CN104899554A - Vehicle ranging method based on monocular vision - Google Patents

Abstract

The invention discloses a vehicle ranging method based on monocular vision, and belongs to the field of target detection and ranging. The method comprises the steps of installing a monocular camera on a vehicle, measuring the height and the pitching angle of the camera, and determining focal distance parameters of the camera; acquiring a video image in an expressway environment; performing preliminary de-noising and filtration on the video image by adopting Gaussian filtration; performing interest region segmentation preprocessing before detection of a target vehicle on the video image; performing vehicle detection in a segmented video image region, wherein an Haar feature for increasing a wheel feature and a tail feature is adopted in the target vehicle detection, so as to effectively improve the accuracy of target vehicle identification; and measuring the distance of the target vehicle, wherein a ranging method based on pin-hole imaging is adopted within a short-distance range, while a ranging method of data fitting is adopted within a long-distance range, so that the error rate is reduced, and a real-time ranging effect can be achieved. The method has the advantages of high detection speed, high accuracy, strong real-time property and low cost.

Description

A kind of vehicle odometry method based on monocular vision

Art

The invention belongs to target detection and range finding field, particularly relate to a kind of vehicle odometry method based on monocular vision.

Background technology

Along with the development of highway communication particularly freeway facility, road accident rate also presents ascendant trend, and traffic safety more and more becomes the focus that people pay close attention to.Therefore, research vehicle safety assists driving technology, for vehicle provides safety assistant driving function, thus provides intellectual technology service for reducing the traffic hazard caused because of driver's subjective factor.Computer vision because of its provide contain much information, the factor such as good stability, become the research emphasis that vehicle safety assists driving technology gradually, along with the development of computer vision technique, its effect in Intelligent Vehicle System is constantly perfect, computer vision technique is applied in vehicle detection, great effect is created to the raising of vehicle safety.

Utilize monocular cam to carry out the detection of moving target (such as vehicle), its development roughly experienced by three phases: passive learning stage---Active Learning stage---the adaptive learning stage.In the passive learning stage, mainly according to the feature of image, carry out matching to the object of all existence, the target that after comparing the difference of front and back two field picture, district office will detect, its main algorithm has mixed Gauss model, background subtraction, Kalman filter etc.In the Active Learning stage, mainly for specific moving target, study its inherent feature, by the study to feature, moving target is detected.At present for the vehicle of motion, conventional inherent feature comprises shade that vehicle bottom produces, the entropy of shade, the symmetry of vehicle edge, the brightness of vehicle pixel, the texture etc. of vehicle.In adaptive learning, the detection of moving target is roughly divided into three steps: the first step is feature extraction, here feature mainly mathematical feature, feature extraction algorithm conventional now has HOG, Haar, SIFT, LBT etc., and this step is the basis of two steps next; Second is the training of sorter, and by inputting a large amount of positive negative sample, obtain the sorter identifying target mathematical feature through training, classifier algorithm conventional now comprises SVM, Adaboost etc.; 3rd step is moving object detection, and this step is mainly used through training the sorter obtained, and carries out moving object detection to the video image of input.

Monocular vision range finding is that the picture utilizing a video camera to obtain draws depth information, is mainly divided into the measuring method based on geometric relationship and the measuring method based on data message according to the principle measured.Measuring method based on geometric relationship refers to that the picture utilizing the structure of video camera and video camera to obtain records depth distance.Utilizing theory on computer vision and method, carrying out the front vehicles in runway, on quick detection and the basis demarcated in advance video camera, utilizing camera parameters and road geometric model, obtaining front vehicles distance.The shortcoming of above-mentioned measurement is the coupling will carrying out unique point to a width or a few width picture, and matching error has obvious impact to measurement result, and the processing time is long simultaneously, for multiple image, then must need more computing time.

Measuring method based on data message refers to that the Target Photo utilizing video camera to obtain under the condition of known object information obtains depth distance.The shortcoming of these class methods needs the accurate information utilizing image to measure, and easily causes the inaccurate of measurement because of the inaccurate of image information.

Summary of the invention

For the deficiency that existing method exists, the present invention proposes a kind of vehicle odometry method based on monocular vision.

Technical scheme of the present invention is achieved in that

Based on a vehicle odometry method for monocular vision, vehicle odometry object is the front vehicles travelled in the same way, comprises the steps:

Step 1: install monocular cam on vehicle, measures camera height and the angle of pitch thereof, and determines camera focal length parameter;

First monocular cam is fixed on vehicle front, determines this camera distance height on ground and the angle of axis and horizontal direction thereof, be i.e. the vertical height of this monocular cam and the angle of pitch;

Step 2: the video image under utilizing described monocular cam to gather highway environment;

Step 3: the video image preprocessing process before target vehicle detection;

Step 3.1: adopt gaussian filtering to carry out preliminary denoising, filtering process to video image;

Step 3.2: the segmentation of the region-of-interest before target vehicle detection pre-service is carried out to the video image after step 3.1 rough handling;

Step 4: target vehicle detection process;

Carry out vehicle detection in video image region after singulation, and the target vehicle detected is marked in real time on picture;

Step 5: target vehicle ranging process;

Measurement target vehicle distances also shows this target vehicle distance in real time on video pictures.

Described step 3.2 comprises following concrete steps:

Step 3.2.1: sky areas segmentation is carried out to the video image gathered;

Adopt the video image gathered and carry out sky areas segmentation based on color space, method is specific as follows: first obtain the histogram of video image in HIS (tone, color saturation, brightness), RGB (red, green, blue three look), YIQ (brightness, tone, saturation degree) and YCbCr (colour brightness, blueness and red color side-play amount) four kinds of color spaces; Then in these four kinds of color spaces, determine the distribution range of each color component of sky areas respectively, calculate and the variance of more determined Four composition cloth range data and extreme value, choose wherein variance and the minimum also YCbCr color space that namely distribution of each component is the most concentrated of extreme value as the color space of sky spatial segmentation; Finally binary conversion treatment is carried out to video image, determine the UNICOM region of sky and calculate its area, and adopting Otsu Adaptive Thresholding, automatically adjusting segmentation threshold, sky portion is removed from image;

Step 3.2.2: to the video image after the segmentation of sky areas, adopt Minimum error threshold method inspection vehicle diatom, namely near the road edge line of image border, and two-dimentional straight-line equation is set up to the lane line detected, and the region of removing based on this equation outside edge line, reduce vehicle detection region area further;

Described step 4 comprises following concrete steps:

Step 4.1: gather positive and negative sample image (positive sample refers to vehicle rear picture, and negative sample refers to other any image, but can not comprise vehicle rear), be normalized as onesize to all positive and negative samples pictures;

Step 4.2: increase tailstock characteristic sum back wheels of vehicle feature in Haar feature, and according to this Haar feature, adopt Adaboost algorithm to train positive and negative sample set, obtain cascade classifier;

Step 4.3: utilize the cascade classifier obtained, target vehicle detection is carried out to the video image of monocular cam collection, and the target vehicle detected is marked in real time on picture;

In target vehicle detection process in described step 4.3, multiple dimensioned windowhood method (multiscale approach) video image to monocular cam collection is utilized to carry out Scanning Detction;

In described step 5, the method for measurement target vehicle distances is as follows:

If target vehicle distance is within 30 meters, then set up video camera projection model according to pinhole imaging system principle, world coordinate system is projected in image coordinate system, sets up vehicle odometry geometric relationship model by relation corresponding between Two coordinate system, ask for objects ahead vehicle distances; If target vehicle distance is greater than 30 meters, then first obtains the mapping relations between real road sample point and picture plane by data fitting method, and ask for objects ahead vehicle distances according to these mapping relations.

Advantage of the present invention is: the suitable environment of the vehicle odometry method based on monocular vision of the present invention is highway, first the method carries out the acquisition of call parameter to the monocular cam be arranged on vehicle, then after adopting gaussian filtering to carry out rough handling to the video image of this camera acquisition, again pre-service is carried out to video image: be first adopt the sky areas dividing method based on color space, rational segmentation threshold is found by the method for adaptive adjustment threshold value, tell sky areas, decrease the scan area of image; Next territory, track is split, further reduce the scan area of image; The Haar feature being the increase in wheel characteristics and tailstock feature adopted in the process of target vehicle detection, effectively improves the accuracy of target vehicle identification; In the process of target vehicle detection (identification), adopt multiple dimensioned windowhood method to carry out Scanning Detction to the video image of monocular cam collection, also can significantly improve target vehicle detection speed.In target vehicle range observation, in short range, (within 30 meters) adopt the distance-finding method based on pinhole imaging system; The scope of long distance (is greater than 30 meters), adopts the distance-finding method of data fitting (linear difference), reduce error rate, the effect of real time distance can be reached.And method of the present invention only adopts a camera collection video, and equipment is simple.Therefore, method of the present invention has that detection speed is fast, and accuracy rate is high, the comparatively strong and advantage that cost is low of real-time.

Accompanying drawing explanation

Fig. 1 is the vehicle odometry method flow diagram based on monocular vision of one embodiment of the present invention;

Fig. 2 is the Haar characteristic set figure of one embodiment of the present invention;

Fig. 3 is that the Adaboost of one embodiment of the present invention trains process flow diagram;

Fig. 4 is the multiple dimensioned windowhood method scanning process schematic diagram of one embodiment of the present invention;

Fig. 5 is the pinhole imaging system principle schematic of one embodiment of the present invention;

Fig. 6 is schematic diagram when tested point departs from center in the pinhole imaging system of one embodiment of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail.

Fig. 1 is the vehicle odometry method flow diagram based on monocular vision of present embodiment, and the method comprises following concrete steps:

Step 1: install monocular cam on vehicle, measures camera height, knows camera focal length parameter;

First monocular cam is fixed on vehicle front, determines this camera distance height on ground and the angle of its axis and horizontal direction thereof, be i.e. the vertical height of this monocular cam and the angle of pitch;

Step 2: the video image under utilizing described monocular cam to gather highway environment;

Step 3: the video image preprocessing process before target vehicle detection;

Step 3.1: adopt gaussian filtering to carry out preliminary denoising, filtering process to video image;

Step 3.2: the segmentation of the region-of-interest before target vehicle detection pre-service is carried out to the video image after step 3.1 rough handling, needs processing region to reduce, improve target vehicle detection speed;

Step 3.2.1: sky areas segmentation is carried out to the video image gathered;

Adopt the method for carrying out sky areas segmentation based on color space, specific as follows: first to obtain the histogram of video image in HIS (tone, color saturation, brightness), RGB (red, green, blue three look), YIQ (brightness, tone, saturation degree) and YCbCr (colour brightness, blueness and red color side-play amount) four kinds of color spaces; Then in these four kinds of color spaces, determine the distribution range of each color component of sky areas respectively, the variance of calculating and more determined Four composition cloth range data and extreme value, choose the color space of the minimum also YCbCr color space that namely distribution of each component is the most concentrated of wherein variance and extreme value as sky spatial segmentation, experiment finds, for YCbCr color space, Y and Cr distribution is more concentrated, the distributed areas of sky can be recognized preferably, and less on the impact of entire image, therefore adopt YCbCr color space as the color space removing sky; Finally binary conversion treatment is carried out to video image, determine the UNICOM region of sky and calculate its area, and adopting Otsu Adaptive Thresholding, automatically adjusting segmentation threshold, sky portion is removed from image; Present embodiment is to remove the harmful effect of the undesirable elements such as illumination to COLOR COMPOSITION THROUGH DISTRIBUTION region in sky spatial segmentation process, adopts Otsu Adaptive Thresholding automatically suitably to adjust threshold value thus improves the accuracy of sky segmentation;

Formula (1) is the process function to sky areas in experiment, the gray-scale value that F (x, y) is image.When sky areas meets the color element scope in formula (1), the gray-scale value of sky areas is set to 255, removes sky areas.

$F(x,y)=\left\{\begin{array}{cc}255& Y\∈\left[\mathrm{0.627,0.702}\right]\\ \mathrm{and}& \mathrm{Cr}\∈\left[\mathrm{0.494,0.564}\right]\\ \mathrm{and}& \mathrm{Cb}\∈\left[\mathrm{0.407,0.502}\right]\\ 0& \mathrm{others}\end{array}\right.---\left(1\right)$

In order to get rid of the distracter (color, building etc. as car) consistent with sky color that may exist in picture, when only having the area when UNICOM region to be greater than 1/10 of image, small pieces just can be removed, and eliminate disturbing factor so to a great extent.After the image of sky areas detects, corresponding cutting process can be carried out, for follow-up vehicle detection process.

Pixel coverage due to sky is vulnerable to the impact of the undesirable elements such as illumination, can be found by the histogram distribution of image, the color gamut of sky obeys Gauss normal distribution, if suitable threshold value can be found, image is divided into sky and non-sky portion, so just can well solves have impact on of the problems such as light.

According to Otsu Adaptive Thresholding, with the separation of threshold value ψ as binaryzation, what be greater than ψ is classified as a class, is less than being classified as of ψ another kind of, if the average of two classes is respectively v ₁and v ₂, the distribution probability of two classes is respectively u ₁and u ₂.

α ²＝u ₁u ₂(v ₂-v ₁) ²(2)

By adjusting the size of threshold value ψ, allow variance α ²obtain maximal value, ψ at this time can as the threshold value of sky areas segmentation.

Present embodiment adopts adaptivenon-uniform sampling algorithm, and for the sky that process illumination is undesirable, effect is ideal.

Step 3.2.2: to the video image after the segmentation of sky areas, adopt Minimum error threshold method inspection vehicle diatom, namely near the road edge line of image border, and two-dimentional straight-line equation is set up to the lane line detected, and the region of removing based on this equation outside edge line, reduce vehicle detection region area further; Most of extraneous areas can be removed through sky areas segmentation and this twice segmentation of track regional partition process, reduce image-region to be dealt with in vehicle detection, thus improve the speed of vehicle detection.

Present embodiment adopts Minimum error threshold method inspection vehicle diatom, and determine road area and non-rice habitats region with this.Extraneous areas in every frame picture is got rid of, only retains road sections.Minimum error threshold algorithm has good noise immunity, and Detection results is desirable.Compared to two-dimentional Minimum error threshold algorithm, the calculated amount of one dimension least error algorithm is less, greatly can shorten the processing time, and can meet the requirement of the image process in early stage for target vehicle identification (detection).

Detailed process is as follows:

Be the image of M × N for a width size, if the gray-scale value that coordinate is the pixel of (x, y) be h (x, y) ∈ G=[0,1 ..., 255].Represent the one dimensional histograms of image by function f (g), it represent the frequency that in image, each gray-scale value occurs, for gray threshold m ∈ G=[0,1 ..., 255], the function of minimum classification error is

Z(m)＝1+2[P ₀(m)lnσ ₀(m)+P ₁(m)lnσ ₁(m)]-2[P ₀(m)lnP ₀(m)+P ₁(m)lnP ₁(m)] (3)

Wherein, $P_0\left(m\right)=\underset{g=0}{\overset{m}{\mathrm{\&Sigma;}}}f\left(g\right)$ And $P_1\left(m\right)=\underset{g=m+1}{\overset{255}{\mathrm{\&Sigma;}}}f\left(g\right)$ It is the prior probability of two son distributions; ${\mathrm{\&mu;}}_0\left(m\right)=\frac{\underset{g=0}{\overset{m}{\mathrm{\&Sigma;}}}g\&CenterDot;f\left(g\right)}{P_0\left(m\right)}$ With ${\mathrm{\&mu;}}_0\left(m\right)=\frac{\underset{g=m+1}{\overset{255}{\mathrm{\&Sigma;}}}g\&CenterDot;f\left(g\right)}{P_1\left(m\right)}$ It is the average of two son distributions; ${{\mathrm{\&sigma;}}^2}_0\left(m\right)=\frac{\underset{g=0}{\overset{t}{\mathrm{\&Sigma;}}}{(g-{\mathrm{\&mu;}}_0\left(m\right))}^{2}f\left(g\right)}{P_0\left(m\right)}$ With it is the variance of two son distributions; Optimum gradation threshold value m ^*for the threshold value making Z (m) get minimum value, at this moment there is m ^*=arg (minZ (m)) (0 < m < 255); The classifying mode of the pixel after therefore can obtaining Minimum error threshold is:

$\stackrel{\&OverBar;}{h}(x,y)=\left\{\begin{array}{ccc}0& \mathrm{when}& h(x,y)<m^{*}\\ 25& \mathrm{when}& h(x,y)\&GreaterEqual;m^{*}\end{array}\right.$

Wherein, for the pixel after Minimum error threshold; H (x, y) is the pixel before Minimum error threshold.

Experiment proves, Minimum error threshold method can be good at by lane detection out.Track and non-track can be separated by the lane line equation set up in two dimensional surface, a large amount of extraneous areas can be removed, thus the workload of vehicle detection process can be reduced.

Step 4: target vehicle detection process;

Carry out vehicle detection in video image region after singulation, and the target vehicle detected is marked in real time on picture;

Step 4.1: gather positive and negative sample image; Wherein positive sample refers to vehicle rear picture, and negative sample refers to other any image, but can not comprise vehicle rear, is all normalized all positive and negative samples pictures, and present embodiment is that all pictures are all normalized to 20*20.

Step 4.2: increase tailstock characteristic sum wheel characteristics in Haar feature, and according to amended Haar feature, adopt Adaboost algorithm to train positive and negative sample set, obtain cascade classifier.

Eigenwert according to feature templates defines, white rectangle pixel and deduct black rectangle pixel and, because integrogram has that computing velocity is fast and calculated value is only relevant with edge, so, adopt integrogram to calculate Haar eigenwert herein.

Present embodiment, on the basis of original Haar feature, adds wheel characteristics and tailstock feature as shown in Figure 2.The accuracy rate of target vehicle detection can be improved.The computing method of Haar feature be white rectangle pixel and deduct black rectangle pixel and, because integrogram has that computing velocity is fast and calculated value is only relevant with edge, so, adopt integrogram to calculate Haar eigenwert to improve computing velocity herein.

Present embodiment, further according to Haar feature, utilizes Adaboost algorithm to align negative sample and trains, obtain cascade classifier.As shown in Figure 3, cascade classifier is made up of multiple Weak Classifier training process flow diagram, and every one-level is all complicated than previous stage.Each sorter can allow nearly all positive example pass through, the negative example of the major part of filtering simultaneously.The positive example to be detected of so every one-level is just few than previous stage, eliminates a large amount of non-detection targets, greatly can improve detection speed.

Step 4.3: utilize the cascade classifier obtained, the video image of monocular cam collection is detected, and in testing process, utilizes multiple dimensioned windowhood method (multiscale approach) to scan;

In order to reduce the time spent when calculating rectangular characteristic, present embodiment have employed multiple dimensioned windowhood method and scans, namely respectively Subarea detecting is carried out to image with the subwindow of different size, and calculate the Haar feature of zones of different simultaneously, scanning schematic diagram as shown in Figure 4, according to the feature of vehicle detection, utilize the vehicle feature that the tailstock varies in size when different distance, image is divided into four regions, and with the subwindow of four different sizes, Subarea detecting is carried out to image, and calculate the Haar feature of four subwindows simultaneously.

Step 5: measurement target vehicle distances also shows this target vehicle distance in real time on video pictures.

If target vehicle distance is within 30 meters, then set up video camera projection model according to pinhole imaging system principle, world coordinate system is projected in image coordinate system, sets up vehicle odometry geometric relationship model by relation corresponding between Two coordinate system, ask for objects ahead vehicle distances; According to pinhole imaging system principle, object in world coordinate system is projected in image coordinate system through small holes, as shown in Figure 5, the reality imagery of plane bchq corresponding to world coordinate system, the projection image of plane BCHQ corresponding to image coordinate system, in plane bchq, hq limit and bc limit extend intersection point is s, and simultaneously s point is also the intersection point of road extended line, and road surface is actual, and that be observed is trapezoidal bchq.Representated by M point is the aperture that convex lens are formed.

If do not consider the impact of camera distortion, optical axis Ii is the center of two planes at the intersection point of two corresponding flats, I as shown in Figure 5, i 2 point.

As shown in Figure 5, in world coordinate system, get a t, through the projection of a M, obtain the some T in image coordinate system.∠ TMI and the ∠ α of Δ MIT can be found ₂be equal, and line segment MI is the focal distance f of lens in video camera, so there is tan α ₂in=TI/f, Fig. 5, MN is horizontal line, ∠ α ₁the angle of pitch formed by optical axis and horizontal line, the height that camera is installed is δ=MW, and observed some t is S=Wi to the distance of photocentre, has:

$\tan {\mathrm{\&alpha;}}_1=\frac{\mathrm{\&delta;}}{S}---\left(4\right)$

Thus can in the hope of range formula:

$d_1=\stackrel{\&OverBar;}{\mathrm{Mt}}=\frac{h}{\tan [{\mathrm{\&alpha;}}_1+\arctan (\mathrm{TI}/f)]}---\left(5\right)$

As shown in Figure 6, when tested point departs from center time, need calculating off-centered lateral separation so measured actual range is

In order to ask for d ₂point, chooses b ₁point is test point, b ₁point is B at the subpoint of image coordinate system ₁; The another one test point crossing with sides aligned parallel and with center line is a ₁, the intersection point projecting to image coordinate system is accordingly A ₁; By the existence of projection theory and vertical plane, ∠ MA can be known ₁b ₁with ∠ Ma ₁b ₁be 90 °, owing to also there are a pair vertical angle, so Δ MA ₁b ₁with Δ Ma ₁b ₁similar.The criterion proportional according to similar triangles corresponding sides, obtains following formula:

$\frac{A_1{B}_1}{a_1{b}_1}=\frac{{\mathrm{MA}}_1}{{\mathrm{Ma}}_1}---\left(6\right)$

For A ₁b ₁, in image coordinate system, if the distance between horizontal ordinate pixel and ordinate pixel is respectively d _xand d _y, then A ₁b ₁between actual range be then wherein represent A ₁b ₁pixel value difference.

In like manner for MA ₁, can push away according to the relation between pixel and actual range and obtain its expression formula and be:

${\mathrm{MA}}_1=\sqrt{f^2+{(\stackrel{\&OverBar;}{A_{1}I}*d_y)}^2}---\left(7\right)$

Wherein that represent is A ₁the pixel value difference of I.

For Ma ₁, owing to having tried to achieve d by Fig. 5 ₁, the Wa namely in figure ₁, the geometric relationship from figure:

${\mathrm{Ma}}_1=\sqrt{h^2+{{\mathrm{Wa}}_1}^2}=\sqrt{{\mathrm{MW}}^2+{{\mathrm{Wa}}_1}^2}---\left(8\right)$

A ₁b ₁be the parameter d of requirement ₂.In sum, the formula calculating distance is:

$d=\sqrt{d_1^2+d_2^2}=\mathrm{sqrt}\left(\begin{array}{c}{\left(\frac{h}{\tan [{\mathrm{\&alpha;}}_1+\arctan (\mathrm{TI}/f)]}\right)}^2+...\\ \frac{\stackrel{\&OverBar;}{A_1{B}_1}\&CenterDot;d_x}{\sqrt{f^2+{(\stackrel{\&OverBar;}{A_{1}I}*d_y)}^2}}\&CenterDot;\sqrt{h^2+{{\mathrm{Wa}}_1}^2}\end{array}\right)---\left(9\right)$

In formula (9), the focal distance f of video camera and TI can check in from the specifications parameter of camera, and setting height(from bottom) h can measure and obtain, α ₁can try to achieve with formula (5).

If target vehicle distance is greater than 30 meters, then first obtains the mapping relations between real road sample point and picture plane by data fitting method, and ask for objects ahead vehicle distances according to these mapping relations.For the video camera laid perpendicular to ground that a frame specification is determined, when its angle of pitch is determined, the distance range of the road ahead that it can detect is certain, position coordinates in the world coordinate system that on image area, each point is corresponding is also certain, as long as so find the position corresponding relation between image coordinate system and world coordinate system just can ask for front vehicles distance by this relation in reality range finding.This method effectively overcomes the harmful effect that the path error such as the lens distortion existed in actual imaging cause, and the problem such as video camera modes of emplacement, surface conditions, road type (straight way or structured road).

The relation matching of Two coordinate system needs to obtain data by experiment to realize.What the data obtained in experiment were formed is discrete point, and present embodiment is the relation of the method construct Two coordinate system adopting cubic spline interpolation.

Claims (6)

1., based on a vehicle odometry method for monocular vision, vehicle odometry object is the front vehicles travelled in the same way, it is characterized in that: comprise the steps:

Step 1: install monocular cam on vehicle, measures camera height and the angle of pitch thereof, and determines camera focal length parameter;

First monocular cam is fixed on vehicle front, determines this camera distance height on ground and the angle of axis and horizontal direction thereof, be i.e. the vertical height of this monocular cam and the angle of pitch;

Step 2: the video image under utilizing described monocular cam to gather highway environment;

Step 3: the video image preprocessing process before target vehicle detection;

Step 3.1: adopt gaussian filtering to carry out preliminary denoising, filtering process to video image;

Step 3.2: the segmentation of the region-of-interest before target vehicle detection pre-service is carried out to the video image after step 3.1 rough handling;

Step 4: target vehicle detection process;

Carry out vehicle detection in video image region after singulation, and the target vehicle detected is marked in real time on picture;

Step 5: target vehicle ranging process;

Measurement target vehicle distances also shows this target vehicle distance in real time on video pictures.

2. the vehicle odometry method based on monocular vision according to claim 1, is characterized in that: described step 3.2 comprises following concrete steps:

Step 3.2.1: sky areas segmentation is carried out to the video image gathered;

Step 3.2.2: to the video image after the segmentation of sky areas, adopt Minimum error threshold method inspection vehicle diatom, namely near the road edge line of image border, and two-dimentional straight-line equation is set up to the lane line detected, and the region of removing based on this equation outside edge line, reduce vehicle detection region area further.

3. the vehicle odometry method based on monocular vision according to claim 1, is characterized in that: described step 4 comprises following concrete steps:

Step 4.1: gather positive and negative sample image, is normalized as onesize to all positive and negative samples pictures; Wherein positive sample refers to vehicle rear picture, and negative sample refers to other any image, but can not comprise vehicle rear;

Step 4.2: increase tailstock characteristic sum back wheels of vehicle feature in Haar feature, and according to this Haar feature, adopt Adaboost algorithm to train positive and negative sample set, obtain cascade classifier;

Step 4.3: utilize the cascade classifier obtained, target vehicle detection is carried out to the video image of monocular cam collection, and the target vehicle detected is marked in real time on picture.

4. the vehicle odometry method based on monocular vision according to claim 2, it is characterized in that: adopt the video image gathered in described step 3.2.1 and carry out sky areas segmentation based on color space, method is specific as follows: first obtain the histogram of video image in HIS, RGB, YIQ and YCbCr tetra-kinds of color spaces; Then in these four kinds of color spaces, determine the distribution range of each color component of sky areas respectively, calculate and the variance of more determined Four composition cloth range data and extreme value, choose wherein variance and the minimum also YCbCr color space that namely distribution of each component is the most concentrated of extreme value as the color space of sky spatial segmentation; Finally binary conversion treatment is carried out to video image, determine the UNICOM region of sky and calculate its area, and adopting Otsu Adaptive Thresholding, automatically adjusting segmentation threshold, sky portion is removed from image.

5. the vehicle odometry method based on monocular vision according to claim 3, is characterized in that: in the target vehicle detection process in described step 4.3, utilizes multiple dimensioned windowhood method to carry out Scanning Detction to the video image of monocular cam collection.

6. the vehicle odometry method based on monocular vision according to claim 1, is characterized in that: in described step 5, the method for measurement target vehicle distances is as follows:

If target vehicle distance is within 30 meters, then set up video camera projection model according to pinhole imaging system principle, world coordinate system is projected in image coordinate system, sets up vehicle odometry geometric relationship model by relation corresponding between Two coordinate system, ask for objects ahead vehicle distances; If target vehicle distance is greater than 30 meters, then first obtains the mapping relations between real road sample point and picture plane by data fitting method, and ask for objects ahead vehicle distances according to these mapping relations.