CN101089917A - Quick identification method for object vehicle lane changing - Google Patents

Abstract

A method for quickly identifying lanes-change condition of object vehicle includes utilizing machine vision and radar to collect horizontal and longitudinal distances, relative speed and the same lane possibility between object vehicle and local vehicle, making comparison and quantization on said data collected each time, making compensation and tracking prediction on relevant data, calculating confidence of each vehicle by using two layers of neural network training to obtain object vehicle weight and setting object vehicle to be dangerous object once calculated confidence coefficient is over confidence threshold.

Description

A kind of target carriage is changed the method for quickly identifying under the operating mode

Technical field:

The method for quickly identifying that a kind of target carriage is changed under the operating mode belongs to intelligent vehicle running environment sensing technical field.

Background technology:

In the research field of vertical active safety, picking out risk object or follow the car target from the candidate's vehicle that identifies is the follow-up basis that collision avoidance is reported to the police or driven automatically of carrying out.The vertical active safety system of traditional intelligent vehicle adopts the sensor of radar as target carriage identification mostly, but the alert rate of the mistake of radar is very high, the works of roadside and road top all may be surveyed by the radar flase drop and be the barrier in the running region of this car the place ahead, and its detection visual field is limited, when road had than deep camber, front vehicles was rolled the measurement range of radar easily away from, in addition, when target carriage was changed operation, radar can't be made response timely.

For finishing the target carriage identification mission, document 1 (Knoeppel C, et al.Robust vehicle detection at largedistance using low resolution cameras, Proceedings of the IEEE Intelligent VehiclesSymposium, 2000) use stereoscopic vision that vehicle is discerned, determine virtual lane line according to the driving trace that analog computation goes out, determine track, vehicle place on this basis and finish target carriage identification.Document 2 (Eom T D, et al.Hierarchical object recognition algorithm based on Kalman filter for adaptive cruise controlsystem using scanning laser, 1999) data of obtaining according to the steering wheel angle sensor remove to estimate road curvature, then from radar scanning to vehicle select the most close track center line as target carriage.Above-mentioned two kinds of methods do not have the inspection vehicle diatom, difficult relative position relation between target vehicle and this track are made accurate judgement.Document 3 (CladyX, et al.Object tracking with a Pan-tilt-zoom camera:Application to car driving assistance, 2001) utilize a common camera that the lane line in the place ahead and the vehicle that travels are discerned, and utilize the video camera (Pan-Tilt-Zoom Camera) of a focal length, variable-angle that distant object is followed the tracks of, but system need work in coordination with demarcation to two video cameras, and equipment composition and algorithm are all very complicated.Document 4 (CHU Jiangwei, et al.Studyon method of detecting preceding vehicle based on monocular camera, 2004) at first lane line is detected, utilize then and detect the delta-shaped region of two straight lines formations that obtains, carry out target carriage identification in this zone, problem is to be difficult for the vehicle of cutting this track is made timely reaction.

Said method since merge to utilize fully horizontal, vertically perception information causes the stability of target carriage recognition and tracking not enough and change when operating when target carriage, algorithm is untimely to the reaction of target carriage state variation.

Summary of the invention:

The objective of the invention is to, propose a kind of target carriage and change method for quickly identifying under the operating mode.This method merges and utilizes a plurality of features of vehicle to finish the target carriage identification in conjunction with the recognition result of vehicle and lane line, can realize the accurate recognition and tracking to target carriage, and improves the capability of fast response to the target carriage state variation.

The method is characterized in that this method realizes successively according to the following steps in computing machine:

Step (1), if: each is taken turns and detects the vehicle number that obtains is n, each front truck of taking turns that identification obtains in the detection is i, i=1, ..., n, the front truck vehicle bottom mid point that definition identifies from the image that the method with machine vision obtains is a lateral separation apart from the distance of this car track center line, then: each as the front truck of target carriage through relatively and quantize lateral separation D afterwards _HiFor:

$D_{\mathrm{Hi}}=\frac{D_{\mathrm{HI}\min }}{D_{\mathrm{HIi}}}\&CenterDot;\frac{4-{\mathrm{<figure-callout\; id="4"\; label="D\; HIi"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">D</figure-callout>}}_{\mathrm{HIi}}}{4},$ If D _Hi＜0, then make D _Hi=0,

D _HIminBe the lateral separation that adds up to minimum in n the front truck that calculates according to image information, D _HIiIt is original lateral separation for the front truck i that calculates according to image information;

Each front truck process as target carriage compares and quantification fore-and-aft distance D afterwards _LiFor:

$D_{\mathrm{Li}}=(\frac{D_{\mathrm{LI}\min }-{\mathrm{<figure-callout\; id="20"\; label="D\; LIi"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">D</figure-callout>}}_{\mathrm{LIi}}}{20}+1)/2,$

If: D _Li＜0, D then _Li=0,

D _Li＞1, D then _Li=1,

Each as front truck of target carriage through relatively and after quantizing with respect to the relative velocity V of this car _iFor:

$V_i=(\frac{V_{I\min }-{\mathrm{<figure-callout\; id="10"\; label="V\; Ii"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">V</figure-callout>}}_{\mathrm{Ii}}}{10}+1)/2,$

If: V _i＜0, V then _i=0,

V _i＞1, V then _i=1,

Wherein: D _LIminAnd V _IminThe conduct that is respectively the lateral separation minimum detect target front truck fore-and-aft distance and with respect to the speed of this car;

D _LIiAnd V _IiBe respectively the original fore-and-aft distance and the relative velocity that calculate according to image information and radar data;

Each is in possibility P with the track as the front truck of target carriage and this car _IiObtain according to detections of radar;

Above-mentioned four characteristic quantity D _Hi, D _Li, V _i, P _IiSpan in [0,1] codomain space;

Step (2), the same track possibility data P that step (1) is obtained _IiRevise by following formula, revised is P with the track possibility _i:

$P_i=P_{\mathrm{Ii}}+(P_{\max }-P_{\mathrm{Ii}})\&CenterDot;\sqrt{{\left(\frac{D_{H\max }}{D_{\mathrm{HIi}}}\right)}^2+{\left(\frac{D_{L\max }-D_{\mathrm{LIi}}}{D_{L\max }}\right)}^2},$

If: D _LIi〉=D _Lmax, P then _i=P _Ii,

D _HIi≤ D _Hmax, P then _i=P _Max,

${\left(\frac{D_{H\max }}{D_{\mathrm{HIi}}}\right)}^2+{\left(\frac{D_{L\max }-D_{\mathrm{LIi}}}{D_{L\max }}\right)}^2>1,$ P then _i=P _Max,

Wherein: P _Max, D _Hmax, D _LmaxBe respectively to same take turns radar detection in the detection to a plurality of vehicle targets compare after, same track possibility, lateral separation and the fore-and-aft distance of the target carriage of the same track possibility maximum that selects;

Step (3), the vehicle parameter when utilizing kalman filter method to the generation omission carries out tracking prediction, and vertically the state equation of Kalman filter is:

s(t+Δt)＝A·s(t)+B·w(t)，

Observation equation is:

z(t)＝H·s(t)+v(t)，

Wherein: s (t) is the system state vector of current time, s (t+ Δ t) is next moment state vector of prediction, Δ t is a time step, and z (t) is the systematic observation vector of current time, and A is a system matrix, B is the system noise influence matrix, H is the systematic observation matrix, and w (t) is a system noise, and its variance matrix is Q, v (t) is an observation noise, and its variance matrix is R;

s(t)＝[d _L，v _L，a _L] ^T，z(t)＝[d _L，v _L] ^T，

$A=\left[\begin{array}{ccc}1& \mathrm{\&Delta;t}& \frac{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="t"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}}^2}{2}\\ 0& 1& \mathrm{\&Delta;<figure-callout\; id="0"\; label="t"\; filenames="A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}\\ 0& 0& 1\end{array}\right],$ $B=[\frac{{\mathrm{\&Delta;<figure-callout\; id="3"\; label="t"\; filenames="A20071009996400201.png"\; state="\{\{state\}\}">t</figure-callout>}}^3}{6},\frac{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="t"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}}^2}{2},\mathrm{\&Delta;t}],$ $H=\left[\begin{array}{ccc}1& 0& 0\\ 0& 1& 0\end{array}\right],$

Q＝1， $R=\left[\begin{array}{cc}0.01& 0\\ 0& 0.01\end{array}\right],$

d _LRefer to the relative fore-and-aft distance between front truck and this car, v _LRefer to the relative longitudinal velocity between front truck and this car, a _LRefer to the relative longitudinal acceleration between front truck and this car;

Laterally the form of the state equation of Kalman filter and observation equation is identical with vertical Kalman filter, but quantity of state, observed quantity, system matrix, system noise influence matrix, observing matrix, system noise variance matrix, observation noise variance matrix difference:

s(t)＝[d _H，v _H] ^T，z(t)＝[d _H]

$A=\left[\begin{array}{cc}1& \mathrm{\&Delta;<figure-callout\; id="0"\; label="t"\; filenames="A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}\\ 0& 1\end{array}\right],$ $B=[\frac{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="t"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}}^2}{2},\mathrm{\&Delta;t}],$ H＝[1?0]，Q＝1，R＝0.01，

d _HRefer to the relative lateral separation between front truck and this car, v _HRefer to the relative transverse velocity between front truck and this car;

Step (4), the characteristic quantity of the input p of the described two-layer BP neural network of conduct of each car during weight matrix iw, the threshold vector b that obtains with two-layer BP neural metwork training and the transition function logsig of setting take turns each is handled, and calculates the target carriage degree of confidence W that obtains each car _i:

W _i＝logsig(iw ²logsig(iw ¹p _i+b ¹)+b ²)，

Wherein, p _i=[D _Hi, D _Li, V _i, P _i] ^T

Iw ¹Be the weight matrix of ground floor BP neural network, for:

${\mathrm{<figure-callout\; id="1"\; label="iw"\; filenames="A20071009996400161.png,A20071009996400201.png"\; state="\{\{state\}\}">iw</figure-callout>}}^1=\left[\begin{array}{cccc}-2.41& -6.65& -4.75& 0.28\\ 4.56& 3.62& -1.43& 5.75\\ -0.57& 3.31& 3.50& -6.87\\ -8.12& -0.77& 3.66& -0.61\\ -2.10& 2.77& -0.52& -7.70\end{array}\right],$

Iw ²Be the weight matrix of second layer BP neural network, for:

iw ²＝[-8.58，9.54，-3.06，-5.89，-3.31]，

b ¹Be the threshold vector of ground floor BP neural network, for:

b ¹＝[7.14，-6.03，3.11，3.16，-0.15] ^T，

b ²Be the threshold vector of second layer BP neural network, for:

b ²＝5.15；

Step (5), relatively the degree of confidence W of all vehicle targets _i, the target vehicle of degree of confidence maximum as the candidate target car;

Step (6), when the degree of confidence of candidate target car greater than 0.5 the time, this target carriage is risk object, it is followed the tracks of, otherwise this is taken turns does not have target carriage in the detection.

Evidence: under normal road condition, method among the present invention can be accurately, recognition objective car stably, when target carriage was changed operation, than the target carriage identification of simple dependence radar, the method among the present invention can effectively improve the responding ability to the target carriage state variation.

Description of drawings:

Vehicle data contrast before and after Fig. 1, Kalman follow the tracks of, wherein (a) is the contrast before and after fore-and-aft distance is followed the tracks of, and (b) is the contrast before and after lateral separation is followed the tracks of, and (c) is the contrast before and after vertical relative velocity is followed the tracks of;

Fig. 2, two-layer BP neural network model;

Fig. 3, the training performance curve;

Fig. 4, the target carriage recognition effect, wherein (a) is the target carriage identification (target is arranged) of expressway, (b) is the target carriage identification (driftlessness) of expressway, (c) is the target carriage identification (target carriage incision) of urban road;

Fig. 5 sails out of the target carriage parameter under the operating mode, and wherein (a) is the left and right boundary position of target carriage in image, (b) be the upper and lower boundary position of target carriage in image, (c) being the fore-and-aft distance of target carriage, (d) is the relative velocity of target carriage, (e) is the lateral separation of target carriage;

Fig. 6, the target carriage parameter under the incision operating mode, wherein (a) is the lateral separation of target carriage, (b) is the fore-and-aft distance of target carriage;

Fig. 7, the FB(flow block) of program run.

Embodiment:

Method among the present invention needs 3 steps to finish altogether, at first on comprehensive utilization vehicle and lane identification result's basis, set up the vehicle parameter feature set, utilize the extraction and the optimization method of characteristic parameter that the characteristic quantity in the feature set has been carried out optimization process subsequently, obtained characteristic parameter by two-layer BP neural metwork training at last and influenced weights and threshold value for target carriage identification, and training result is applied to the identification of target carriage, effectively improved the accuracy of target carriage recognizer.

(1) foundation of vehicle parameter feature set

Compare with existing target carriage recognition methods, this method does not rely on identification and the judgement that the single unit vehicle parameter is carried out target carriage not isolatedly, but applied in any combination a plurality of vehicle parameter features of obtaining in the middle of machine vision and the radar data, the advantage of this disposal route comprises two aspects: the first, and utilize the redundancy of discriminative information to improve the stability and the accuracy of target carriage recognition and tracking; The second, owing to merged laterally and environment sensing information longitudinally, make and when operation such as this track can be cut, sail out of to algorithm in target carriage the state variation of target carriage is made timely response.

From the vehicle data information that machine vision and laser radar obtain, can obtain lateral separation, fore-and-aft distance and relative velocity between front truck and this car, and the radar possibility data (In-Lane Probabilty) that also provide front truck to be in this car track, therefore choose this 4 characteristic quantities in the middle of the vehicle data feature set, be labeled as D respectively _Hi(lateral separation relatively), D _Li(fore-and-aft distance relatively), V _i(relative velocity) and P _i(being in the possibility in this track), i=1 ..., n, n are that each takes turns the vehicle fleet that identification obtains in the detection.If directly there are two problems in input as characteristic quantity the raw data such as relative distance between front truck and this car: the data variation scope of (1) several characteristic amount is very big, do not have unified quantization in a codomain scope, be unfavorable for follow-up neural metwork training; (2) when human driver carries out target carriage identification, be actually each is taken turns and detect the result that the candidate's vehicle obtain compares, be not whether isolated ground is the judgement of target to single unit vehicle,, can't reflect this contrast factor if raw data is used as characteristic quantity.For solving this two problems, in the selected characteristic amount, merely process is not chosen in the raw data substitution, but according to each recognition result of taking turns, the comparative result of having introduced data such as relative distance is as characteristic quantity, and 4 characteristic quantity unified quantization have been arrived [0,1] codomain space.

(a) lateral separation index

The vehicle that goes out for radar and Machine Vision Recognition, can both provide the lateral separation between this vehicle and this car in radar data and the view data, but because laser radar given target lateral distance in the detection of a target is to be come out according to the algorithm computation of radar inner setting by a plurality of target afterbody reflection spots that scan, consider the factor of front truck overall width, this is worth common out of true, and in the lateral separation data computing process that radar provides, do not consider lane line information yet.Therefore in the selection of this characteristic quantity of lateral separation, adopted image information, the vehicle bottom mid point that promptly identifies in the computed image is apart from the distance of track center line.If each is taken turns and detects the number of vehicles that obtains is n, for each car, the original lateral separation that calculates according to image information is D _HIi, minimum lateral separation is D in the middle of n the car _HIminThe fore-and-aft distance and the relative velocity of the target carriage of lateral separation minimum are designated as: D _LIminAnd V _IminEach car is D through comparing and quantizing lateral separation afterwards _Hi, it calculates as the formula (1).

$D_{\mathrm{Hi}}=\frac{D_{\mathrm{HI}\min }}{D_{\mathrm{HIi}}}\&CenterDot;\frac{4-D_{\mathrm{HIi}}}{4}---\left(1\right)$

if D _Hi＜0，D _Hi＝0

D _HiCharacterized the hazard level of target aspect lateral separation, its value is big more, and hazard level is high more.In the comparison quantizing process of lateral separation, at first considered the same contrast of taking turns in the testing process, promptly more near the target of minimal transverse distance, the lateral separation index is big more; Also considered simultaneously the absolute lateral separation of vehicle, promptly the value of being somebody's turn to do is more little, and hazard level is high more.As lateral separation D _HIiSurpass after the 4m lateral separation index D _HiBe designated as 0.Through type (1) can be original lateral separation D _HIiComparative quantity turns to lateral separation index D _Hi, and the codomain scope of this index is [0,1].

(b) fore-and-aft distance and relative velocity index

The longitudinal direction of car data that laser radar obtains are more accurate and stable, therefore with the foundation of radar data as fore-and-aft distance and relative velocity in the extraction feature set.When radar generation omission, when having recognized the vehicle of omission by image is additional, because the data of this vehicle can't provide by radar, this moment, fore-and-aft distance and relative velocity were provided by image information.Vehicle bottom mid point that fore-and-aft distance directly identifies in the computed image and the distance between the Ben Che.The fore-and-aft distance data in the middle of this two field picture and continuous 3 two field pictures have in the past been used in the calculating of relative velocity, calculate 3 velocity amplitudes then, use the wave filter of [0.25,0.25,0.5] to obtain relative velocity.

Original fore-and-aft distance is designated as D _LIi, the fore-and-aft distance index D after relatively quantizing _LiIt calculates as the formula (2).

$D_{\mathrm{Li}}=(\frac{D_{\mathrm{LI}\min }-{\mathrm{<figure-callout\; id="20"\; label="D\; LIi"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">D</figure-callout>}}_{\mathrm{LIi}}}{20}+1)/2$

if?D _Li＜0，D _Li＝0 (2)

if?D _Li＞1，D _Li＝1

The comparison other of fore-and-aft distance index is the target carriage of lateral separation minimum.

Original relative velocity is designated as V _Ii, the relative velocity index V after relatively quantizing _iIt calculates as the formula (3).

$V_i=(\frac{V_{I\min }-{\mathrm{<figure-callout\; id="10"\; label="V\; Ii"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">V</figure-callout>}}_{\mathrm{Ii}}}{10}+1)/2$

if?V _i＜0，V _i＝0 (3)

if?V _i＞1，V _i＝1

The comparison other of relative velocity index equally also is the target carriage of lateral separation minimum.

After relatively quantizing, can introduce feature set to same relative hazard degree of taking turns between many cars that identify in the middle of the detection, and the characteristic quantity unified quantization has been arrived [0,1] space.

(c) with the track possibility

Radar provides " In-Lane Probability " data for the vehicle target that detects, this data characterization vehicle target be in the possibility of this car with the track, so these data have also been taken in the foundation of selecting as risk object in the middle of the vehicle characteristics collection.When radar generation omission, when having recognized the vehicle of omission by image is additional, these data are in this track by the vehicle that recognizes in the image width obtains divided by the car load width.

(2) optimization process of vehicular characteristics data

(a) DCF with track possibility data revises

Radar can not be made timely reaction to the incision vehicle usually, show with on these data of track possibility, following situation will appear: as two cars, fore-and-aft distance vehicle far away is in the centre in this track, the near vehicle of fore-and-aft distance has begun to cut this track, and in radar data, the same track possibility numerical value of this two cars may be mutually far short of what is expected, fore-and-aft distance vehicle far away its with the track possibility near 1, but the near vehicle of fore-and-aft distance is but near 0.These data do not conform to the hazard level of actual road conditions and vehicle target.

At document 5 (LIU Y, et al.Nash strategies with distance discount factor in target selectionproblems, 2004) under the inspiration of designed DDF (Distance Discount Factor) method, designed the compensated distance factor (DCF, Distance Compensation Factor), distance factor to vehicle target takes in, and revises same track possibility in the middle of the radar data with this, makes it can realistic better road conditions.

At first, to same take turns detect in the middle of radar detection to a plurality of vehicle targets compare, choose target with track possibility maximum, its same track possibility, lateral separation, fore-and-aft distance is designated as respectively: P _Max, D _Hmax, D _LmaxSame track possibility in the middle of the raw data is designated as P _Ii, be designated as P with the track possibility through DCF is revised _iThe DCF correction formula as the formula (4).

$P_i=P_{\mathrm{Ii}}+(P_{\max }-P_{\mathrm{Ii}})\&CenterDot;\sqrt{{\left(\frac{D_{H\max }}{D_{\mathrm{HIi}}}\right)}^2+{\left(\frac{D_{L\max }-D_{\mathrm{LIi}}}{D_{L\max }}\right)}^2}$

if?D _LIi≥D _Lmax，P _i＝P _Ii

if?D _HIi≤D _Hmax，P _i＝P _max (4)

if ${\left(\frac{D_{H\max }}{D_{\mathrm{HIi}}}\right)}^2+{\left(\frac{D_{L\max }-D_{\mathrm{LIi}}}{D_{L\max }}\right)}^2>1,$ P _i＝P _max

Than DDF, it is the one dimension modifying factor that DCF has following characteristics: DDF, and DCF has adopted two-dimentional modifying factor, has taken all factors into consideration the influence factor of lateral separation and fore-and-aft distance; DDF is a discount factor, promptly by adopting DDF to revise, can reduce the importance of distant object.And DCF is a compensating factor, the mxm. of this track possibility normally accurately in the middle of the radar raw data, but vehicle for incision, radar can not in time reflect this variation, make that this track possibility numerical value of incision vehicle is very low, DCF considers distance factor, and the same track possibility data of target have been carried out compensation correction.

(b) Kalman's tracking prediction

Using Kalman filtering mainly is to carry out tracking prediction for the vehicle parameter when omission takes place.

The information that fusion machine vision and laser radar obtain is carried out in the process of vehicle identification, the situation of vehicle omission may take place, for improving the accuracy rate and the stability of vehicle identification, when carrying out vehicle identification, for initial probe to vehicle it is not defined as vehicle target, this car is just determined its target identities when the second time, continuous probe arrived, and it is listed among the tracking-files.Vehicle in tracking-files is not identified then, can at once it not deleted from archives yet, but continue tracking prediction is carried out in its position, up to failing to survey this car 10 times continuously, just it is deleted from archives.

In the present invention, suppose between this car and the front truck relative longitudinal acceleration and relatively the variation of transverse velocity be stably, therefore selected for use 3 rank Kalman filter that longitudinal data is carried out tracking prediction, 2 rank Kalman filter are carried out tracking prediction to horizontal data.

Vertically the state equation of Kalman filter as the formula (5), observation equation is as the formula (6).

s(t+Δt)＝A·s(t)+B·w(t) (5)

z(t)＝H·s(t)+v(t) (6)

Wherein: s (t) is the system state vector of current time, s (t+ Δ t) is next moment state vector of prediction, Δ t is a time step, and z (t) is the systematic observation vector of current time, and A is a system matrix, B is the system noise influence matrix, H is the systematic observation matrix, and w (t) is a system noise, and its variance matrix is Q, v (t) is an observation noise, and its variance matrix is R;

s(t)＝[d _L，v _L，a _L] ^T，z(t)＝[d _L，v _L] ^T

$A=\left[\begin{array}{ccc}1& \mathrm{\&Delta;t}& \frac{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="t"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}}^2}{2}\\ 0& 1& \mathrm{\&Delta;<figure-callout\; id="0"\; label="t"\; filenames="A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}\\ 0& 0& 1\end{array}\right],$ $B=[\frac{{\mathrm{\&Delta;<figure-callout\; id="3"\; label="t"\; filenames="A20071009996400201.png"\; state="\{\{state\}\}">t</figure-callout>}}^3}{6},\frac{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="t"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}}^2}{2},\mathrm{\&Delta;t}],$ $H=\left[\begin{array}{ccc}1& 0& 0\\ 0& 1& 0\end{array}\right]$

Q＝1， $R=\left[\begin{array}{cc}0.01& 0\\ 0& 0.01\end{array}\right]$

d _LRefer to the relative fore-and-aft distance between front truck and this car, v _LRefer to the relative longitudinal velocity between front truck and this car, a _LRefer to the relative longitudinal acceleration between front truck and this car.

Laterally the form of the state equation of Kalman filter and observation equation is identical with vertical Kalman filter, but quantity of state, observed quantity, system matrix, system noise influence matrix, observing matrix, system noise variance matrix, observation noise variance matrix difference.

s(t)＝[d _H，v _H] ^T，z(t)＝[d _H]

$A=\left[\begin{array}{cc}1& \mathrm{\&Delta;<figure-callout\; id="0"\; label="t"\; filenames="A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}\\ 0& 1\end{array}\right],$ $B=[\frac{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="t"\; filenames="A20071009996400161.png,A20071009996400191.png"\; state="\{\{state\}\}">t</figure-callout>}}^2}{2},\mathrm{\&Delta;t}],$ H＝[1?0]，Q＝1，R＝0.01

d _HRefer to the relative lateral separation between front truck and this car, v _HRefer to the relative transverse velocity between front truck and this car.

In the specific implementation process, under the situation that omission does not take place, (promptly there is observation data to arrive), utilize observed quantity that Kalman filtering gain matrix and evaluated error association square matrix are carried out the state renewal, thereby obtain the optimal estimation of current system state value, when omission (not having observation data) has taken place the time, need current system state value be carried out step forecast estimate based on last round of system state estimation value.The form of state renewal equation and predictive equation can referring to document 6 (Hou Dezao. automobile longitudinal is the research of anti-collision system initiatively, 2004) in formula (6-19) to (6-23).

The tracking prediction effect of Kalman filter as shown in Figure 1.As can be seen from the figure, lost target in the raw data for some time, the situation of omission has promptly taken place, by Kalman's tracking prediction, can estimate the data of vehicle preferably, thereby improve the stability of vehicle identification, reduce loss, the characteristic quantity that the data after the process Kalman filtering also can be discerned as target carriage better.

(3) BP neural metwork training and target carriage judgement

Through vehicle longitudinal spacing, lateral separation, the relative velocity of Kalman's tracking prediction and the feature set of having formed target carriage identification through these 4 characteristic quantities of same track possibility that DCF revises jointly, for obtaining the weighing factor that characteristic quantity is discerned for target carriage in the feature set, document 7 (Zhou Kaili have been used, Deng. neural network model and the design of MATLAB simulated program thereof, 2005) the two-layer BP neural network model in is trained sample set, and its structure as shown in Figure 2.

The meaning of each symbol is as follows among the figure: R: the dimension of input vector; S: neuronic number; P: input vector; Iw: weight matrix; B: threshold vector; N: weighted sum; A: output vector; F: transition function.

The superscript of each symbol is represented the sequence number of network layer.In concrete the application, input vector is 4 dimensions, and output vector is 1 dimension, wherein:

p＝[D _H，D _L，V，P] ^T (7)

a ¹＝f ¹(iw ¹p+b ¹) (8)

a ²＝f ²(iw ²a ¹+b ²) (9)

In formula (7), the input of neural network comprises 4 parameters, is respectively the lateral separation, fore-and-aft distance of vehicle target, longitudinal velocity and be in possibility with the track, the output that formula (8) and (9) have been calculated two-layer neural network respectively, a with this car relatively ²Also be the output of whole network, this vector is made of single-element, and whether directly be used for adjudicating vehicle target is the target carriage of travelling in the place ahead.

The training parameter that needs to determine in the use of neural network comprises the number of hidden nodes, two-layer transition function and training function.In the selection of the number of hidden nodes, still there is not good analytic expression to find the solution at present, utilize the experimental formula in the document 7 (Zhou Kaili is etc. neural network model and the design of MATLAB simulated program thereof, 2005) to select usually.

$k=\sqrt{k_i+k_o}+a---\left(10\right)$

In the formula (10): k: the number of hidden nodes; k _i: input number of nodes; k _o: the output node number; Constant between a:1～10.

The number of hidden nodes is set at 5, and two-layer transition function has all been selected the logsig function, and the training function is set at traingdx, comprises 646 vehicle datas in the sample set altogether.The input sample is corresponding with input vector, and first element of vehicle sample is a lateral separation, and second element is fore-and-aft distance, and the 3rd element is relative longitudinal velocity, and the 4th element is that target and Ben Che are in the possibility with the track.Weights that obtain after the training and threshold value are as the formula (11).

${\mathrm{iw}}^1=\left[\begin{array}{cccc}-2.41& -6.65& -4.75& 0.28\\ 4.56& 3.62& -1.43& 5.75\\ -0.57& 3.31& 3.50& -6.87\\ -8.12& -0.77& 3.66& -0.61\\ -2.10& 2.77& -0.52& -7.70\end{array}\right]---\left(11\right)$

b ¹＝[7.14，-6.03，3.11，3.16，-0.15] ^T

iw ²＝[-8.58，9.54，-3.06，-5.89，-3.31]

b ²＝5.15

The performance curve of training process as shown in Figure 3.

Weights that obtain after the training and threshold value can be used in the identification of target carriage.In actual application, each is taken turns all vehicles that recognize in the detection, according to weights and threshold value the characteristic quantity of each car is handled respectively, calculate the target carriage degree of confidence that obtains each car, as the formula (12).

w _i＝logsig(iw ²logsig(iw ¹p _i+b ¹)+b ²) (12)

In the formula: p _i: the parameter vector of i vehicle target; W _i: the target carriage degree of confidence of i vehicle target.

The target carriage degree of confidence of all vehicles relatively, the vehicle of degree of confidence maximum as the candidate target car.Compare the degree of confidence of candidate target car and the magnitude relationship of decision threshold at last, during greater than decision threshold, this candidate target car is exactly with car target or risk object, otherwise this wheel does not have target carriage in detecting.The decision threshold of choosing among the present invention is 0.5, and judging threshold also can connect with driver's personal habits, and conservative driver can set decision threshold smallerly.

Specific embodiment:

At home on the Jetta AT type car platform, finish the perception of environmental information by monocular-camera and laser radar, monocular-camera has been selected German BaslerA601fCMOS video camera for use, the CAN bus radar that laser radar has selected for use Denso company to produce.

The effect of target carriage identification as shown in Figure 4.Among the figure, black line is the lane line position of identifying, and white box is the vehicle location that identifies, and black box is the target carriage position that identification obtains.As can be seen, after identifying a plurality of vehicles, algorithm can select exactly with car target or risk object, when many cars that travel on the road surface all do not constitute a threat to this car, algorithm also can be made accurately and judging, and in car when incision, arranged forwardly, algorithm can be made response comparatively in time, with the vehicle of incision as following the car target.

The data of target carriage in the test have been carried out record, target carriage sail out of under this track situation parameter as shown in Figure 5.As can be seen from the figure, along with this this car of spacing is more and more far away, distance between its border, left and right sides and the distance between the up-and-down boundary obviously diminish, and have reflected target near big and far smaller characteristic in image.The target choosing method that among Fig. 5 (e) target carriage is sailed out of under this track situation contrasts, A line among the figure is to utilize many Feature Fusion method of the present invention to judge that former target carriage no longer is the switching time with the car target, the B line then is the switching time that only utilizes radar information to obtain, as can be seen, when target carriage is sailed out of this track gradually, fusion method can in time be found sailing out of of target carriage, and more realistic road conditions only rely on radar data that long sluggishness then can take place.

When vehicle was cut this track, the data and curves of target vehicle as shown in Figure 6.As can be seen from the figure, when only relying on radar information, the vehicle of finding incision is than later, switch to the new car target of following and also bigger sluggishness can take place, the method of utilizing the present invention to design then can be found the vehicle of incision intention rapidly and will switch to the vehicle of new incision with the car target, thereby further improve the safety coefficient of longitudinal drive person's backup system.

The program flow chart of entire method as shown in Figure 7.

Claims (2)

1. a target carriage is changed the method for quickly identifying under the operating mode, it is characterized in that this method realizes in computing machine successively according to the following steps:

Step (1), if: each is taken turns and detects the vehicle number that obtains is n, each front truck of taking turns that identification obtains in the detection is i, i=1, ..., n, the front truck vehicle bottom mid point that definition identifies from the image that obtains with machine vision method is a lateral separation apart from the distance of this car track center line, then: each as the front truck of target carriage through relatively and the lateral separation D after quantizing _HiFor:

$D_{\mathrm{Hi}}=\frac{D_{\mathrm{HI}\min }}{D_{\mathrm{HIi}}}\&CenterDot;\frac{4-D_{\mathrm{HIi}}}{4},$ If D _Hi＜0, then make D _Hi=0,

D _HIminBe the lateral separation that adds up to minimum in n the front truck that calculates according to image information, D _HIiIt is original lateral separation for the front truck i that calculates according to image information;

Each front truck process as target carriage compares and quantification fore-and-aft distance D afterwards _LiFor:

$D_{\mathrm{Li}}=(\frac{D_{\mathrm{LI}\min }-D_{\mathrm{LIi}}}{20}+1)/2,$

If: D _Li＜0, D then _Li=0,

D _Li＞1, D then _Li=1,

Each as front truck of target carriage through relatively and after quantizing with respect to the relative velocity V of this car _iFor:

$V_i=(\frac{V_{I\min }-V_{\mathrm{Ii}}}{10}+1)/2,$

If: V _i＜0, V then _i=0,

V _i＞1, V then _i=1,

Wherein: D _LIminAnd V _IminThe conduct that is respectively the lateral separation minimum detect target front truck fore-and-aft distance and with respect to the speed of this car;

D _LIiAnd V _IiBe respectively the original fore-and-aft distance and the relative velocity that calculate according to image information and radar data;

Each is in possibility P with the track as the front truck of target carriage and this car _IiObtain according to detections of radar;

Above-mentioned four characteristic quantity D _Hi, D _Li, V _i, P _IiSpan in [0,1] codomain space;

Step (2), the same track possibility data P that step (1) is obtained _IiRevise by following formula, revised is P with the track possibility _i:

$P_i=P_{\mathrm{Ii}}+(P_{\max }-P_{\mathrm{Ii}})\&CenterDot;\sqrt{{\left(\frac{D_{H\max }}{D_{\mathrm{HIi}}}\right)}^2+{\left(\frac{D_{L\max }-D_{\mathrm{LIi}}}{D_{L\max }}\right)}^2},$

If: D _LIi〉=D _Lmax, P then _i=P _Ii,

D _HIi≤ D _Hmax, P then _i=P _Max,

${\left(\frac{D_{H\max }}{D_{\mathrm{HIi}}}\right)}^2+{\left(\frac{D_{L\max }-D_{\mathrm{LIi}}}{D_{L\max }}\right)}^2>1,$ P then _i=P _Max,

Wherein: P _Max, D _Hmax, D _LmaxBe respectively to same take turns radar detection in the detection to a plurality of vehicle targets compare after, same track possibility, lateral separation and the fore-and-aft distance of the target carriage of the same track possibility maximum that selects;

Step (3), the vehicle parameter when utilizing kalman filter method to the generation omission carries out tracking prediction, and vertically the state equation of Kalman filter is:

s(t+Δt)＝A·s(t)+B·w(t)，

Observation equation is:

z(t)＝H·s(t)+v(t)，

Wherein: s (t) is the system state vector of current time, s (t+ Δ t) is next moment state vector of prediction, Δ t is a time step, and z (t) is the systematic observation vector of current time, and A is a system matrix, B is the system noise influence matrix, H is the systematic observation matrix, and w (t) is a system noise, and its variance matrix is Q, v (t) is an observation noise, and its variance matrix is R;

s(t)＝[d _L，v _L，a _L] ^T，z(t)＝[d _L，v _L] ^T，

$A=\left[\begin{array}{ccc}1& \mathrm{\&Delta;t}& \frac{\mathrm{\&Delta;}{t}^2}{2}\\ 0& 1& \mathrm{\&Delta;t}\\ 0& 0& 1\end{array}\right],$ $B=[\frac{\mathrm{\&Delta;}{t}^3}{6},\frac{\mathrm{\&Delta;}{t}^2}{2},\mathrm{\&Delta;t}],$ $H=\left[\begin{array}{ccc}1& 0& 0\\ 0& 1& 0\end{array}\right],$

Q＝1 $R=\left[\begin{array}{cc}0.01& 0\\ 0& 0.01\end{array}\right],$

d _LRefer to the relative fore-and-aft distance between front truck and this car, v _LRefer to the relative longitudinal velocity between front truck and this car, a _LRefer to the relative longitudinal acceleration between front truck and this car;

Laterally the form of the state equation of Kalman filter and observation equation is identical with vertical Kalman filter, but quantity of state, observed quantity, system matrix, system noise influence matrix, observing matrix, system noise variance matrix, observation noise variance matrix difference:

s(t)＝[d _H，v _H] ^T，z(t)＝[d _H]，

$A=\left[\begin{array}{cc}1& \mathrm{\&Delta;t}\\ 0& 1\end{array}\right],$ $B=[\frac{\mathrm{\&Delta;}{t}^2}{2},\mathrm{\&Delta;t}],$ H＝[1?0]，Q＝1，R＝0.01，

d _HRefer to the relative lateral separation between front truck and this car, v _HRefer to the relative transverse velocity between front truck and this car;

Step (4), the characteristic quantity of the input p of the described two-layer BP neural network of conduct of each car during weight matrix iw, the threshold vector b that obtains with two-layer BP neural metwork training and the transition function logsig of setting take turns each is handled, and calculates the target carriage degree of confidence W that obtains each car _i:

W _i＝logsig(iw ²logsig(iw ¹p _i+b ¹)+b ²)，

Wherein, p _i=[D _Hi, D _Li, V _i, P _i] ^T

Iw ¹Be the weight matrix of ground floor BP neural network, for:

${\mathrm{iw}}^1=\left[\begin{array}{cccc}-2.41& -6.65& -4.75& 0.28\\ 4.56& 3.62& -1.43& 5.75\\ -0.57& 3.31& 3.50& -6.87\\ -8.12& -0.77& 3.66& -0.61\\ -2.10& 2.77& -0.52& -7.70\end{array}\right],$

Iw ²Be the weight matrix of second layer BP neural network, for:

iw ²＝[-8.58，9.54，-3.06，-5.89，-3.31]，

b ¹Be the threshold vector of ground floor BP neural network, for:

b ¹＝[7.14，-6.03，3.11，3.16，-0.15] ^T，

b ²Be the threshold vector of second layer BP neural network, for:

b ²＝5.15；

Step (5), relatively the degree of confidence W of all vehicle targets _i, the target vehicle of degree of confidence maximum as the candidate target car;

Step (6), when the degree of confidence of candidate target car greater than 0.5 the time, this target carriage is risk object, it is followed the tracks of, otherwise this is taken turns does not have target carriage in the detection.

2. a kind of target carriage according to claim 1 is changed the method for quickly identifying under the operating mode, it is characterized in that, when calculating target carriage, use the fore-and-aft distance in the middle of this two field picture and continuous three two field pictures in the past to the relative velocity of this car, calculate three velocity amplitudes then, use [0.25 again, 0.25,0.5] wave filter obtain relative velocity.

Images