CN106485715A - A kind of unstructured road recognition methods - Google Patents

Abstract

The invention discloses a kind of unstructured road recognition methods, comprises the following steps：S1, acquisition mobile robot are running over the coloured image of Cheng Qian side, and the coloured image to obtaining is pre-processed；S2, region segmentation is carried out to the image that obtains of pretreatment；S3, rim detection is carried out to pretreated image；S4, the image obtained by rim detection carry out fusion extraction, obtain road area.The present invention obtains the coloured image in front of mobile robot first, and the coloured image to obtaining is pre-processed；Then region segmentation is carried out to image, while extracting edge feature；Last integration region segmentation and edge feature information, obtain road area.OTSU and multidirectional sobel operator edge detection algorithm is respectively adopted in region segmentation and rim detection, with less time complexity, algorithm also has preferable recognition effect for new environment, and its robustness is greatly enhanced, and also improves accuracy and the validity of calculating.

Description

A kind of unstructured road recognition methods

Technical field

The invention belongs to computer vision field, more particularly to a kind of unstructured road recognition methods.

Background technology

Currently, the world is in one " intellectualization times ", including intelligent information acquisition, Intelligent Information Processing and intelligence Human-computer interaction technology will also continue the deep life for changing people in interior intellectual technology.China proposes " made in China 2025 " strategic plan, its core are also the manufacturing intelligent level of lifting.

Traditional intelligence technology has many development bottlenecks while great variety is brought to economic society, also.These Bottleneck is difficult to tackle the requirement that advanced information society processes magnanimity complex information, the desire that therefore New Generation of Intelligent revolution is just being exhaled Go out.In recent years, as brain is cognitive and the development of Neuscience, domestic and international academia is all it has been realized that intellectual technology can be from Brain science and Neuscience are obtained and are inspired, and develop new theory and method, improve the level of intelligence of machine.It is based on to human brain information Treatment mechanism and the research of human intelligence, it would be possible to develop the theoretical and technology of a set of class brain intelligence computation, the Fashion of Future information Technology is strided forward to intelligentized developing direction.Self-navigation robot (Automatic navigation robot, ANR) becomes The focus of intelligent age study, and the road Identification of view-based access control model is key technology therein.By vision road Identification technology The road area that also can be walked in front of robot is calculated, robot is according to the state of this regional planning next step.

Road Identification is broadly divided at present：Structurized road Identification, unstructured road are recognized.Structured road refers to It is edge comparison rule, road surface evenness, the carriage way for having obvious lane line and other handmarkings.For example：Highway, Arterial street etc..Identification for structured road has the technical scheme of comparative maturity.The feature of unstructured road is：Road Edge line is degenerated, road surface has the boundary on covering, Fei Lu and road unobvious etc..These complex environments cause destructuring Lane detection technology is extremely difficult.

Being currently based on non-structured road Identification has mode identification technology Bayes's classification, the road based on neutral net Tracking, road vanishing Point Detection Method etc..Although said method in terms of the unstructured road identification can some effects, there is also Some recognize the problems such as inaccurate, restricted application, calculating time are long.It would therefore be desirable to one kind is accurate in complex scene Really, quick road Identification technology.

The patent application of Application No. " 201510443994.9 " discloses a kind of based on non-structured road Identification side Method and device.Its roads recognition method obtains the forward image in vehicle travel process first；Described image is pre-processed, Obtain road profile image；Extract the link characteristic information in the road profile image；According in the road profile image Link characteristic information, recognize the road in the road profile image.This application directly obtains contour images by pretreatment, Not through effectively processing, cause contour images by noise pollution than more serious, it is impossible to effective reaction road area Profile；With monochrome information as road and off-highroad profile is distinguished, easily affected by environment；Increased using template matches Workload, the validity leaf to new environment algorithm will weaken, simultaneously because unstructured road there may be illumination, shade, The complex situations such as water mark, fallen leaves, substantially reduce the effect of template, and detection speed also cannot be guaranteed.

Extracted based on road profile and the roads recognition method of template matches has certain defect.Natural environment in the wild Under, the situation of unstructured road is more complicated, and road area and non-rice habitats area limit be not obvious, and colour information is to discriminate between Road and the key character of Fei Lu.Again as unpredictable and unmanageable weather environment factor affects, the change of brightness is very Greatly, and the change of colourity is then relatively small.Therefore, it can emphasis and enter row threshold division to chrominance information, and by monochrome information only Refer to.

Content of the invention

It is an object of the invention to overcoming the deficiencies in the prior art, one kind is provided based on integration region feature and edge feature Unstructured road recognizer, be respectively adopted OTSU and multidirectional sobel operator in region segmentation and rim detection Edge detection algorithm, with less time complexity, algorithm also has preferable recognition effect for new environment, its robustness Greatly enhance, improve the accuracy of calculating and the unstructured road recognition methods of validity.

The purpose of the present invention is achieved through the following technical solutions：A kind of unstructured road recognition methods, including Following steps：

S1, acquisition mobile robot are running over the coloured image of Cheng Qian side, and the coloured image to obtaining is pre-processed；

S2, region segmentation is carried out to the image that obtains of pretreatment；

S3, rim detection is carried out to pretreated image；

S4, the image obtained by rim detection carry out fusion extraction, obtain road area.

Further, step S1 includes following sub-step：

S11, the foreground image from front of mobile robot within 20 meters of camera acquisition；

S12, by obtain image initial size normalization be 1280 pixels of width and height 960 pixels；

Be averaging in S13, the window by image with 2x2, the image to obtaining carries out double sampling again, last image big Little it is fixed on 320x240；

S14, the image to 320x240 are sharpened operation with Laplacian Second Order Differential Operator；

S15, HSI color space is transformed into, RGB image is transformed into HSI space, by the value of the pixel value of H and S component Scope is normalized to [0,255]；

RGB color is transformed into the formula of HSI color space：

Wherein, V₁、V₂It is the intermediate variable comprising shade of color information, H, S, I represent the component of HSI color space respectively, R, G, B represent the component of RGB color respectively.

Further, step S2 includes following sub-step：

S21, H and S component in HSI is separated, respectively OTSU partitioning algorithm is adopted to the two components；

S22, morphologic closing operations are carried out to S component segmentation result, morphology is carried out to H component segmentation result and opens behaviour Make；

S23, merging S and H component segmentation result：If road area H Zhi Bifei road region H value is low, S component and H are divided Amount segmentation result is carried out and operation, i.e., the corresponding region for being road area in S component and H component is just judged to road area； If road area H Zhi Bifei road region H value is high, by H component segmentation figure as being carried out and operation with S component after inverse again；

S24, the image obtained by S23 carry out medium filtering.

Further, step S3 includes following sub-step：

S31, the local maximum intensity of tri- components of HSI is calculated using multi-direction sobel operator, sobel operator level Direction, vertical direction template, calculate the gradient for (x, y) position in image, obtain based on sobel operator gradient edge detection knot Really, edge strength computing formula is：

VS (x, y)=| I (x-1, y-1)+2I (x-1, y)+I (x-1, y+1)

-I(x+1,y-1)-2I(x+1,y)-I(x+1,y+1)|

HS (x, y)=| I (x-1, y-1)+2I (x, y-1)+I (x+1, y-1)

-I(x-1,y+1)-2I(x,y+1)-I(x+1,y+1)|

Wherein, the edge strength in VS (x, y) expression pixel (x, y) vertical direction, HS (x, y) expression pixel (x, Y) edge strength in horizontal direction, I (x, y) represent the pixel value of pixel (x, y)；

Then maximum value in each (x, y) local strength is defined as the local maxima pixel of the point, obtains cromogram Local maximum intensity figure as each component of HSI space；

S32, calculating local maximum intensity threshold value T, obtain threshold value, i.e., when edge pixel and non-edge picture using comentropy When the comentropy sum of the two classifications plain obtains maximum, corresponding T is required threshold value；

Local maximum intensity in figure is judged to edge pixel when certain pixel local maximum intensity is not less than T, is otherwise non-side Edge pixel；

S33, employing or computing merge tri- edge features of H, S, I, obtain in three components of colour edging, i.e. (x, y) One be edge, then merge image in corresponding points just be edge.

Further, step S4 includes following sub-step：

S41, integration region segmentation and edge detection results, revise erroneous judgement region；The cromogram that integration region segmentation is obtained The colour edging that the road area of picture and rim detection are obtained：Start to scan up from the lower end of region segmentation image, work as road When width is less than 100 pixel values, start to test the change of width, width reaches minimum and corresponding edge detection results and is Edge, is judged as judging region by accident, the road is revised as non-rice habitats region；

S42, the image obtained by S41 carry out continuity testing, obtain final road area.

The invention has the beneficial effects as follows：Propose a kind of based on integration region feature and the unstructured road of edge feature Recognizer；Coloured image mobile robot in front of is obtained first, and the coloured image to obtaining is pre-processed；Then to figure As carrying out region segmentation, while extracting edge feature；Last integration region segmentation and edge feature information, obtain road area. OTSU and multidirectional sobel operator edge detection algorithm is respectively adopted in region segmentation and rim detection, with less Time complexity, algorithm also have preferable recognition effect for new environment, and its robustness is greatly enhanced, can exclude illumination, Shade, water mark, fallen leaves etc. impact of the environmental factor to result of calculation, improves the accuracy and validity for calculating.

Description of the drawings

Fig. 1 is the unstructured road recognition methods flow chart of the present invention.

Specific embodiment

The present invention adopts integration region feature and edge feature road Identification exactly.For the road conditions that field is complicated (as illumination, shade, water mark, fallen leaves, partial occlusion), obtains high resolution graphics as far as possible when mobile robot obtains image Picture, in the case that light application ratio is dark, camera can automatically open up infrared lamp, and the information content of acquisition is just more.In pretreatment rank Section, needs the above-mentioned image to acquisition on the basis of road Identification is had substantially no effect on, and carries out change of scale, normalization pixel value Span, eliminate impact of the partial noise to image, change the color space that needs to us of image, to the mileage chart Image intensifying, becomes apparent from rim detection effect below.In the region segmentation stage, due to unpredictable in natural environment and can not The weather environment factor impact of manipulation, the change of brightness is very big, and the change of colourity is then relatively small.Therefore, it can emphasis Chrominance information is carried out value segmentation is closed, and monochrome information is only referred to.The monochrome information and colour information of HSI color space be point Open.There is above-mentioned two aspects reason, the present invention carries out road image using the colourity tone of HSI color space and saturation degree and closes Value segmentation, partitioning algorithm adopt auto-thresholding algorithm (OTSU) algorithm based on colourity maximum between-cluster variance.At edge Detection-phase, as coloured image can provide more information than gray level image, so present invention contemplates that examine from colour edging More detailed marginal information is extracted during survey, and the present invention is in the method with gradient operator (sobel) is based on.Respectively at three points Rim detection is carried out in amount, and the method for then merging three-component can not meet unstructured road inspection under complicated road environment Survey and require.In order to more accurately extract road image marginal information, algorithm should have Detection results good enough, while do not have again Too big amount of calculation.Invention is using the optimal threshold edge detection algorithm of multi-direction sobel operator herein.Rank is extracted in fusion Section, in conjunction with region segmentation image above and edge feature image, obtains last road region in the picture.With reference to attached Figure further illustrates technical scheme.

As shown in figure 1, a kind of unstructured road recognition methods, comprises the following steps：

S1, acquisition mobile robot are running over the coloured image of Cheng Qian side, and the coloured image to obtaining is pre-processed； Including following sub-step：

S11, the foreground image from front of mobile robot within 20 meters of camera acquisition；

S12, by obtain image initial size normalization be 1280 pixels of width and height 960 pixels；

Be averaging in S13, the window by image with 2x2, the image to obtaining carries out double sampling again, last image big Little be fixed on 320x240, final result reduces resolution ratio and noise on the premise of not fuzzy data, well；

S14, the image Laplacian second-order differential in order to improve the profile of road and non-road further, to 320x240 Operator is sharpened operation；

S15, HSI color space is transformed into, RGB image is transformed into HSI space, by the value of the pixel value of H and S component Scope is normalized to [0,255]；

RGB color is transformed into the formula of HSI color space：

Wherein, V₁、V₂It is the intermediate variable comprising shade of color information, H, S, I represent the component of HSI color space respectively, R, G, B represent the component of RGB color respectively；

S2, region segmentation is carried out to the image that obtains of pretreatment；Including following sub-step：

S21, H and S component in HSI is separated, respectively OTSU partitioning algorithm is adopted to the two components；

S22, morphologic closing operations are carried out to S component segmentation result, morphology is carried out to H component segmentation result and opens behaviour Make；

S23, merging S and H component segmentation result：If road area H Zhi Bifei road region H value is low, S component and H are divided Amount segmentation result is carried out and operation, i.e., the corresponding region for being road area in S component and H component is just judged to road area； If road area H Zhi Bifei road region H value is high, by H component segmentation figure as being carried out and operation with S component after inverse again；

S24, the image obtained by S23 carry out medium filtering.

S3, rim detection is carried out to pretreated image；Including following sub-step：

S31, the local maximum intensity of tri- components of HSI is calculated using multi-direction sobel operator, sobel operator level Direction, vertical direction template, calculate the gradient for (x, y) position in image, obtain based on sobel operator gradient edge detection knot Really, edge strength computing formula is：

VS (x, y)=| I (x-1, y-1)+2I (x-1, y)+I (x-1, y+1)

-I(x+1,y-1)-2I(x+1,y)-I(x+1,y+1)|

HS (x, y)=| I (x-1, y-1)+2I (x, y-1)+I (x+1, y-1)

-I(x-1,y+1)-2I(x,y+1)-I(x+1,y+1)|

Wherein, the edge strength in VS (x, y) expression pixel (x, y) vertical direction, HS (x, y) expression pixel (x, Y) edge strength in horizontal direction, I (x, y) represent the pixel value of pixel (x, y)；

Then maximum value in each (x, y) local strength is defined as the local maxima pixel of the point, obtains cromogram Local maximum intensity figure as each component of HSI space；

S32, calculating local maximum intensity threshold value T, obtain threshold value, i.e., when edge pixel and non-edge picture using comentropy When the comentropy sum of the two classifications plain obtains maximum, corresponding T is required threshold value；

Local maximum intensity in figure is judged to edge pixel when certain pixel local maximum intensity is not less than T, is otherwise non-side Edge pixel；

S33, employing or computing merge tri- edge features of H, S, I, obtain in three components of colour edging, i.e. (x, y) One be edge, then merge image in corresponding points just be edge.

S4, the image obtained by rim detection carry out fusion extraction, obtain road area；Including following sub-step：

S41, integration region segmentation and edge detection results, revise erroneous judgement region；The cromogram that integration region segmentation is obtained The colour edging that the road area of picture and rim detection are obtained：Start to scan up from the lower end of region segmentation image, work as road When width is less than 100 pixel values, start to test the change of width, width reaches minimum and corresponding edge detection results and is Edge, is judged as judging region by accident, the road is revised as non-rice habitats region；

S42, the image obtained by S41 carry out continuity testing, obtain final road area.

Claims (5)

1. a kind of unstructured road recognition methods, it is characterised in that comprise the following steps：

S1, acquisition mobile robot are running over the coloured image of Cheng Qian side, and the coloured image to obtaining is pre-processed；

S2, region segmentation is carried out to the image that obtains of pretreatment；

S3, rim detection is carried out to pretreated image；

S4, the image obtained by rim detection carry out fusion extraction, obtain road area.

2. unstructured road recognition methods according to claim 1, it is characterised in that step S1 includes following son Step：

S11, the foreground image from front of mobile robot within 20 meters of camera acquisition；

S12, by obtain image initial size normalization be 1280 pixels of width and height 960 pixels；

It is averaging in S13, the window by image with 2x2, the image to obtaining carries out double sampling again, the size of last image is solid It is scheduled on 320x240；

S14, the image to 320x240 are sharpened operation with Laplacian Second Order Differential Operator；

S15, HSI color space is transformed into, RGB image is transformed into HSI space, by the span of the pixel value of H and S component It is normalized to [0,255]；

RGB color is transformed into the formula of HSI color space：

$\left[\begin{array}{c}I\\ V_1\\ V_2\end{array}\right]=\left[\begin{array}{ccc}\frac{1}{3}& \frac{1}{3}& \frac{1}{3}\\ \frac{-1}{\sqrt{6}}& \frac{-1}{\sqrt{6}}& \frac{2}{\sqrt{6}}\\ \frac{1}{\sqrt{6}}& \frac{-2}{\sqrt{6}}& \frac{1}{\sqrt{6}}\end{array}\right]\·\left[\begin{array}{c}R\\ G\\ B\end{array}\right]$

$\begin{array}{cc}H=arctan\frac{V_2}{V_1}& S=\sqrt{V_1^2+V_2^2}\end{array}.$

Wherein, V₁、V₂Be the intermediate variable comprising shade of color information, H, S, I represent the component of HSI color space respectively, R, G, B represents the component of RGB color respectively.

3. unstructured road recognition methods according to claim 2, it is characterised in that step S2 includes following son Step：

S21, H and S component in HSI is separated, respectively OTSU partitioning algorithm is adopted to the two components；

S22, morphologic closing operations are carried out to S component segmentation result, morphology is carried out to H component segmentation result and opens operation；

S23, merging S and H component segmentation result：If road area H Zhi Bifei road region H value is low, S component and H component are divided Cutting result is carried out and operation, i.e., the corresponding region for being road area in S component and H component is just judged to road area；If Road area H Zhi Bifei road region H value is high, by H component segmentation figure as being carried out and operation with S component after inverse again；

S24, the image obtained by S23 carry out medium filtering.

4. unstructured road recognition methods according to claim 3, it is characterised in that step S3 includes following son Step：

S31, the local maximum intensity of tri- components of HSI is calculated using multi-direction sobel operator, sobel operator horizontal direction, Vertical direction template, calculates the gradient for (x, y) position in image, obtains based on sobel operator gradient edge testing result, side Edge strength calculation formula is：

VS (x, y)=| I (x-1, y-1)+2I (x-1, y)+I (x-1, y+1)

-I(x+1,y-1)-2I(x+1,y)-I(x+1,y+1)|

HS (x, y)=| I (x-1, y-1)+2I (x, y-1)+I (x+1, y-1)

-I(x-1,y+1)-2I(x,y+1)-I(x+1,y+1)|

Wherein, VS (x, y) represents the edge strength in pixel (x, y) vertical direction, and HS (x, y) represents pixel (x, y) water Edge strength square upwards, I (x, y) represent the pixel value of pixel (x, y)；

Then maximum value in each (x, y) local strength is defined as the local maxima pixel of the point, obtains coloured image HSI The local maximum intensity figure of each component of space；

S32, calculate local maximum intensity threshold value T, obtain threshold value using comentropy, i.e., when edge pixel and non-edge pixels this When the comentropy sum of two classifications obtains maximum, corresponding T is required threshold value；

Local maximum intensity in figure is judged to edge pixel when certain pixel local maximum intensity is not less than T, is otherwise non-edge picture Element；

S33, employing or computing merge tri- edge features of H, S, I, obtain in three components of colour edging, i.e. (x, y) Individual for edge, then it is just edge to merge corresponding points in image.

5. unstructured road recognition methods according to claim 4, it is characterised in that step S4 includes following son Step：

S41, integration region segmentation and edge detection results, revise erroneous judgement region；The coloured image that integration region segmentation is obtained The colour edging that road area and rim detection are obtained：Start to scan up from the lower end of region segmentation image, work as road width During less than 100 pixel values, start to test the change of width, width reaches minimum and corresponding edge detection results for side Edge, is judged as judging region by accident, the road is revised as non-rice habitats region；

S42, the image obtained by S41 carry out continuity testing, obtain final road area.