CN102155955A - Stereoscopic vision mile meter and measuring method - Google Patents

Abstract

The invention discloses a stereoscopic vision mile meter and a measuring method, belongs to the field of autonomous navigation of vehicles, and particularly relates to a stereoscopic vision mile meter. The measuring method for the stereoscopic vision mile meter comprises the following steps of: image acquisition, preprocessing, characteristic extraction and matching, motion estimation, result output and the like. The stereoscopic vision mile meter comprises a hardware platform for integrating high-speed image acquisition, transmission, analysis and processing, and a high-speed image mode identification algorithm suitable for the platform, wherein the algorithm provides necessary software for developing an autonomous navigation system of a vehicle; the hardware provides high-speed stable operation function; the software algorithms are high parallel to make full use of multiple parallel instruction streamlines of a digital signal processor (DSP); and the core algorithm is realized by adopting field programmable gate array (FPGA) hardware and organically matched with the algorithm of the DSP so as to provide an effective supplement method for the traditional mile meter.

Description

The stereoscopic vision mileage is taken into account measuring method

[technical field]

The invention belongs to the vehicular autonomous navigation field, relate in particular to a kind of stereoscopic vision odometer.

[background technology]

In the navigator fix process of autonomous vehicle, the odometer role is most important, and it can draw the distance and the directional information of vehicle '.Usually to calculate wheel revolutions or angular velocity of rotation based on traditional odometer of wheel encoder device and determine that vehicle gait of march, the greatest problem of this method are exactly the counting that causes in the time of can't overcoming wheel-slip or measure mistake.Because the smooth degree in friction factor, ground of athletic ground and wheel, angle of inclination etc. all can't obtain usually, therefore the generation and the degree of sliding is unforeseen often, especially in fields such as military affairs, surveies of deep space, because degree that the soil is porous and wheel slip and terrain environment is unpredictable, make the reading of scrambler exist, can't correctly reflect the situation of body movement than mistake; Simultaneously gps signal also can be because of disturbing etc. former thereby becoming unreliable, and the lunar rover that is operated in menology does not then have the auxiliary of GPS information fully, need take other assisting navigation location technology, to guarantee to obtain enough accurate car body position and attitude estimation.In addition; there is the problem of " drift " in time in the inertial navigation unit that vehicle is commonly used; if lack the correction of overall locating information such as GPS; will cause dead reckoning to have the defective that precision is relatively poor and error increases with roaming distance increasing run-up; need take other assisting navigation location technology, to guarantee to obtain enough accurate car body position and attitude estimation.Therefore, make up a kind of autonomous navigation system, have crucial meaning in the accurate location and the estimation of long-range navigation under the destructuring environment or under the accurate structured environment realizing autonomous vehicle with good robustness and precision.

[summary of the invention]

The present invention is based on the vision odometer of stereoscopic vision, is a kind of by carry the binocular vision device on autonomous vehicle, according to the vision input, and the movement estimation system that the position and the motion state of vehicle are estimated.Purpose is to provide a kind of effective compensation process for traditional odometer, especially under some destructuring environment as contour sliding zone, sandy slope, stereoscopic vision odometer and wheel encoder device and inertial navigation equipment are worked in coordination, and can obtain more accurate car body location.The stereoscopic vision mileage system integrates technology such as Flame Image Process, pattern-recognition, embedded system based on the embedded system fundamental construction.

The present invention obtains sequence of stereoscopic images by the binocular solid video camera, tries to achieve the exercise data of autonomous vehicle through steps such as the three-dimensional coupling of feature extraction, unique point, feature point tracking coupling, coordinate transform and estimation, and its system framework as shown in Figure 1.

Stereoscopic vision odometer of the present invention is connected successively by following modules:

1). image capture module, adopt binocular CCD camera to take measurand, then with the data output that collects;

2). signal transmission module, the data that the CCD camera collects are carried out preliminary hardware filtering, adopt A/D converter, signal amplifier that the signal of image capture module output is changed amplification, the signal transformation that makes simulation is digital signal, carries out data processing for processor;

3). image analysis module, adopt the embedded system that constitutes by DSP+FPGA as central processing unit, carry out image analysis processing, obtain image feature information, and then through characteristic matching, estimation obtains the body movement parameter;

4). output control module, result to be exported by output control module, output control module comprises LCD, driving pulse.

The method of utilizing the stereoscopic vision odometer to measure comprises the steps:

A, image acquisition: the view data of gathering autonomous vehicle environment of living in real time by binocular CCD camera; Data stream deposits among the SDRAM through the EDMA passage, finishes image acquisition; The image sequence that collects will pass through pre-service (denoising, filtering, background compensation etc.) just can carry out subsequent analysis processing, and the collection effect better image can be used through pre-service seldom, and this has just saved the processing time.

B, pre-service: fundamental purpose is to remove the background and context interference noise, adopts FPGA to realize;

C, feature extraction and coupling: obtain its characteristics of image after the image sequence process denoising that collects, image normalization, dynamic binaryzation, filtering, the refinement, adopt the SIFT algorithm that the image after handling is carried out feature extraction and coupling;

D, estimation: the some feature that obtains after the characteristic matching is carried out three-dimensional reconstruction to sequence, the uncertainty of calculated characteristics point three-dimensional reconstruction; According to unique point locus and uncertain covariance matrix, utilize least square method and maximum-likelihood method that autonomous vehicle is carried out estimation;

Wherein, characteristic matching and estimation are core algorithms of the present invention, adopt the DSP+FPGA software and hardware to realize this core algorithm, have solved the bottleneck of processing speed.

E, result's output: after the motion estimation analysis, the kinematic parameter such as the information such as move distance, direction of car body are exported by output control module.

Technology path that the present invention is concrete and scheme are:

1) use binocular CCD camera to gather autonomous vehicle environment of living in, the view data of gathering is stored in the internal memory through A/D, FPGA, EDMA under the control of FPGA, uses FPGA picture signal to be carried out pre-service such as denoising, correction lens distortion in this process.

2) obtain its characteristics of image after the image sequence process denoising that collects, image normalization, dynamic binaryzation, filtering, the refinement.

3) select the extracting method of SIFT algorithm as unique point.

If input picture be I (x, y), G (x, y σ) are the variable dimension Gaussian function, and then the metric space L of image can obtain by convolution:

L(x，y，σ)＝G(x，y，σ)*I(x，y)

Can travel through whole metric space by changing scale factor σ.The SIFT operator carries out extreme value at metric space and detects,

Obtain more stable candidate feature point; Then the metric space equation is carried out Taylor and launch, unique point is accurately located.Adopt the principal direction of key point neighborhood gradient to characterize this characteristic direction, thereby realize the independence of feature detection to yardstick and direction, such key point will produce the descriptor vector of 128 dimensions.

4) adopt the SIFT algorithm to carry out the solid coupling.The Euclidean distance of the descriptor vector of 128 dimensions that above-mentioned steps 3) obtain can be used as the similarity determination tolerance of unique point in two width of cloth images.Get certain unique point in the piece image wherein, and find out two nearest unique points of Euclidean among itself and another width of cloth figure, in these two unique points, if the ratio of nearest distance and inferior near distance is less than certain proportion threshold value, then accept this a pair of match point, otherwise refusal.

5) estimation.Based on triangulation relation, by the two groups of stereo-pictures in front and back are drawn parameters such as betwixt move distance of car body, direction to carrying out processing such as coordinate transform, and long apart from suitable robustness should arranged; Adopt least square method and maximum likelihood algorithm for estimating, make error minimize.

6) consider that the parameter that need investigate is many, operand is big, adopts parallel algorithm and FPGA hardware to realize that jointly embedded main process chip is selected the high-speed dsp processor TMS320DM642 of American TI Company for use, adopt the DSP+FPGA software and hardware to realize this algorithm, solved the bottleneck of processing speed.

7) entire equipment does not need computing machine to connect based on Embedded System Design, can be equipped on car body easily.

8) the TCP/IP procotol of the own support standard of equipment, can with seamless the integrating of external network architecture, realize sharing and interconnecting of information.

Take into full account modular design when 9) image capture module and arithmetic section design, can make up according to actual needs, meet the different needs.

Advantage of the present invention is:

Therefore 1) only rely on the vision input information, do not exist, need not the prior imformation of scene and motion because of the error that the scrambler reading is inaccurate, sensor accuracy reduces or factor such as inertial navigation drift causes.

2) working in coordination with wheel encoder device and inertial navigation equipment, can obtain more accurate car body location, is effectively replenishing of classic method.

3) adopt the passive vision sensor, simple in structure, power consumption is little.

4) combine with the embedded system technology that with the high-speed dsp is core, realize that the parallel multi-stage pipeline of DSP Processing Algorithm is carried out and the parallel pipeline processing ability of maximized performance DSP;

[description of drawings]

Fig. 1 is a system framework of the present invention;

Fig. 2 is a modular structure of the present invention.

[embodiment]

The present invention is further described below in conjunction with drawings and Examples.

Embodiment 1

1) image acquisition: the view data of gathering autonomous vehicle environment of living in real time by binocular CCD camera.Data stream deposits among the SDRAM through the EDMA passage, finishes image acquisition.The image sequence that collects will pass through pre-service (denoising, filtering, background compensation etc.) just can carry out subsequent analysis processing, and the collection effect better image can be used through pre-service seldom, and this has just saved the processing time.

2) pre-service: fundamental purpose is to remove the background and context interference noise, can adopt software and hardware to realize under the situation of fully analyzing noise source and characteristics thereof, and the present invention adopts FPGA hardware to realize.

3) feature extraction and coupling: obtain its characteristics of image after the image sequence process denoising that collects, image normalization, dynamic binaryzation, filtering, the refinement, adopt the SIFT algorithm that the image after handling is carried out feature extraction and coupling.

4) estimation: the some feature that obtains after the characteristic matching is carried out three-dimensional reconstruction to sequence, the uncertainty of calculated characteristics point three-dimensional reconstruction.According to unique point locus and uncertain covariance matrix, utilize least square method and maximum-likelihood method that autonomous vehicle is carried out estimation.

Wherein, characteristic matching and estimation are core algorithms of the present invention, adopt the DSP+FPGA software and hardware to realize this core algorithm, have solved the bottleneck of processing speed.

5) result's output: after the motion estimation analysis, the kinematic parameter such as the information such as move distance, direction of car body are exported by output module.

Concrete steps are as follows:

(1) gathers the realtime graphic signal by ccd image sensor.

(2) under the FPGA sequential control, carry out digital sample through high-speed a/d and quantize.

(3) obtain its characteristics of image after the image sequence process denoising that collects, image normalization, dynamic binaryzation, filtering, the refinement, adopt the SIFT algorithm that the image after handling is carried out feature extraction and coupling; The point feature that obtains after the characteristic matching is carried out three-dimensional reconstruction to sequence, store among the RAM by the EDMA passage, DSP carries out calculation process in conjunction with FPGA to image afterwards.

(4) result is exported by output module.Output module comprises LCD, driving pulse as a result.LCD adopts the LCM12864ZK module of Beijing high official position company, and driving pulse adopts FPGA to add the mode of light-coupled isolation.

Claims (2)

1. a stereoscopic vision odometer is characterized in that, is connected successively by following modules:

1). image capture module, adopt binocular CCD camera to take measurand, then with the data output that collects;

2). signal transmission module, the data that the CCD camera collects are carried out preliminary hardware filtering, adopt A/D converter, signal amplifier that the signal of image capture module output is changed amplification, the signal transformation that makes simulation is digital signal, carries out data processing for processor;

3). image analysis module, adopt the embedded system that constitutes by DSP+FPGA as central processing unit, carry out image analysis processing, obtain image feature information, and then through characteristic matching, estimation obtains the body movement parameter;

4). output control module, result to be exported by output control module, output control module comprises LCD, driving pulse.

2. a method of utilizing the described stereoscopic vision odometer of claim 1 to measure is characterized in that, comprises the steps:

A, image acquisition: the view data of gathering autonomous vehicle environment of living in real time by binocular CCD camera; Data stream deposits among the SDRAM through the EDMA passage, finishes image acquisition;

B, pre-service: fundamental purpose is to remove the background and context interference noise, adopts FPGA to realize;

C, feature extraction and coupling: obtain its characteristics of image after the image sequence process denoising that collects, image normalization, dynamic binaryzation, filtering, the refinement, adopt the SIFT algorithm that the image after handling is carried out feature extraction and coupling;

D, estimation: the some feature that obtains after the characteristic matching is carried out three-dimensional reconstruction to sequence, the uncertainty of calculated characteristics point three-dimensional reconstruction; According to unique point locus and uncertain covariance matrix, utilize least square method and maximum-likelihood method that autonomous vehicle is carried out estimation;

E, result's output: after the motion estimation analysis, the kinematic parameter such as the information such as move distance, direction of car body are exported by output control module.

Images