CN104063711A - Corridor vanishing point rapid detection algorithm based on K-means method - Google Patents

Abstract

The invention provides a corridor vanishing point rapid detection algorithm based on a K-means method. The method is characterized by carrying out earlier stage processing comprising down-sampling, gray processing, histogram equalization and Canny edge detection after obtaining image data returned by a robot in real time; carrying out line extraction on the image obtained in the previous step by utilizing a probability Hough transformation algorithm; and finally, classifying the detected lines into four kinds by utilizing the K-means algorithm according to the slope, calculating the mean value of the midpoints of each cluster of lines, then, establishing four lines to replace the lines detected in the previous step by utilizing the slope obtained by cluster and the midpoint mean values, dividing the four lines into two groups randomly, calculating the intersection point respectively, and taking the midpoint of the intersection point as a vanishing point. The method can rapidly detect the position of the vanishing point, and then, the heading direction of the robot is corrected by utilizing the vanishing point, so that the navigation of the robot can be realized. Compared with the conventional vanishing point detection method, the vanishing point rapid detection algorithm in the invention has the advantages of being simple and efficient, good in real-time performance and high in stability.

Description

A kind of corridor end point fast algorithm of detecting based on K-means method

Technical field

The present invention relates to robot autonomous navigation field, be specially a kind of corridor Vanishing Point Detecting Algorithm based on K-means method, can be the accurately real-time image obtaining from robot, detect the position coordinates of end point, and then provide navigation information for the motion of robot.The present invention also can be applicable to microminiature unmanned plane autonomous flight task along corridor in buildings.

Background technology

One group, space parallel lines, on the plane of delineation, imaging has and only has an intersection point, is called end point, and it has comprised the directional information of parallel lines on image, therefore can be used to as robot provides navigation information, and this technology is widely used.When robot moves in corridor, utilize the realtime graphic information obtaining from camera, calculate the coordinate of end point in corridor, constantly revise accordingly the working direction of robot, finally can realize the autokinetic movement in Robot corridor.

In order to use end point to navigate to the motion of robot, Vanishing Point Detecting Algorithm is had to two important requirements: (1) calculated amount is little, and real-time is high; (2) the end point position detecting is accurate.

Popular several Vanishing Point Detecting Algorithm are basic thought mainly with Hough transformation greatly at present, and select different parameter spaces, and main parameter space has Gaussian sphere, hough space and image space.Barnard, Stephen T. " Interpreting perspective images. " Artificial intelligence21.4 (1983): propose to using Gaussian sphere centered by camera optics center in 435-462. as accumulation space.It is a circle that straight line in image is mapped in Gaussian sphere, and end point is corresponding a point in Gaussian sphere.Because the unlimited plane of delineation is mapped to a limited Gaussian sphere space, so the end point of unlimited distance is mapped to a point equally.In the cumulative voting stage of determining end point, Gaussian sphere is divided into a plurality of accumulation junior units, the accumulation unit ballot of each circle to its process, and votes accumulation unit how is corresponding to end point.This method all can effectively detect the end point of limit remote and unlimited distance.Tuytelaars, Tinne, et al. " The cascaded Hough transform as an aid in aerial imageinterpretation. " Computer Vision, 1998.Sixth International Conference on.IEEE, 1998. adopt and the similar thought of Barnard, use Hough transformation that the straight line parameter in plane is converted into hough space, at hough space, carry out cumulative voting, determine end point, reduced to a certain extent calculated amount.Rother, Carsten. " Anew approach to vanishing point detection in architectural environments. " Image and VisionComputing20.9 (2002): 647-655. will detect the candidate target of the intersection point between two of straight line as end point in all planes of delineation, and provided some constraint conditions and determine end point, this method can obtain result more accurately.

But, there is following some deficiency in above-mentioned several method: (1) is in Gaussian sphere space and hough space method, after Gaussian sphere conversion and Hough transformation, reduced the spatial positional information of line segment and end point, simultaneously, the get involved impact of product unit size of the end point precision that these two kinds of methods detect is larger, when accumulation unit is larger, the error that obtains end point is larger, when accumulation unit is too small, calculated amount will be multiplied, can not requirement of real time; (2) when the plane of delineation carries out the detection of end point, all possible straight-line intersection to be calculated, the precision of this method is higher, but when practical application, owing to detecting, straight line quantity is larger, and algorithm complex is high, calculated amount is large, so real-time is poor, cannot meet actual needs.

Summary of the invention

In order to overcome the deficiencies in the prior art, the present invention proposes a kind of corridor end point fast algorithm of detecting based on K-means method, adopt K-means method to carry out vanishing Point Detection Method, only need two intersection points of calculated line, can calculate fast and accurately the position of end point in image.

Technical scheme of the present invention is:

Described a kind of corridor end point fast algorithm of detecting based on K-means method, is characterized in that: adopt following steps:

Step 1: the image gathering is carried out to pre-service, get rid of the redundant information in image, the impact of noise decrease; Preprocessing process is followed successively by: the image gathering is carried out to down-sampled, gray processing processing, histogram equalization, employing Canny operator and carry out rim detection;

Step 2: adopt the image after probability Hough transformation algorithm detecting step 1 is processed, extract the lines in image;

Step 3: the lines cluster that adopts K-means method according to lines slope, step 2 to be extracted is four classes, obtains respectively the average of four class lines mid points; According to every class lines slope and corresponding lines mid point average, determine altogether four straight lines, four straight random are divided into two groups, and every group of two straight lines obtain an intersection point, get the mid point of two intersection points as end point.

Further preferred version is:

Described a kind of corridor end point fast algorithm of detecting based on K-means method, is characterized in that: the lines cluster that adopts K-means method according to lines slope, step 2 to be extracted in step 3 is that the process of four classes is as follows:

Step 3.1: the slope of the n bar lines that calculation procedure 2 detects, and using the n an obtaining slope value as object to be clustered; From n object, select arbitrarily four as initial cluster center; Initial cluster center is for being polymerized to the initial center of four classifications;

Step 3.2: calculate each object to be clustered to the Euclidean distance of all cluster centres, and each clustering object is divided into apart from its nearest cluster centre, as a member of the corresponding classification of this cluster centre;

Step 3.3: recalculate the average of each member in each classification, as the new cluster centre of this classification;

Step 3.4: when each object to be clustered to it under the distance sum of classification cluster centre can not reduce again time, cluster finishes, otherwise returns to step 3.2.

Beneficial effect

The present invention has realized a kind of corridor Vanishing Point Detecting Algorithm of precise and high efficiency, in the image that the method can obtain in robot, the position of end point detected fast, then utilize end point to revise the working direction of robot, realize the navigation to robot.Compare with existing vanishing Point Detection Method method, the method that the present invention proposes has the feature simply efficient, real-time is good, stability is strong.

Why the present invention has its reason of above-mentioned beneficial effect is: the corridor image that first method that the present invention proposes obtains robot carries out pre-service, get rid of and do not have helpful view data to calculating end point, only retain the view data part that determines corridor end point, greatly reduce the quantity of information of required processing, reduced the complexity of calculating.And because probability Hough transformation is a process relatively consuming time, step (1) is removed after a large amount of extraneous data, the lines leaching process of step (2) has also been saved and has been assessed the cost, and the lines that detect are all useful to determining of end point.When definite end point position, do not adopt a large amount of straight lines method of find intersection between two, but take into full account the architectural feature in corridor, adopting K-means method that the straight line detecting in image is gathered is four classes, four straight lines that utilization obtains replace a large amount of straight lines that detect above, and these four straight lines just can determine the position of end point, adopt and have further simplified in this way calculating, final in the situation that only needing two intersection points of calculated line, just can obtain the end point in corridor.

Accompanying drawing explanation

Fig. 1: method flow diagram of the present invention;

Fig. 2: the processing image in embodiment.

Embodiment

Below in conjunction with drawings and Examples, the present invention is further described.

Accompanying drawing 1 has shown that the present invention realizes the overall flow of vanishing Point Detection Method in corridor, and this flow process has comprised each required key step of definite end point.The image information that the object of the invention is to utilize the robot that moves in corridor to return, calculates the position of end point in corridor fast, and utilizes the end point detecting to navigate to the motion of Robot corridor.Accompanying drawing 2 has been shown and is utilized the method that the present invention proposes in the corridor image of a width random shooting, to detect each process of end point.

Be concrete performing step below, the input picture of take in accompanying drawing 2 describes whole process as example.

One, the pre-service in early stage of image

When robot returns after a pictures, first use the method for computer vision to carry out relevant treatment to image, to reduce the interference of noise in surrounding environment, get rid of determining that end point does not have effective view data, and the edge in image is detected, for next step, to extract lines ready.

(1) the down-sampled processing of image

First the image collecting is carried out down-sampled processing, reduce need to process the size of image, in practical operation, we are by the length of image and widely all become original 1/2.This step has reduced needs quantity of information to be processed, has reduced the time cost in calculating process.

(2) gray level image

Because the method for computer vision is easy to be subject to the impact of illumination condition in environment of living in, for this impact is down to minimum, we are converted into gray level image the down-sampled RGB imagery exploitation formula (1) obtaining, and reduce the impact of illumination variation on vanishing Point Detection Method.Obtain image shown in accompanying drawing 2-b.

Gray＝0.229R+0.587G+0.114B (1)

(3) histogram equalization

Histogram equalization is the method for utilizing image histogram to adjust picture contrast in image processing field.The gray level image obtaining for step (2), allows n _irepresent the number of times that gray scale i occurs, the occurrence rate of the pixel that in this sampled images, gray scale is i is

$p_i=\frac{n_i}{n},i\∈\mathrm{0,1}\·\·\·L----\left(2\right)$

In formula, L is grey all in image, and n is gray level image number of pixels altogether, and p is the histogram of image, is normalized to 0 ..., 1.The cumulative probability function that definition c is p, is defined as:

$c\left(i\right)=\underset{j=0}{\overset{i}{\mathrm{\Σ}}}{p}_j---\left(3\right)$

Define a transforming function transformation function that form is y=T (x), each the pixel x in original image, just can produce a y, and the cumulative probability function of y just can carry out linearization within the scope of all values like this, and definition conversion formula is:

y _i＝T(x _i)＝c(i) (4)

T is different grade mappings to 0 ..., 1 territory, for these values are shone upon go back to their initial territories, needs application conversion below in result:

y _i'＝y _i(max-min)+min (5)

Wherein max is the maximal value in step (2) gained gray level image set of pixels, and min is the minimum value of pixel wherein.By histogram equalization, improved the contrast of image, be conducive to the rim detection in image.Obtain image shown in accompanying drawing 2-c.

(4) adopt Canny operator to carry out rim detection

The imagery exploitation Canny operator that step (3) is obtained carries out rim detection.Canny operator carries out rim detection can be summarized as four steps:

1, use Gaussian smoothing function to carry out smoothing processing to image, to eliminate the impact of noise;

The present invention adopts following Gaussian smoothing function to process image:

$H(x,y,\mathrm{\σ})=\frac{1}{\sqrt{2\mathrm{\π}}\mathrm{\σ}}{e}^{\frac{x^2+y^2}{2{\mathrm{\σ}}^2}}---\left(6\right)$

With the image f (x, y) that gaussian kernel function H (x, y, σ) obtains step (3), carry out smoothing processing, can obtain the image g (x, y) after level and smooth, concrete formula is:

g(x,y)＝H(x,y,σ)*f(x,y) (7)

In formula, * represents convolution algorithm.

2, by the finite difference of single order local derviation assign to amplitude and the direction of compute gradient, the object strengthening to reach edge;

The gradient of gradation of image value can be used the first difference of image pixel to assign to be similar to, and so just can obtain image two matrixes of partial derivative in the x and y direction.

In the present invention, we use the convolution operator shown in following table to carry out convolution algorithm to image, and what embodied in following table left side is the difference of image in x direction, and what embodied on right side is the difference of image in y direction.

Obtaining image is respectively at single order partial derivative matrix P and the Q of x direction, y direction:

P(x,y)＝[f(x+1,y)-f(x,y)+f(x+1,y+1)-f(x,y+1)]/2 (8)

Q(x,y)＝[f(x,y+1)-f(x,y)+f(x+1,y+1)-f(x+1,y)]/2 (9)

Obtain amplitude and the deflection of gradient, the mathematic(al) representation of gradient magnitude M and gradient direction θ is simultaneously:

$M(x,y)=\sqrt{P{(x,y)}^2+Q{(x,y)}^2}---\left(10\right)$

θ(x,y)＝arctan(P(x,y)/Q(x,y)) (11)

M (x, y) has reflected the intensity at edge, and θ (x, y) has reflected the direction at edge.Making M (x, y) obtain the deflection θ (x, y) of local maximum, is just the direction at this edge.

By this step, obtain the image being reinforced in marginal position place feature.

3, the peaked point of partial gradient in computed image, weeds out the point at most non-edge with this;

For each pixel in image, make its picture plain M of value (x, y) compare with two pixel values along gradient line, if the gray-scale value of central pixel point is large unlike the gray-scale value of two adjacent image points along gradient line, making central pixel point gray-scale value is 0, otherwise remains unchanged.This operation has retained the point of partial gradient value maximum, and the pixel value of other point has all been set to 0.

4, carry out bilinearity threshold value, obtain the marginal point of image;

To the peaked image of partial gradient value that retained obtained in the previous step, select two threshold alpha ₁and α ₂, α wherein ₁for low threshold value, α ₂for high threshold.We are less than α by Grad ₁the gray-scale value of pixel be set to 0, obtain image 1, then Grad be less than to α ₂the gray-scale value of pixel be set to 0, obtain image 2.Because the threshold value of image 2 is higher, removed most of noise, but also lost useful marginal information simultaneously.And the threshold value of image 1 is lower, retained more information, we take image 2 as basis, take image 1 as supplementing to link the marginal point of image.Concrete edge Connection Step is as follows: first image 2 is scanned, when running into the non-zero pixel p (x, y) of gray-scale value, follow the tracks of and take the outline line that p (x, y) is starting point, until the terminal q (x, y) of outline line.Then, in image under consideration 1 with image 2 in 8 adjacent domains of some s (x, y) corresponding to q (x, y) some position.If have non-zero pixels s (x, y) to exist in 8 adjacent domains of s (x, y) point, be included in image 2, as r (x, y) point.From r (x, y), start, continue search forward, until we cannot continue in image 1 and image 2.After completing the link of the outline line to comprising p (x, y), this outline line is labeled as and is accessed.Other edge is found in continuation in image, until can not find new edge.

In the present invention by the height threshold alpha of Canny algorithm ₁and α ₂ratio be made as 5:2.Obtain image as shown in accompanying drawing 2-d.

The detected edge of this step is the basis of extracting lines from image, and has reduced the hunting zone of next step probability Hough transformation, makes it to existing the place of marginal point to search for, reduce time cost.

Two, the extraction of lines in image

The picture that robot passes back in real time, through above step process to have removed correlation noise, reduced quantity of information to be dealt with simultaneously, below just can carry out the detection of lines.The present invention adopts probability Hough transformation algorithm to carry out the detection of lines in image, and this algorithm is highly suitable for certain real-time system of response time.

Straight line on the plane of delineation can be expressed as under polar coordinate system: ρ=xcos θ+ysin θ, and in formula, ρ is that initial point is to the distance of straight line, θ is the angle between ρ and x axle, and on a known straight line, ρ and θ are constant, (x, y) is the coordinate of any point on straight line.A point (x, y) in image space, is mapped to the parameter space of ρ and θ, just becomes a curve.Straight line in image space, is mapped to ρ, θ parameter space, just changes many curves that intersect at a point into, therefore, finds the operation of straight line in image space, is just converted into the problem of searching the intersection point of curve in ρ, θ parameter space.

The concrete steps of probability Hough transformation algorithm are as follows:

1, by the ρ under polar coordinates, the discrete minizone that turns to a plurality of equalizations of θ parameter space, each interval corresponding totalizer A (ρ, θ), its initial value is zero; Then all marginal points of previous step Canny rim detection gained are put into pending edge point set;

Whether be empty, if it is algorithm finishes if 2, detecting pending edge point set; Otherwise from pending marginal point, concentrate and get a pixel at random, be projected to ρ, θ parameter space obtains a curve, all minizones to this Curves process, calculate the corresponding θ value of ρ value in each minizone, and totalizer A (ρ, θ) corresponding to each minizone added to 1;

3, from pending marginal point, concentrate and delete the point of getting;

4, whether totalizer A (ρ, the θ) value after judgement renewal is greater than predefined threshold value threshold, if be less than, gets back to second step; Otherwise go to next step;

If totalizer A (ρ in 54, θ) value is greater than predefined threshold value threshold, and explanation has many curves through these minizones, when 1 interval of dividing hour, can think that these many curve intersections are in a bit, this point is interval enough hour corresponding point.So ground of equal value can obtain, these pass through ρ of this minizone, the curve of θ parameter space, and what in image (x, y) plane, form is a line segment, on this line segment, the number of pixel is the curve number through this minizone, and both are at least threshold value threshold.Our record detects the starting point of line segment and terminal as the record of a line segment.

Then delete the concentrated point on this line segment, this totalizer zero clearing of being positioned at of pending marginal point;

6, return the 2nd step.

Image with probability Hough transformation algorithm after to Canny rim detection is processed, and can detect the straight line that can determine end point in image, and every straight line is determined by 2 on its of record.Obtain image shown in accompanying drawing 2-e.

Three, determine the end point in image

The straight line detecting according to probability Hough transformation algorithm, can determine in two steps the position of end point:

(1) end point in corridor is by determined along the straight line of corridor direction, and the straight line being the most easily detected in corridor is the straight line that vertical wall and ground and ceiling intersect formation, by to these straight lines the projection in the plane of delineation analyze, find all along plane of delineation diagonal, to distribute on these straight basics, roughly can be divided into four bunches, its slope also can be divided into four classes.In the straight line detecting by step 2, these straight lines account for the overwhelming majority, and therefore, it is four classes that the straight line that utilizes K-means algorithm that step 2 is detected gathers.

K-means algorithm is a kind of widely used clustering algorithm, n object is divided into the class number of appointment, so that have higher similarity in class, and similarity between class is as far as possible little.The calculating of similarity is carried out according to the mean value of object in each class.In the present invention, use this algorithm steps as follows:

The slope of the n bar straight line that 1. calculation procedure two detects, and using the n an obtaining slope value as object to be clustered, then from this n object, select arbitrarily four as initial cluster center, namely to be polymerized to the initial center of four classifications;

2. calculate each object to be clustered to the Euclidean distance of all cluster centres; And each clustering object is divided into apart from its nearest cluster centre, as such other member;

3. recalculate the average of each classification, as the new cluster centre of this classification;

4. when the distance sum of each object to be clustered to its affiliated classification cluster centre can not reduce again, when algorithm has been restrained, stop; Otherwise get back to step 2..

(2) utilize the result of cluster, calculate respectively the average of every bunch of straight line mid point, utilize the cluster straight slope obtaining and the mid point average of obtaining to build four straight lines according to point slope form, utilize these four straight lines to replace the straight line detecting above.Four straight random are divided into two groups, and every group of two straight lines all can obtain an intersection point, and the mid point of getting two intersection points is end point.

The end point detecting in input picture is as shown in accompanying drawing 2-f.

In conjunction with first's algorithm of the present invention, be summarized as follows:

(1) after obtaining the view data that robot returns in real time, first it is carried out to processing in early stage, comprise down-sampled, gray processing processing, histogram equalization and Canny rim detection;

(2) previous step is obtained to imagery exploitation probability Hough transformation algorithm and carry out the extraction of lines;

(3) utilizing K-means algorithm to gather according to slope the straight line detecting is four classes, calculate the average of every bunch of straight line mid point, then utilize slope and the mid point average that cluster obtains to build the straight line that four straight lines replace previous step to detect, four straight random are divided into two groups, calculate respectively its intersection point, get the mid point of intersection point as end point.

Evaluation index

In order to evaluate the result of vanishing Point Detection Method performance of the present invention and comparative experiments, by following index, illustrate: the deviation of end point coordinate and operation time.Wherein the deviation of end point coordinate is defined as: wherein, (x ₁, y ₁) be actual end point coordinate, (x ₂, y ₂) be the end point coordinate that algorithm to be tested (algorithm that uses the present invention to propose) detects.

Effect of the present invention can further illustrate by following emulation experiment:

In order to further illustrate the validity of put forward the methods of the present invention, designed below and verified:

Experimental program: several sizes of random shooting are about the picture of 350x450 in different corridors, utilize the algorithm that the present invention proposes automatically to detect end point to every width picture, and record the time used, then with manual measurement to end point (can be considered actual end point) compare, calculate end point grid deviation, then get its mean value and compare.Select general Vanishing Point Detecting Algorithm to carry out vanishing Point Detection Method: first to use the straight line in the Hough transformation algorithm detected image of standard, then the direct method acquisition end point to intersection between lines point.The end point coordinate that records this method time used and detect, and compare with the result of put forward the methods of the present invention.

Experimental result: by the threshold of Hough transformation in the present invention (when the intersection point number of every zonule curve is greater than threshold, think that parameter corresponding to this region shows as a line segment on the plane of delineation) be made as respectively 30,45 and at 60 o'clock, the experimental result obtaining is as shown in Table 1.

By table one, can be found out, the algorithm that the present invention proposes has very high end point computational accuracy, and average end point deviation all in 15 pixels, and has very low assessing the cost, and can make robot can reserve the more time and carry out other operations.Compare and utilize Hough transformation detection of straight lines, then ask for straight-line intersection as the method for end point, the method that the present invention proposes has very high robustness to choosing of Hough transformation threshold value threshold, when threshold value changes, substantially remain unchanged computing time, in the end point grid deviation detecting remains on very among a small circle.Therefore the method that the present invention proposes has good environmental suitability, has very high practical value.

Table one vanishing Point Detection Method experimental result

Claims (2)

1. the corridor end point fast algorithm of detecting based on K-means method, is characterized in that: adopt following steps:

Step 1: the image gathering is carried out to pre-service, get rid of the redundant information in image, the impact of noise decrease; Preprocessing process is followed successively by: the image gathering is carried out to down-sampled, gray processing processing, histogram equalization, employing Canny operator and carry out rim detection;

Step 2: adopt the image after probability Hough transformation algorithm detecting step 1 is processed, extract the lines in image;

Step 3: the lines cluster that adopts K-means method according to lines slope, step 2 to be extracted is four classes, obtains respectively the average of four class lines mid points; According to every class lines slope and corresponding lines mid point average, determine altogether four straight lines, four straight random are divided into two groups, and every group of two straight lines obtain an intersection point, get the mid point of two intersection points as end point.

2. a kind of corridor end point fast algorithm of detecting based on K-means method according to claim 1, is characterized in that: the lines cluster that adopts K-means method according to lines slope, step 2 to be extracted in step 3 is that the process of four classes is as follows:

Step 3.1: the slope of the n bar lines that calculation procedure 2 detects, and using the n an obtaining slope value as object to be clustered; From n object, select arbitrarily four as initial cluster center; Initial cluster center is for being polymerized to the initial center of four classifications;

Step 3.2: calculate each object to be clustered to the Euclidean distance of all cluster centres, and each clustering object is divided into apart from its nearest cluster centre, as a member of the corresponding classification of this cluster centre;

Step 3.3: recalculate the average of each member in each classification, as the new cluster centre of this classification;

Step 3.4: when each object to be clustered to it under the distance sum of classification cluster centre can not reduce again time, cluster finishes, otherwise returns to step 3.2.