statistics set D, get with centered by, the square gray level image region that max (maxX-minX, maxY-minY) is the length of side, uses the character description method of step 21 that this image-region is changed into proper vector V;

Step 35: the similarity P calculating V and K cluster respectively _k(V), if all P _k(V) all P is met _k(V) <T _k, then think and find FOD object, send FOD and report to the police, and target respective coordinates is provided

Step 36: repeat step 32 to 35 successively, until set C is null set, completes the detection to FOD in an image;

Step 37: repeat step 32 to 36 successively, detect until complete to the FOD of all images, realizes detecting the complete FOD of runway monitor area.