CN105631880A - Lane line segmentation method and apparatus - Google Patents

Abstract

The invention discloses a lane line segmentation method and apparatus. The method comprises: a lane line image is collected; the lane line image is processed by using a convolution neural network to obtain a coarse segmentation result; and according to the coarse segmentation result and the lane line image, an image segmentation model is constructed to carry out subdivision on the lane line image, thereby determine a lane line area. According to the invention, when the image segmentation model is constructed, information of the whole lane line image is used and segmentation of the image segmentation model becomes precise, so that lane line segmentation precision is improved.

Description

Track line dividing method and device

Technical field

The embodiment of the present invention relates to Cartographic Technique, particularly relates to a kind of track line dividing method and device.

Background technology

In natural scene, line segmentation in accurate track can help the generation of graph key traffic key element accurately, is automatic Pilot and the auxiliary main dependence technology driven.

In prior art, the segmentation for track line is generally the method based on image procossing, first orients the track line in the image collected, asks for the image gradient near the line of track, will respond the border of place the strongest as track line.

Prior art utilizes the information of a regional area to ask for image gradient, it is easy to be subject to influence of noise, when there is shade, block or during the problem such as track line is smudgy error bigger; For non-single root track line (such as two-wire, real dotted line etc.), accurate track line cannot be oriented. Therefore, prior art also exists the coarse defect of track line of segmentation.

Summary of the invention

In view of this, the embodiment of the present invention provides a kind of track line dividing method and device, to improve the tolerance range of the track line of segmentation.

First aspect, embodiments provides a kind of track line dividing method, and described method comprises:

Gather track line image;

Utilize convolutional neural networks to be processed by described track line image, obtain coarse segmentation result;

According to described coarse segmentation result and described track line image, design of graphics cuts model and is segmented by described track line image, it is determined that line region, track.

Second aspect, the embodiment of the present invention additionally provides a kind of track line cutting device, and described device comprises:

Track line acquisition module, for gathering track line image;

Coarse segmentation module, for utilizing convolutional neural networks to be processed by described track line image, obtains coarse segmentation result;

Module is cut in segmentation, for according to described coarse segmentation result and described track line image, design of graphics cuts model and segmented by described track line image, it is determined that line region, track.

The technical scheme of the embodiment of the present invention, by after collecting track line image, utilize convolutional neural networks to be processed by described track line image, obtain the coarse segmentation result to track line, then according to coarse segmentation result and track line image, design of graphics cuts model and is segmented by track line image, determine line region, track, due to when design of graphics cuts model, it may also be useful to the information of whole track line image, and figure cuts the more accurate of model segmentation, it is to increase the tolerance range of the track line of segmentation.

Accompanying drawing explanation

Fig. 1 is the schema of a kind of track line dividing method that the embodiment of the present invention one provides;

Fig. 2 is the example schematic network structure of the convolutional neural networks in the track line dividing method of embodiment of the present invention offer;

Fig. 3 is the example image of the track line image used in the track line dividing method of embodiment of the present invention offer;

Track line image in Fig. 3 is utilized convolutional neural networks to identify the image of the coarse segmentation result obtained in the embodiment of the present invention by Fig. 4;

Fig. 5 is the schema of a kind of track line dividing method that the embodiment of the present invention two provides;

Fig. 6 is the schema of a kind of track line dividing method that the embodiment of the present invention three provides;

Fig. 7 is the structural representation of a kind of track line cutting device that the embodiment of the present invention four provides.

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in further detail. It should be appreciated that specific embodiment described herein is only for explaining the present invention, but not limitation of the invention. It also should be noted that, for convenience of description, accompanying drawing illustrate only part related to the present invention and not all content.

Track line dividing method described in the embodiment of the present invention and device, the positioning result based on track line is carried out, and the localization method of track line has a lot, specifically how to position not within the scope of the discussion of the embodiment of the present invention.

Embodiment one

Fig. 1 is the schema of a kind of track line dividing method that the embodiment of the present invention one provides, the present embodiment is applicable in the situation splitting track line wherein according to the track line image navigated to, the method can be performed by track line cutting device, described track line cutting device can be integrated in the terminal such as computer or mobile terminal, specifically comprises as follows:

S110, gathers track line image.

When navigating to track line, by control camera collection track line image.

S120, utilizes convolutional neural networks to be processed by described track line image, obtains coarse segmentation result.

Wherein, convolutional neural networks (ConvolutionalNeuralNetwork, CNN) is a kind of feedforward neural network, and its artificial neuron unit can respond the surrounding cells in part coverage, has outstanding performance for large-scale image procossing. The pixel contact that the space relationship of image is also local is comparatively tight, and distant pixel interdependence is then more weak. Thus, each neurone there is no need global image is carried out perception in fact, it is only necessary to local is carried out perception, then comprehensively gets up just to obtain the information of the overall situation in the more high-rise information by local.

Fig. 2 is the example schematic network structure of the convolutional neural networks in the track line dividing method of embodiment of the present invention offer, as shown in Figure 2, convolutional neural networks is the neural network of a multilayer, every layer is made up of multiple two dimension plane, and each plane is made up of multiple independent neurone, input picture is by carrying out convolution with three wave filters that can train with being biased, three Feature Mapping figure are produced at convolutional layer C1 after convolution, then four pixels often organized in Feature Mapping figure are sued for peace again, weighted value, it is biased, the Feature Mapping figure of three pond layer S2 is obtained by a Sigmoid function. these mapping graphs entered filtering again and obtained convolutional layer C3. this level structure is the same with pond layer S2 again produces pond layer S4. finally, these pixels value is rasterized, and connects into a vector and be input to traditional neural network B5, obtains Output rusults.

A kind of typical convolutional neural networks structure be " input picture-convolutional layer-pond layer-convolutional layer-pond layer-... the articulamentum-Output rusults of the articulamentum of-pond layer-entirely-entirely ". Wherein, convolutional layer passes through convolution algorithm, it is possible to the signal feature in original image is strengthened, and reduces noise; Pond layer utilizes the principle of image local correlation, and image is carried out sub-sampling, it is possible to reduce data processing amount remain with by information simultaneously; Full articulamentum is for learning the feature of local and the overall situation.

Before utilizing convolutional neural networks to be processed by described track line image, utilize a large amount of samples that convolutional neural networks is carried out training study, obtain the convolutional neural networks trained, in convolutional neural networks line image input in track trained again, Output rusults is just coarse segmentation result, in coarse segmentation result, the codomain of each pixel is 0-1, represent the possibility belonging to track line, more close to 1, represent that the probability belonging to track line is more big.

S130, according to described coarse segmentation result and described track line image, design of graphics cuts model and is segmented by described track line image, it is determined that line region, track.

Wherein, figure cuts model and namely utilizes figure to cut (GraphCut) algorithm to realize Iamge Segmentation. Iamge Segmentation can regard element marking problem as, and the mark of target is set to 1, and the mark of background is set to 0, and this process can be cut minimization of energy function by minimumization figure and be obtained, it is intended that separated by object and background. Figure cuts algorithm only to be needed respectively to draw strokes as input in prospect (target) and background place, and algorithm will set up each pixel tax weight graph similar to prospect background, and minimum cuts differentiation prospect and background by solving. This algorithm is the method based on Color Statistical sampling, and the image effect that prospect background difference is bigger is better.

According to described coarse segmentation result, determine the absolute foreground area of track line image and absolute background area, and utilize figure to cut algorithm design of graphics and cut model, track line image is segmented, obtain the tax weight graph that each pixel in the line chart of track is similar to prospect background, and then according to weights, prospect and background segment are opened, it is determined that prospect be line region, track.

The technical scheme of the present embodiment, by after collecting track line image, utilize convolutional neural networks to be processed by described track line image, obtain the coarse segmentation result to track line, then according to coarse segmentation result and track line image, design of graphics cuts model and is segmented by track line image, determine line region, track, due to when design of graphics cuts model, it may also be useful to the information of whole track line image, and figure cuts the more accurate of model segmentation, it is to increase the tolerance range of the track line of segmentation.

On the basis of technique scheme, utilize convolutional neural networks to be processed by described track line image, obtain coarse segmentation result and preferably include:

Described track line image is divided into the square image blocks of same size;

Extract the view data of described square image blocks in color space model, and described view data is normalized;

The convolutional neural networks that view data input after normalization method has been trained is identified, obtains recognition result;

The image that described recognition result is normalized to size same with described square image blocks, obtains coarse segmentation result.

First, the track line image of entirety being carried out piecemeal, be divided into the size of every block to be the square image blocks of M �� M pixel, wherein, M can be arranged according to the network configuration of convolutional neural networks; For a square image blocks, extract the data of multicolour space model, comprise at least two in RGB, HSV, YCrCb and Lab etc. kind, undertaken merging (namely piecing together) by the channel data of these the different color spaces extracted, and under normalizing to identical size; The convolutional neural networks that view data input after the normalization method of each square image blocks has been trained is identified, the size of the recognition result obtained is N �� N pixel, in order to avoid producing poor fitting problem, general requirement N �� N is less than M �� M, wherein, N is determined by the network configuration of convolutional neural networks; Described recognition result is normalized to M �� M pixel, each pixel represents the possibility belonging to track line, codomain is 0-1, more close to 1, represents that the probability belonging to track line is more big, predetermined threshold value can be set, when the value of pixel is more than or equal to predetermined threshold value, it is determined that this pixel belongs to track line, after the recognition result of each square image blocks is normalized to M �� M pixel, according to original sequential concatenation to together, obtaining coarse segmentation result. But owing to, during to recognition result normalization method, having carried out interpolation, so coarse segmentation result resolving power is not enough, therefore, in addition it is also necessary to coarse segmentation result is segmented. As shown in Figure 3 and Figure 4, Fig. 3 is the example image of the track line image used in the track line dividing method of embodiment of the present invention offer, and the track line image in Fig. 3 is utilized convolutional neural networks to identify the image of the coarse segmentation result obtained in the embodiment of the present invention by Fig. 4. Wherein, when track line image is continued piecemeal, when can not form square image blocks to last remaining image, it is possible to a part overlapping with previous image block so that with remaining image sets dreit) image block.

Wherein, rgb color space model adopts R (Red, red), G (Green, green) and B (Blue, blue) three primary colours represent a kind of color, each color can be added in different ratios by R, G, B three primary colours and mix; HSV color space model adopts conical space model to describe, wherein, H (Hue, tone) measures by angle, span is 0 �㡫360 ��, S (Saturation, saturation ratio) is a ratio value, and scope is from 0 to 1, its represents the ratio between the purity of selected color and the maximum purity of this color, V (Value, brightness) represents the light levels of color, and scope is from 0 to 1; YCrCb color space and YUV color space, it is mainly used in optimizing the transmission of colourvideo signal, make its compatible old-fashioned monochrome television backward, wherein, " Y " represents lightness (Luminance or Luma), is also exactly grey decision-making, and that " U " and " V " represents is then colourity (Chrominance or Chroma), effect describes colors of image and saturation ratio, is used to specify the color of pixel; Lab color space model a, b tri-key elements by lightness (L) with about color form, L represents lightness (Luminosity), a represents the scope from carmetta to green, b represents the scope from yellow to blueness, the codomain of L by be just equivalent to when 0 to 100, L=50 50% black, the codomain of a and b is all by+127 to-128, wherein+127a is exactly red, just becomes green when being gradually transitioned into-128a; Same principle ,+127b is yellow, and-128b is blue.

Embodiment two

Fig. 5 is the schema of a kind of track line dividing method that the embodiment of the present invention two provides, the present embodiment in embodiment one according to described coarse segmentation result and described track line image, design of graphics cuts model and is segmented by described track line image, determine that line region, track is optimized, specifically comprise as follows:

S510, gathers track line image.

S520, utilizes convolutional neural networks to be processed by described track line image, obtains coarse segmentation result.

S530, according to described coarse segmentation result, it is determined that absolute foreground area, definitely background area and the uncertain region in the line image of described track.

This step mainly realizes the initialize that figure cuts model, namely by the process to described coarse segmentation result, point out in the line image of described track, which pixel belongs to absolute foreground area (track line), which pixel belongs to absolute background area (non-track line), and which pixel belongs to uncertain region.

Wherein, according to described coarse segmentation result, it is determined that absolute foreground area, definitely background area and uncertain region in the line image of described track preferably include:

Described coarse segmentation result is carried out binary conversion treatment, obtains binary image;

Described binary image is carried out corrosion operation, it is determined that the absolute foreground area of track line image;

Described binary image is carried out expansive working, it is determined that the absolute background area of track line image;

Using the uncertain region of region except described absolute foreground area and absolute background area in the line image of described track as track line image.

First described coarse segmentation result is carried out binary conversion treatment, obtain binary image I; Described binary image I is carried out corrosion operation, the result I after being corroded_erode, by the result I after binary image I and corrosion_erodeCarrying out and computing, its result is absolute foreground area, can determine the absolute foreground area of track line image by the location of pixels of this absolute foreground area; Described binary image I is carried out expansive working, the result I after being expanded_dilate, by the result I after binary image I and expansion_dilateCarrying out or computing, its result is absolute background area, can determine the absolute background area of track line image by the location of pixels of this absolute background area; Using the uncertain region of region except described absolute foreground area and absolute background area in the line image of described track as track line image, see following formula:

Wherein, I_maskFor the mark to track line image, fg is the absolute foreground area of track line image, and bg is the absolute background area of track line image, and unknown is the uncertain region of track line image, and I is binary image, I_erodeFor the result after corrosion, I_dilateFor the result after expansion.

S540, according to described absolute foreground area, definitely background area and described track line image, design of graphics cuts model and solves, and obtains thin segmentation result.

Absolute foreground area can being labeled as 1, namely definitely the pixel of foreground area belongs to the probability of prospect is 1; Absolute background area is labeled as 0, and namely definitely the pixel of background area belongs to the probability of prospect is 0. According to described absolute foreground area, definitely background area and described track line image, the figure building track line image cuts model, this figure cuts model and represents by the energy function of track line image, figure is cut model solve, namely the pixel when value solving energy function reaches minimum belongs to the probability of prospect, the probability belonging to prospect according to pixel just can determine whether pixel belongs to prospect, namely whether belongs to line region, track, and this result is thin segmentation result.

Wherein, according to described absolute foreground area, definitely background area and described track line image, design of graphics cuts model and solves, and obtains thin segmentation result and preferably includes:

The figure building described track line image according to following formula cuts model:

$\begin{array}{c}\underset{b}{\arg \min }E=E_{data}+E_{smoothness}\\ =\underset{p=\left(x,y\right)}{\mathrm{\Σ}}{\mathrm{\ω}}_p{\mathrm{\α}}_p+\underset{q\∈N_p}{\mathrm{\Σ}}{\mathrm{\ω}}_{pq}^c\left|\right|b_q-b_p\left|\right|\end{array}$

Wherein, b is the probability that the pixel in the line image of track belongs to prospect, and E is the energy function of described track line image, E_dataFor the data item in the energy function of track line image, E_smoothnessFor the level and smooth item in the energy function of track line image, ��_pFor pixel p belongs to the probability of prospect, ��_pFor pixel p belongs to the weight of prospect, (x, y) is the coordinate of pixel p in the line image of track,For the weight of the color distortion of neighbor pixel p and q, N_pFor the neighborhood of pixel p, b_pFor pixel p belongs to the probability of prospect, b_qFor pixel q belongs to the probability of prospect, wherein,c_pFor the value of color of pixel p, c_qFor the value of color of pixel q, ��_cFor default weighting threshold value;

Build the gauss hybrid models of definitely foreground area and absolute background area in the line image of track respectively, determine that pixel p belongs to the probability �� of prospect according to following formula_pAnd weights omega_p:

��_p=GMM_fg/(GMM_fg+GMM_bg)

Wherein, GMM_fgFor the gauss hybrid models of foreground area absolute in the line image of track, GMM_bgFor the gauss hybrid models of background area absolute in the line image of track, fg is the absolute foreground area in the line image of track, and bg is the absolute background area in the line image of track;

Utilizing max-flow min-cut algorithm that the figure of track line image is cut model to solve, the pixel obtained in the line image of track belongs to the probability b of prospect, when the probability that pixel belongs to prospect is greater than predetermined probabilities threshold value, it is determined that this pixel belongs to prospect;

When at least two neighbor pixels belong to prospect, form connected domain.

Wherein, gauss hybrid models accurately quantizes things with Gaussian probability density function (normal distribution curve), and a things is decomposed into some models formed based on Gaussian probability density function (normal distribution curve). The fundamental theorem of max-flow min-cut algorithm is the max-flow that the minimum energy sum cut equals in network. Figure cuts model and image segmentation problem is associated with the minimum problem of cutting of figure, therefore by max-flow min-cut algorithm, figure can be cut model and solve.

S550, it is determined that the largest connected territory in described thin segmentation result is line region, track.

Relatively the connected domain in described thin segmentation result, finds out largest connected territory wherein, this largest connected territory is defined as line region, track.

The technical scheme of the present embodiment, by utilizing convolutional neural networks track line image carried out after process obtains coarse segmentation result, according to this coarse segmentation result, determine the absolute foreground area in the line image of track, definitely background area and uncertain region, design of graphics cuts model and solves, obtain thin segmentation result, it is determined that the largest connected territory in described thin segmentation result is line region, track, further increase the tolerance range of the track line of segmentation.

Embodiment three

Fig. 6 is the schema of a kind of track line dividing method that the embodiment of the present invention three provides, and embodiment one is optimized by the present embodiment, on the basis of embodiment one, adds the content that the edge point to line region, track verifies, specifically comprises as follows:

S610, gathers track line image.

S620, utilizes convolutional neural networks to be processed by described track line image, obtains coarse segmentation result.

S630, according to described coarse segmentation result and described track line image, design of graphics cuts model and is segmented by described track line image, it is determined that line region, track.

S640, verifies the edge point in line region, track, it is determined that the profile of track line.

Utilize RANSAC (RANdomSAmpleConsensus, stochastic sampling is consistent) the edge point in line region, track verifies by algorithm, reject wild point, namely the pixel in line region, track it is judged as and really belongs to line region, track, so that it is determined that accurate line region, track, the edge in this line region, track constitutes the profile of track line. Wherein, RANSAC algorithm is the sample data set comprising abnormal data according to a group, calculates the mathematical model parameter of data, obtains the algorithm of effective sample data.

Wherein, the edge point in line region, track is verified, it is determined that the profile of track line preferably includes:

Gather two edge points at the same edge being positioned at line region, track;

Described two edge points are carried out fitting of a straight line;

Calculate the number of the edge point meeting fitting of a straight line result;

Determine that the fitting of a straight line result that described number is maximum is the profile of track line.

First random acquisition is positioned at two edges point (i.e. edge pixel) at the same edge in line region, track, a straight line can be determined by these two edge points, namely fitting of a straight line result is obtained, calculate the number of the edge point in fitting of a straight line result, the edge point at this edge being positioned at line region, track is all carried out above-mentioned process, wild point is affirmed than being all that the number of edge points in the fitting of a straight line result that becomes of group of edge points is few with the number of the edge point in the fitting of a straight line result of edge point (or wild point), therefore, wild point can be rejected by the method, select the fitting of a straight line result that described number is maximum, using the profile of this fitting of a straight line result as track line. the profile of the track line determined like this is more accurate.

The technical scheme of the present embodiment, by after collecting track line image, convolutional neural networks is utilized to be processed by described track line image, obtain the coarse segmentation result to track line, again according to coarse segmentation result and track line image, design of graphics cuts model and is segmented by track line image, determine line region, track, the edge point in line region, track is verified, determine the profile of track line, due to when design of graphics cuts model, employ the information of whole track line image, and figure cuts the more accurate of model segmentation, improve the tolerance range of the track line of segmentation, compared with embodiment one, the present embodiment is it may also be determined that go out the profile of more accurate track line. show through practice, the present embodiment when track line exist smudgy, block or other interference, the accurate segmentation of track line can both be completed.

Embodiment four

Fig. 7 is the structural representation of a kind of track line cutting device that the embodiment of the present invention four provides, and as shown in Figure 7, the track line cutting device described in the present embodiment comprises: module 730 is cut in track line acquisition module 710, coarse segmentation module 720 and segmentation.

Wherein, track line acquisition module 710 is for gathering track line image;

For utilizing, described track line image is processed coarse segmentation module 720 by convolutional neural networks, obtains coarse segmentation result;

Segmentation cuts module 730 for according to described coarse segmentation result and described track line image, design of graphics cuts model and segmented by described track line image, it is determined that line region, track.

Preferably, described coarse segmentation module 720 comprises:

Divide module unit, for described track line image being divided into the square image blocks of same size;

Image data extraction unit, for extracting the view data of described square image blocks in color space model, and is normalized described view data;

Recognition unit, identifies for the view data after normalization method being inputted in the convolutional neural networks trained, obtains recognition result;

Coarse segmentation unit, for described recognition result is normalized to the image of size same with described square image blocks, obtains coarse segmentation result.

Preferably, described segmentation is cut module 730 and is comprised:

Initialization unit, for according to described coarse segmentation result, it is determined that absolute foreground area, definitely background area and the uncertain region in the line image of described track;

Thin cutting unit, for according to described absolute foreground area, definitely background area and described track line image, design of graphics cuts model and solves, and obtains thin segmentation result;

Track line area determination unit, the largest connected territory for determining in described thin segmentation result is line region, track.

Preferably, described initialization unit specifically for:

Described coarse segmentation result is carried out binary conversion treatment, obtains binary image;

Described binary image is carried out corrosion operation, it is determined that the absolute foreground area of track line image;

Described binary image is carried out expansive working, it is determined that the absolute background area of track line image;

Using the uncertain region of region except described absolute foreground area and absolute background area in the line image of described track as track line image.

Preferably, described thin cutting unit specifically for:

The figure building described track line image according to following formula cuts model:

$\begin{array}{c}\underset{b}{\arg \min }E=E_{data}+E_{smoothness}\\ =\underset{p=\left(x,y\right)}{\mathrm{\Σ}}{\mathrm{\ω}}_p{\mathrm{\α}}_p+\underset{q\∈N_p}{\mathrm{\Σ}}{\mathrm{\ω}}_{pq}^c\left|\right|b_q-b_p\left|\right|\end{array}$

Wherein, b is the probability that the pixel in the line image of track belongs to prospect, and E is the energy function of described track line image, E_dataFor the data item in the energy function of track line image, E_smoothnessFor the level and smooth item in the energy function of track line image, ��_pFor pixel p belongs to the probability of prospect, ��_pFor pixel p belongs to the weight of prospect, (x, y) is the coordinate of pixel p in the line image of track,For the weight of the color distortion of neighbor pixel p and q, N_pFor the neighborhood of pixel p, b_pFor pixel p belongs to the probability of prospect, b_qFor pixel q belongs to the probability of prospect, wherein,c_pFor the value of color of pixel p, c_qFor the value of color of pixel q, ��_cFor default weighting threshold value;

Build the gauss hybrid models of definitely foreground area and absolute background area in the line image of track respectively, determine that pixel p belongs to the probability �� of prospect according to following formula_pAnd weights omega_p:

��_p=GMM_fg/(GMM_fg+GMM_bg)

Wherein, GMM_fgFor the gauss hybrid models of foreground area absolute in the line image of track, GMM_bgFor the gauss hybrid models of background area absolute in the line image of track, fg is the absolute foreground area in the line image of track, and bg is the absolute background area in the line image of track;

Utilizing max-flow min-cut algorithm that the figure of track line image is cut model to solve, the pixel obtained in the line image of track belongs to the probability b of prospect, when the probability that pixel belongs to prospect is greater than predetermined probabilities threshold value, it is determined that this pixel belongs to prospect;

When at least two neighbor pixels belong to prospect, form connected domain.

Preferably, this track line cutting device also comprises:

Track line profile determination module, for, after determining line region, track, verifying the edge point in line region, track, it is determined that the profile of track line.

Preferably, described track line profile determination module specifically for:

Gather two edge points at the same edge being positioned at line region, track;

Described two edge points are carried out fitting of a straight line;

Calculate the number of the edge point meeting fitting of a straight line result;

Determine that the fitting of a straight line result that described number is maximum is the profile of track line.

The said products can perform the track line dividing method that any embodiment of the present invention provides, and possesses manner of execution corresponding function module and useful effect.

Note, above are only the better embodiment of the present invention and institute's application technology principle. It is understood by those skilled in the art that and the invention is not restricted to specific embodiment described here, various obvious change can be carried out for a person skilled in the art, readjust and substitute and protection scope of the present invention can not be departed from. Therefore, although being described in further detail invention has been by above embodiment, but the present invention is not limited only to above embodiment, when not departing from present inventive concept, other equivalence embodiments more can also be comprised, and the scope of the present invention is determined by appended right.

Claims (14)

1. a track line dividing method, it is characterised in that, described method comprises:

Gather track line image;

Utilize convolutional neural networks to be processed by described track line image, obtain coarse segmentation result;

According to described coarse segmentation result and described track line image, design of graphics cuts model and is segmented by described track line image, it is determined that line region, track.

2. method according to claim 1, it is characterised in that, utilize convolutional neural networks to be processed by described track line image, obtain coarse segmentation result and comprise:

Described track line image is divided into the square image blocks of same size;

Extract the view data of described square image blocks in color space model, and described view data is normalized;

The convolutional neural networks that view data input after normalization method has been trained is identified, obtains recognition result;

The image that described recognition result is normalized to size same with described square image blocks, obtains coarse segmentation result.

3. method according to claim 1, it is characterised in that, according to described coarse segmentation result and described track line image, design of graphics cuts model and is segmented by described track line image, it is determined that line region, track comprises:

According to described coarse segmentation result, it is determined that absolute foreground area, definitely background area and the uncertain region in the line image of described track;

According to described absolute foreground area, definitely background area and described track line image, design of graphics cuts model and solves, and obtains thin segmentation result;

The largest connected territory determined in described thin segmentation result is line region, track.

4. method according to claim 3, it is characterised in that, according to described coarse segmentation result, it is determined that absolute foreground area, definitely background area and uncertain region in the line image of described track comprise:

Described coarse segmentation result is carried out binary conversion treatment, obtains binary image;

Described binary image is carried out corrosion operation, it is determined that the absolute foreground area of track line image;

Described binary image is carried out expansive working, it is determined that the absolute background area of track line image;

Using the uncertain region of region except described absolute foreground area and absolute background area in the line image of described track as track line image.

5. method according to claim 3, it is characterised in that, according to described absolute foreground area, definitely background area and described track line image, design of graphics cuts model and solves, and obtains thin segmentation result and comprises:

The figure building described track line image according to following formula cuts model:

$\begin{array}{c}\underset{b}{\arg \min }E=E_{data}+E_{smoothness}\\ =\underset{p=\left(x,y\right)}{\mathrm{\Σ}}{\mathrm{\ω}}_p{\mathrm{\α}}_p+\underset{q\∈N_p}{\mathrm{\Σ}}{\mathrm{\ω}}_{pq}^c\left|\right|b_q-b_p\left|\right|\end{array}$

Wherein, b is the probability that the pixel in the line image of track belongs to prospect, and E is the energy function of described track line image, E_dataFor the data item in the energy function of track line image, E_smoothnessFor the level and smooth item in the energy function of track line image, ��_pFor pixel p belongs to the probability of prospect, ��_pFor pixel p belongs to the weight of prospect, (x, y) is the coordinate of pixel p in the line image of track,For the weight of the color distortion of neighbor pixel p and q, N_pFor the neighborhood of pixel p, b_pFor pixel p belongs to the probability of prospect, b_qFor pixel q belongs to the probability of prospect, wherein,c_pFor the value of color of pixel p, c_qFor the value of color of pixel q, ��_cFor default weighting threshold value;

Build the gauss hybrid models of definitely foreground area and absolute background area in the line image of track respectively, determine that pixel p belongs to the probability �� of prospect according to following formula_pAnd weights omega_p:

��_p=GMM_fg/(GMM_fg+GMM_bg)

Wherein, GMM_fgFor the gauss hybrid models of foreground area absolute in the line image of track, GMM_bgFor the gauss hybrid models of background area absolute in the line image of track, fg is the absolute foreground area in the line image of track, and bg is the absolute background area in the line image of track;

Utilizing max-flow min-cut algorithm that the figure of track line image is cut model to solve, the pixel obtained in the line image of track belongs to the probability b of prospect, when the probability that pixel belongs to prospect is greater than predetermined probabilities threshold value, it is determined that this pixel belongs to prospect;

When at least two neighbor pixels belong to prospect, form connected domain.

6. according to the arbitrary described method of claim 1-5, it is characterised in that, after determining line region, track, also comprise:

The edge point in line region, track is verified, it is determined that the profile of track line.

7. method according to claim 6, it is characterised in that, the edge point in line region, track is verified, it is determined that the profile of track line comprises:

Gather two edge points at the same edge being positioned at line region, track;

Described two edge points are carried out fitting of a straight line;

Calculate the number of the edge point meeting fitting of a straight line result;

Determine that the fitting of a straight line result that described number is maximum is the profile of track line.

8. a track line cutting device, it is characterised in that, described device comprises:

Track line acquisition module, for gathering track line image;

Coarse segmentation module, for utilizing convolutional neural networks to be processed by described track line image, obtains coarse segmentation result;

Module is cut in segmentation, for according to described coarse segmentation result and described track line image, design of graphics cuts model and segmented by described track line image, it is determined that line region, track.

9. device according to claim 8, it is characterised in that, described coarse segmentation module comprises:

Divide module unit, for described track line image being divided into the square image blocks of same size;

Image data extraction unit, for extracting the view data of described square image blocks in color space model, and is normalized described view data;

Recognition unit, identifies for the view data after normalization method being inputted in the convolutional neural networks trained, obtains recognition result;

Coarse segmentation unit, for described recognition result is normalized to the image of size same with described square image blocks, obtains coarse segmentation result.

10. device according to claim 8, it is characterised in that, described segmentation is cut module and is comprised:

Initialization unit, for according to described coarse segmentation result, it is determined that absolute foreground area, definitely background area and the uncertain region in the line image of described track;

Thin cutting unit, for according to described absolute foreground area, definitely background area and described track line image, design of graphics cuts model and solves, and obtains thin segmentation result;

Track line area determination unit, the largest connected territory for determining in described thin segmentation result is line region, track.

11. devices according to claim 10, it is characterised in that, described initialization unit specifically for:

Described coarse segmentation result is carried out binary conversion treatment, obtains binary image;

Described binary image is carried out corrosion operation, it is determined that the absolute foreground area of track line image;

Described binary image is carried out expansive working, it is determined that the absolute background area of track line image;

Using the uncertain region of region except described absolute foreground area and absolute background area in the line image of described track as track line image.

12. devices according to claim 10, it is characterised in that, described thin cutting unit specifically for:

The figure building described track line image according to following formula cuts model:

$\begin{array}{c}\underset{b}{\arg \min }E=E_{data}+E_{smoothness}\\ =\underset{p=\left(x,y\right)}{\mathrm{\Σ}}{\mathrm{\ω}}_p{\mathrm{\α}}_p+\underset{q\∈N_p}{\mathrm{\Σ}}{\mathrm{\ω}}_{pq}^c\left|\right|b_q-b_p\left|\right|\end{array}$

Wherein, b is the probability that the pixel in the line image of track belongs to prospect, and E is the energy function of described track line image, E_dataFor the data item in the energy function of track line image, E_smoothnessFor the level and smooth item in the energy function of track line image, ��_pFor pixel p belongs to the probability of prospect, ��_pFor pixel p belongs to the weight of prospect, (x, y) is the coordinate of pixel p in the line image of track,For the weight of the color distortion of neighbor pixel p and q, N_pFor the neighborhood of pixel p, b_pFor pixel p belongs to the probability of prospect, b_qFor pixel q belongs to the probability of prospect, wherein,c_pFor the value of color of pixel p, c_qFor the value of color of pixel q, ��_cFor default weighting threshold value;

Build the gauss hybrid models of definitely foreground area and absolute background area in the line image of track respectively, determine that pixel p belongs to the probability �� of prospect according to following formula_pAnd weights omega_p:

��_p=GMM_fg/(GMM_fg+GMM_bg)

Wherein, GMM_fgFor the gauss hybrid models of foreground area absolute in the line image of track, GMM_bgFor the gauss hybrid models of background area absolute in the line image of track, fg is the absolute foreground area in the line image of track, and bg is the absolute background area in the line image of track;

Utilizing max-flow min-cut algorithm that the figure of track line image is cut model to solve, the pixel obtained in the line image of track belongs to the probability b of prospect, when the probability that pixel belongs to prospect is greater than predetermined probabilities threshold value, it is determined that this pixel belongs to prospect;

When at least two neighbor pixels belong to prospect, form connected domain.

13. according to the arbitrary described device of claim 8-12, it is characterised in that, also comprise:

Track line profile determination module, for, after determining line region, track, verifying the edge point in line region, track, it is determined that the profile of track line.

14. devices according to claim 13, it is characterised in that, described track line profile determination module specifically for:

Gather two edge points at the same edge being positioned at line region, track;

Described two edge points are carried out fitting of a straight line;

Calculate the number of the edge point meeting fitting of a straight line result;

Determine that the fitting of a straight line result that described number is maximum is the profile of track line.