CN107705577A - A kind of real-time detection method and system based on lane line demarcation vehicle peccancy lane change - Google Patents

Abstract

The present invention relates to a real-time detection method for illegal lane changes of vehicles based on lane line marking, specifically comprising: step 1) marking the lane line L, calculating and obtaining the detection rectangle area D and image I; step 2) the detection rectangle in step 1) In area D, based on the deep convolutional neural network, the vehicle set A of image I is detected = {A ₁ , A ₂ ,...,A _i }; step 3) According to the vehicle set A obtained in step 2), filter out the vehicle set A={A 1 ,A 2 ,...,A i }; Vehicle set B={B ₁ ,B ₂ ,...,B _i } where the lane line L intersects, where, Then match with the tracked vehicle list TL, and update the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j }; step 4) judge whether the tracked vehicle T _j in the updated tracked vehicle list TL is Vehicles changing lanes illegally; if the tracking vehicle T _j is a vehicle changing lanes illegally, mark the position of the tracking vehicle T _j , and delete the information of the illegal lane changing vehicle T _j from the updated tracking vehicle list TL.

Description

A real-time detection method and system for illegal lane changes of vehicles based on lane line calibration

technical field

The invention belongs to the technical field of intelligent transportation systems and image recognition, and in particular relates to a real-time detection method and system for illegally changing lanes of vehicles based on lane line marking.

Background technique

With the progress and development of society, the number of urban motor vehicles is increasing day by day, and the increasing number of motor vehicles makes traffic accidents more and more serious. In recent years, the economic losses and death toll caused by road traffic accidents are also increasing. Controlling the occurrence of traffic accidents has become a problem that traffic management departments pay more and more attention to, and the primary cause of traffic accidents is automobile violations. As a common traffic violation, illegal lane changes will not only cause traffic congestion, but may even cause serious traffic accidents. In order to reduce the incidence of traffic accidents, traffic management departments continue to promote intelligent traffic management systems, which are based on The illegal lane change detection technology of surveillance video is more critical, so it is necessary to study the detection of vehicle illegal lane change based on surveillance video.

There are two main types of existing detection methods for vehicle violations and lane changes: one is based on spatial distance, such as laser detection, infrared detection and ultrasonic detection. The other is based on computer vision technology, which has the advantages of simple installation and maintenance, high visibility, and high detection accuracy.

At present, the detection method of illegal lane-changing vehicles based on computer vision technology is to first extract the moving target set, then find out the moving vehicle set and location information, and finally judge whether the dispersion of each vehicle’s trajectory exceeds the set threshold, so as to judge Whether the vehicle changed lanes illegally. The existing methods have a better detection effect when there are fewer vehicles on the road and the vehicles are moving faster; however, they cannot deal with illegal lane changes when the vehicles are moving slowly or there are many vehicles on the road. In addition, existing methods are susceptible to factors such as illumination, occlusion, shadows, and video jitter; when there are many vehicles, the way to track each vehicle is too computationally intensive to meet real-time requirements. Therefore, for the detection of illegal lane-changing vehicles, there is an urgent need for a real-time detection method for illegal lane-changing vehicles with wide application range, high accuracy, and fast detection speed, which can meet real-time requirements.

Contents of the invention

The purpose of the present invention is to solve the above-mentioned defects in the existing real-time detection method for illegal lane changes of vehicles. The results are stable and reliable, with high accuracy and fast detection speed, which can meet the real-time detection requirements of vehicles changing lanes illegally. In addition, the method can realize fast and accurate identification of illegal lane-changing vehicles in the video.

In order to achieve the above object, the present invention provides a real-time detection method for illegal lane changes of vehicles based on lane line calibration, which specifically includes:

Step 1) mark the lane line L, calculate and obtain the detection rectangular area D and the image I;

Step 2) In the detection rectangular area D of step 1), the vehicle set A={A ₁ , A ₂ ,...,A _i } of the image I is detected based on the deep convolutional neural network;

Step 3) According to the vehicle set A obtained in step 2), filter out the vehicle set B that intersects the lane line L={B ₁ , B ₂ ,...,B _i }, where, Then match with the tracked vehicle list TL, and update the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j };

Step 4) Determine whether the tracking vehicle T _j in the updated tracking vehicle list TL is an illegal lane-changing vehicle; if the tracking vehicle T _j is an illegal lane-changing vehicle, then mark the position of the tracking vehicle T _j , and use Delete the information of illegal lane-changing vehicles T _j from the tracked vehicle list TL.

Described step 1) specifically comprises:

Step 1-1) Mark the lane line L={L ₁ ,L ₂ ,...,L _j }, the position of each lane line L _j includes: starting point pb _j ={x _bj ,y _bj } and end point pe _j ={x _ej , y _ej }, wherein, x _bj , y _bj are the horizontal and vertical coordinates of the starting point respectively, x _ej , y _ej are the horizontal and vertical coordinates of the end point respectively;

Step 1-2) Calculate the detection area rectangle D={x, y, w, h} according to the position of the lane line in step 1-1), where,

x＝min(x _b1 ,x _e1 ,x _b2 ,x _e2 ,...,x _bj ,x _ej )

y=min(y _b1 ,y _e1 ,y _b2 ,y _e2 ,...,y _bj ,y _ej )

w=max(x _b1 ,x _e1 ,x _b2 ,x _e2 ,...,x _bj ,x _ej )-x

h=max(y _b1 ,y _e1 ,y _b2 ,y _e2 ,...,y _bj ,y _ej )-y

Step 2) specifically includes:

According to the detection area D and image I in step 1), the detection area D in the image I is normalized to 576×576 as the input image, and the vehicle set A in the image I is detected based on the deep convolutional neural network = {A ₁ ,A ₂ ,...,A _i }, the position of A _i is recorded as PA _i ={ _xi ,y _i ,w _i ,h _i }, the detection effect is shown in Figure 2(b);

Among them, the convolutional neural network includes: 7 convolutional layers, 4 downsampling layers, 2 fully connected layers and an output layer; the scanning boundary in the convolutional layer is automatically filled with 0, and the Leaky-ReLu function is used to Activation; maximum pooling is used in the downsampling layer.

The convolution kernel size of the convolution layer C1 is 9×9, 32 convolution kernels, the step size is 2, and the size of the generated feature map is 288×288; the window size of the downsampling layer S1 is 4×4, and the step size is 4, The size of the generated feature map is 72×72; the size of the convolutional layer C2 convolution kernel is 1×1, 4 convolution kernels, the step size is 1, and the size of the generated feature map is 72×72; the size of the convolutional layer C3 convolution kernel is 3×3, 8 convolution kernels, the step size is 1, and the size of the generated feature map is 72×72; the size of the downsampling layer S2 window is 2×2, the step size is 2, and the size of the generated feature map is 36×36; The convolution layer C4 convolution kernel size is 1×1, 8 convolution kernels, the step size is 1, and the generated feature map size is 36×36; the convolution layer C5 convolution kernel size is 3×3, 16 convolutions Kernel, the step size is 1, and the size of the generated feature map is 36×36; the window size of the downsampling layer S3 is 2×2, the step size is 2, and the size of the generated feature map is 18×18; the size of the convolution layer C6 convolution kernel is 3×3, 32 convolution kernels, the step size is 1, and the size of the generated feature map is 18×18; the size of the downsampling layer S4 window is 2×2, the step size is 2, and the size of the generated feature map is 9×9; The size of the multilayer C7 convolution kernel is 1×1, 64 convolution kernels, the step size is 1, and the size of the generated feature map is 9×9; the fully connected layer F1 is composed of 256 neurons, and the neurons are processed using the Relu function. Activation; the fully connected layer F2 is composed of 4096 neurons, and the neurons are activated using the Leaky-ReLu function; the output layer is composed of 891 neurons, and the neurons are activated using the Relu function.

Described step 3) specifically comprises:

Step 3-1) Based on the vehicle set A={A ₁ ,A ₂ ,...,A _i } obtained in step 2), select the vehicle set B={B ₁ ,B ₂ , ...,B _i }, where,

Step 3-2) Traverse the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j } according to the vehicle set B obtained in step 3-1), and select the vehicle set C whose coincidence degree U _ij >0.2 , C＝{C ₁ ,C ₂ ,...,C _m }; calculate the Hog features hog _j and hog _k , and then calculate the cosine distance Cos _jk ; then select the largest cosine distance between the tracking vehicle T _j and the vehicle set C Cos _jm and the corresponding vehicle C _m , if Cos _jm > 0.75, then T _j and C _m match each other, and enter step 3-3); if Cos _jm ≤ 0.75, then there is no tracking vehicle T _j in the vehicle set C matching vehicle, and proceed to steps 3-4);

Step 3-3) According to the vehicle C _m obtained in step 3-2), update the vehicle position of the tracking vehicle T _j to the current vehicle position, that is, TP _j =CP _m ; update the tracking vehicle to the current vehicle picture, that is, TM _j =CM _m ; consecutive follow-up times TT _j = 0, the horizontal directional distance TD _j between the tracking point and the lane line L is updated to the horizontal directional distance between the current tracking point and the lane line L, that is, TD _j = CD _m ; at this time C _m Has been matched, whether the current vehicle is matched flag CF _i = true;

Step 3-4) Update the number of consecutive lost times TT _j ' of the tracking vehicle T _j , where TT _j '=TT _j + 1, if the number of consecutive lost times of TT _j ' is greater than 3 times, delete it from the list of tracking vehicles Vehicle T _j ;

Step 3-5) Traversing the vehicle set B, taking the unmatched vehicle B _u ={BP _u ,BM _u ,BF _u ,BQ _u ,BD _u } as the new tracking vehicle T _u ={TP _u ,TM _u , TT _u , TOD _u , TD _u } are added to the tracking vehicle list TL, and the vehicle tracking list TL is updated, wherein,

TP _u = BP _u

TM _u =BM _u

TT _u =0

TOD _u = BD _u

_TDu = _BDu .

The step 3-1) specifically includes:

Step 3-1-1) According to the lane line L obtained in step 1) and step 2) and the vehicle A _i position rectangle PA _i , calculate the intersection of the lane line L _j and the four sides of the vehicle A _i position rectangle PA _i , and the lane line L The intersection points of _j and the left edge, right edge, upper edge, and lower edge of the rectangle are respectively pl _ij ={x _lij ,y _lij }, pr _ij ={x _rij ,y _rij }, pu _ij ={x _uij ,y _uij }, pd _ij ={x _dij ,y _dij };

Among them, {x _lij , y _lij } are the abscissa and ordinate of the left edge intersection respectively, {x _rij ,y _rij } are the abscissa and ordinate of the right edge intersection respectively, {x _uij ,y _uij } are the upper edge respectively The abscissa and ordinate of the intersection point are respectively the abscissa and ordinate of the intersection point of the lower edge.

Step 3-1-2) According to the intersection of the lane line and the rectangle obtained in step 3-1-1), if the intersection is on the top and left of the rectangle and meets the following preset conditions, add the vehicle rectangle A _i to the vehicle set B;

x _i + w _i /4 < x _uij < x _i + w _i

y _i +h _i /2<y _lij <y _i +h _i

Step 3-1-3) According to the intersection point of the lane line and the rectangle obtained in step 3-1-1), if the intersection point is on the upper side and the right side of the rectangle and meets the following preset conditions, add the vehicle rectangle A _i to the vehicle set B;

x _i < x _uij < x _i +3w _i /4

y _i +h _i /2<y _rij <y _i +h _i

Step 3-1-4) According to the intersection point of the lane line and the rectangle obtained in step 3-1-1), if the intersection point is on the upper and lower sides of the rectangle and meets the following preset conditions, add the vehicle rectangle A _i to the vehicle set B;

x _i +w _i /4<x _uij <x _i +3w _i /4

x _i +w _i /4<x _dij <x _i +3w _i /4

The step 3-2) specifically includes:

Step 3-2-1) According to the vehicle set B obtained in step 3-1), initialize the vehicle information, which includes: vehicle position BP _i ={ _xi , y _i , w _i , h _i }, vehicle picture BM _i , Whether the current vehicle is matched with the flag BF _i = false, the tracking point BQ _i = ( _xi + w _i /2, y _i +4h _i /5) and the horizontal directional distance between the tracking point BQ _i and each lane line L _j Bd _ij , that is, vehicle B _i ={BP _i ,BM _i ,BF _i ,BQ _i ,BD _i }, where BD _i ={Bd _i1 ,Bd _i2 ,...,Bd _ij };

Step 3-2-2) Traverse the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j }, T _j ={TP _j ,TM _j ,TT _j ,TOD _j ,TD _j }, T _j ∈TL, where TP _j represents the vehicle position, TM _j represents the vehicle picture, TT _j represents the number of consecutive follow-ups, TOD _j represents the horizontal distance between the initial tracking point and the lane line L, TD _j represents the current tracking point and the lane Horizontal directional distance of line L; TOD _j = {TOd _j1 , TOd _j2 ,..., TOd _jl }, TD _j = {Td _j1 , Td _j2 ,..., Td _jl }; calculate the tracking vehicle T _j The coincidence degree U _ij of the vehicle position TP _j and each unmatched vehicle position BP _i in the vehicle set B:

U _ij ＝S _cij /S _zij

Among them, S _zij is the union of the rectangular areas of BP _i and TP _j , and S _cij is the intersection of the rectangular areas of BP _i and TP _j ;

Select the vehicle set C whose coincidence degree U _ij >0.2, C={C ₁ ,C ₂ ,...,C _m }; if C is empty, there is no vehicle matching the tracking vehicle T _j in the vehicle set B, and enter step 3-4); if C is not If it is empty, go to step 3-2-3);

Step 3-2-3) According to the vehicle set C={C ₁ ,C ₂ ,...,C _m } obtained in step 3-2-2), the vehicle picture TM _j of the tracking vehicle T _j and the vehicle set C The picture CM _k of each car in the CM k is normalized to 64×64 size, and grayscaled, and the Hog features hog _j and hog _k are calculated; wherein, in the calculation, the window size used is 64×64, and the block size is 16×16, the block sliding increment size is 8×8, the cell size is 8×8, and the number of gradient histograms in each cell unit is 9;

Step 3-2-4) Calculate the cosine distance Cos _jk according to the Hog features hog _j and hog _k obtained in step 3-2-3);

Step 3-2-5) According to the cosine distance Cos _jk obtained in step 3-2-4), select the largest cosine distance Cos _jm between the tracking vehicle T _j and the vehicle set C and its corresponding vehicle C _m , if Cos _jm >0.75, then T _j and C _m match each other, and go to step 3-3); if Cos _jm ≤0.75, there is no vehicle matching the tracking vehicle T _j in vehicle set C, and go to step 3-4) .

Described step 4) specifically comprises:

Step 4-1) traverse the updated tracking vehicle list TL obtained according to step 3-5), traverse the updated tracking vehicle list TL, and calculate the horizontal directional distance TOd between the initial tracking point of T _j and the lane line L _j ′ _jj and the product of the horizontal directional distance Td′ _jj between the current tracking point and the lane line L _j , if TOd′ _jj *Td′ _jj <0, it means that the vehicle is located on both sides of the lane line L _j one after another, and T _j is an illegal lane change vehicle;

Step 4-2) According to the illegal lane-changing vehicle T _j obtained in step 4-1), the position of the tracking vehicle T _j is marked, and the information of the illegal lane-changing vehicle T _j is deleted from the updated tracking vehicle list TL.

A real-time detection system for marking vehicle illegal lane changes based on lane lines, the detection system is an intelligent traffic management system, including a memory, a processor and a computer program stored on the memory and operable on the processor, characterized in that , implementing the steps of the detection method when the processor executes the program.

The advantages of the present invention are:

The present invention detects the vehicle set in the image based on a deep convolutional neural network, which has higher accuracy and wider application range than the method of extracting the vehicle set by using motion features. In addition, the method of the present invention screens out the vehicles that intersect with the lane line and then performs matching judgment, without recording and judging the trajectory of each vehicle, the detection speed is fast, and the real-time recognition requirements can be met.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are For some embodiments of the present invention, those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a kind of flow chart of the real-time detection method of vehicle illegal lane change based on lane marking of the present invention;

Fig. 2 (a) is a schematic diagram of the effect of setting lane lines in step 1) of a real-time detection method for vehicle violations and lane changes based on lane line marking in an embodiment of the present invention;

Fig. 2 (b) is a schematic diagram of the effect of detecting a vehicle set in step 2) of a real-time detection method for vehicle violations and lane changes based on lane line marking in an embodiment of the present invention;

Fig. 2 (c) is a schematic diagram of the effect of the set of vehicles intersecting with lane lines screened out in step 3) of a real-time detection method for vehicle violations and lane changes based on lane line marking in an embodiment of the present invention;

Fig. 2(d) is a schematic diagram of the effect of marking the collection of illegal lane changing vehicles in the tracking vehicle list in step 5) of a real-time detection method for vehicle lane change lane marking based on lane line marking in an embodiment of the present invention;

Fig. 2 (e) is a schematic diagram of the effect of step 5) the driving away state of illegal vehicles using a real-time detection method for vehicle violations and lane changes based on lane markings in an embodiment of the present invention;

Fig. 3 is a network structure diagram of the vehicle detection model in the embodiment of the present invention.

detailed description

As shown in Figure 1, the present invention provides a real-time detection method for vehicle violations and lane changes based on lane line calibration, which specifically includes:

Step 1) mark the lane line L, calculate and obtain the detection rectangular area D and the image I;

Step 2) In the detection rectangular area D of step 1), the vehicle set A={A ₁ , A ₂ ,...,A _i } of the image I is detected based on the deep convolutional neural network;

Step 3) According to the vehicle set A obtained in step 2), filter out the vehicle set B that intersects the lane line L={B ₁ , B ₂ ,...,B _i }, where, Then match with the tracked vehicle list TL, and update the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j };

Step 4) Determine whether the tracking vehicle T _j in the updated tracking vehicle list TL is an illegal lane-changing vehicle; if the tracking vehicle T _j is an illegal lane-changing vehicle, then mark the position of the tracking vehicle T _j , and use Delete the information of illegal lane-changing vehicles T _j from the tracked vehicle list TL.

Described step 1) specifically comprises:

Step 1-1) Mark the lane line L={L ₁ ,L ₂ ,...,L _j }, the position of each lane line L _j includes: starting point pb _j ={x _bj ,y _bj } and end point pe _j ={x _ej , y _ej }, as shown in Figure 2(a); wherein, x _bj , y _bj are the abscissa and ordinate of the starting point respectively, x _ej , y _ej are the abscissa and ordinate of the end point respectively;

Step 1-2) Calculate the detection area rectangle D={x, y, w, h} according to the position of the lane line in step 1-1), where,

x＝min(x _b1 ,x _e1 ,x _b2 ,x _e2 ,...,x _bj ,x _ej )

y=min(y _b1 ,y _e1 ,y _b2 ,y _e2 ,...,y _bj ,y _ej )

w=max(x _b1 ,x _e1 ,x _b2 ,x _e2 ,...,x _bj ,x _ej )-x

h=max(y _b1 ,y _e1 ,y _b2 ,y _e2 ,...,y _bj ,y _ej )-y

Step 2) specifically includes:

According to the detection area D and image I in step 1), the detection area D in the image I is normalized to 576×576 as the input image, and the vehicle set A in the image I is detected based on the deep convolutional neural network = {A ₁ ,A ₂ ,...,A _i }, the position of A _i is recorded as PA _i ={ _xi ,y _i ,w _i ,h _i }, the detection effect is shown in Figure 2(b);

Among them, as shown in Figure 3, the convolutional neural network includes: 7 convolutional layers, 4 downsampling layers, 2 fully connected layers, and an output layer; the scanning boundaries in the convolutional layers are automatically filled with 0, all using Leaky -The ReLu function activates the neurons; the maximum pooling is used in the downsampling layer.

The convolution kernel size of the convolution layer C1 is 9×9, 32 convolution kernels, the step size is 2, and the size of the generated feature map is 288×288; the window size of the downsampling layer S1 is 4×4, and the step size is 4, The size of the generated feature map is 72×72; the size of the convolutional layer C2 convolution kernel is 1×1, 4 convolution kernels, the step size is 1, and the size of the generated feature map is 72×72; the size of the convolutional layer C3 convolution kernel is 3×3, 8 convolution kernels, the step size is 1, and the size of the generated feature map is 72×72; the size of the downsampling layer S2 window is 2×2, the step size is 2, and the size of the generated feature map is 36×36; The convolution layer C4 convolution kernel size is 1×1, 8 convolution kernels, the step size is 1, and the generated feature map size is 36×36; the convolution layer C5 convolution kernel size is 3×3, 16 convolutions Kernel, the step size is 1, and the size of the generated feature map is 36×36; the window size of the downsampling layer S3 is 2×2, the step size is 2, and the size of the generated feature map is 18×18; the size of the convolution layer C6 convolution kernel is 3×3, 32 convolution kernels, the step size is 1, and the size of the generated feature map is 18×18; the size of the downsampling layer S4 window is 2×2, the step size is 2, and the size of the generated feature map is 9×9; The size of the multilayer C7 convolution kernel is 1×1, 64 convolution kernels, the step size is 1, and the size of the generated feature map is 9×9; the fully connected layer F1 is composed of 256 neurons, and the neurons are processed using the Relu function. Activation; the fully connected layer F2 is composed of 4096 neurons, and the neurons are activated using the Leaky-ReLu function; the output layer is composed of 891 neurons, and the neurons are activated using the Relu function.

Described step 3) specifically comprises:

Step 3-1) Based on the vehicle set A={A ₁ ,A ₂ ,...,A _i } obtained in step 2), select the vehicle set B={B ₁ ,B ₂ , ...,B _i }, where,

Step 3-2) Traverse the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j } according to the vehicle set B obtained in step 3-1), and select the vehicle set C whose coincidence degree U _ij >0.2 , C＝{C ₁ ,C ₂ ,...,C _m }; calculate the Hog features hog _j and hog _k , and then calculate the cosine distance Cos _jk ; then select the largest cosine distance between the tracking vehicle T _j and the vehicle set C Cos _jm and the corresponding vehicle C _m , if Cos _jm > 0.75, then T _j and C _m match each other, and enter step 3-3); if Cos _jm ≤ 0.75, then there is no tracking vehicle T _j in the vehicle set C matching vehicle, and proceed to steps 3-4);

Step 3-3) According to the vehicle C _m obtained in step 3-2), update the vehicle position of the tracking vehicle T _j to the current vehicle position, that is, TP _j =CP _m ; update the tracking vehicle to the current vehicle picture, that is, TM _j =CM _m ; consecutive follow-up times TT _j = 0, the horizontal directional distance TD _j between the tracking point and the lane line L is updated to the horizontal directional distance between the current tracking point and the lane line L, that is, TD _j = CD _m ; at this time C _m Has been matched, whether the current vehicle is matched flag CF _i = true;

Step 3-4) Update the number of consecutive lost times TT _j ' of the tracking vehicle T _j , where TT _j '=TT _j + 1, if the number of consecutive lost times of TT _j ' is greater than 3 times, delete it from the list of tracking vehicles Vehicle T _j ;

Step 3-5) Traversing the vehicle set B, taking the unmatched vehicle B _u ={BP _u ,BM _u ,BF _u ,BQ _u ,BD _u } as the new tracking vehicle T _u ={TP _u ,TM _u , TT _u , TOD _u , TD _u } are added to the tracking vehicle list TL, and the vehicle tracking list TL is updated, wherein,

TP _u = BP _u

TM _u =BM _u

TT _u =0

TOD _u = BD _u

_TDu = _BDu .

The step 3-1) specifically includes:

Step 3-1-1) According to the lane line L _j obtained in step 1) and step 2) and the vehicle A _i position rectangle PA _i , calculate the intersection of the lane line L _j and the four sides of the vehicle A _i position rectangle PA _i , and the lane line The intersection points of L _j and the left edge, right edge, upper edge, and lower edge of the rectangle are respectively pl _ij ={x _lij ,y _lij }, pr _ij ={x _rij ,y _rij }, pu _ij ={x _uij , y _uij }, pd _ij = {x _dij , y _dij };

Among them, {x _lij , y _lij } are the abscissa and ordinate of the left edge intersection respectively, {x _rij ,y _rij } are the abscissa and ordinate of the right edge intersection respectively, {x _uij ,y _uij } are the upper edge respectively The abscissa and ordinate of the intersection point are respectively the abscissa and ordinate of the intersection point of the lower edge.

Step 3-1-2) According to the intersection of the lane line and the rectangle obtained in step 3-1-1), if the intersection is on the top and left of the rectangle and meets the following preset conditions, add the vehicle rectangle A _i to the vehicle set B;

x _i + w _i /4 < x _uij < x _i + w _i

y _i +h _i /2<y _lij <y _i +h _i

Step 3-1-3) According to the intersection point of the lane line and the rectangle obtained in step 3-1-1), if the intersection point is on the upper side and the right side of the rectangle and meets the following preset conditions, add the vehicle rectangle A _i to the vehicle set B;

x _i < x _uij < x _i +3w _i /4

y _i +h _i /2<y _rij <y _i +h _i

Step 3-1-4) According to the intersection point of the lane line and the rectangle obtained in step 3-1-1), if the intersection point is on the upper and lower sides of the rectangle and meets the following preset conditions, add the vehicle rectangle A _i to the vehicle set B;

x _i +w _i /4<x _uij <x _i +3w _i /4

x _i + w _i /4 < x _dij < x _i +3 w _i /4.

The step 3-2) specifically includes:

Step 3-2-1) According to the vehicle set B obtained in step 3-1), initialize the vehicle information, which includes: vehicle position BP _i ={ _xi , y _i , w _i , h _i }, vehicle picture BM _i , Whether the current vehicle is matched with the flag BF _i = false, the tracking point BQ _i = ( _xi + w _i /2, y _i +4h _i /5) and the horizontal directional distance between the tracking point BQ _i and each lane line L _j Bd _ij , that is, vehicle B _i ={BP _i ,BM _i ,BF _i ,BQ _i ,BD _i }, where BD _i ={Bd _i1 ,Bd _i2 ,...,Bd _il };

Step 3-2-2) Traverse the tracked vehicle list TL={T ₁ , T ₂ ,...,T _j }, T _j ={TP _j ,TM _j ,TT _j ,TOD _j ,TD _j }, T _j ∈TL, where TP _j represents the vehicle position, TM _j represents the vehicle picture, TT _j represents the number of consecutive follow-ups, TOD _j represents the horizontal distance between the initial tracking point and the lane line L, TD _j represents the current tracking point and the lane Horizontal directional distance of line L; TOD _j = {TOd _j1 , TOd _j2 ,..., TOd _jl }, TD _j = {Td _j1 , Td _j2 ,..., Td _jl }; calculate the tracking vehicle T _j The coincidence degree U _ij of the vehicle position TP _j and each unmatched vehicle position BP _i in the vehicle set B:

U _ij ＝S _cij /S _zij

Among them, S _zij is the union of the rectangular areas of BP _i and TP _j , and S _cij is the intersection of the rectangular areas of BP _i and TP _j ;

Select the vehicle set C whose coincidence degree U _ij >0.2, C={C ₁ ,C ₂ ,...,C _m }; if C is empty, there is no vehicle matching the tracking vehicle T _j in the vehicle set B, and enter step 3-4); if C is not If it is empty, go to step 3-2-3);

Step 3-2-3) According to the vehicle set C={C ₁ ,C ₂ ,...,C _m } obtained in step 3-2-2), the vehicle picture TM _j of the tracking vehicle T _j and the vehicle set C The picture CM _k of each car in the CM k is normalized to 64×64 size, and grayscaled, and the Hog features hog _j and hog _k are calculated; wherein, in the calculation, the window size used is 64×64, and the block size is 16×16, the block sliding increment size is 8×8, the cell size is 8×8, and the number of gradient histograms in each cell unit is 9;

Step 3-2-4) Calculate the cosine distance Cos _jk according to the Hog features hog _j and hog _k obtained in step 3-2-3);

Step 3-2-5) According to the cosine distance Cos _jk obtained in step 3-2-4), select the largest cosine distance Cos _jm between the tracking vehicle T _j and the vehicle set C and its corresponding vehicle C _m , if Cos _jm >0.75, then T _j and C _m match each other, and go to step 3-3), the vehicle detection effect is shown in Figure 2(c) if the matching is successful; if Cos _jm ≤0.75, there is no tracking vehicle in vehicle set C T _j matches the vehicle, and enters step 3-4), and the detection effect of the vehicle that fails to match is shown in Figure 2(d).

Described step 4) specifically comprises:

Step 4-1) Traverse the updated tracking vehicle list TL, calculate the horizontal directional distance TOd′ _jj between the starting tracking point of T _j and the lane line L _j and the horizontal directional distance Td between the current tracking point and the lane line L _j The product of ′ _jj , if TOd′ _jj *Td′ _jj <0, means that the vehicles are located on both sides of the lane line L _j one after another, and T _j is an illegal lane-changing vehicle;

Step 4-2) According to the illegal lane-changing vehicle T _j obtained in step 4-1), mark the position of the tracking vehicle T _j , as shown in Figure 2(e), and delete it from the updated tracking vehicle list TL The information of the illegal lane-changing vehicle T _j .

2(a)-2(e) are schematic diagrams showing the effect of using the method in the embodiment of the present invention to detect illegal lane-changing vehicles.

A real-time detection system for marking vehicle illegal lane changes based on lane lines, the detection system is an intelligent traffic management system, including a memory, a processor and a computer program stored on the memory and operable on the processor, characterized in that , implementing the steps of the detection method when the processor executes the program.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than limit them. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent replacements to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all of them should be included in the scope of the present invention. within the scope of the claims.

Claims (8)

1. a kind of real-time detection method based on lane line demarcation vehicle peccancy lane change, is specifically included：

Step 1) demarcates lane line L, calculates and obtains detection rectangular area D and image I；

Step 2) detects image I vehicle set based on deep layer convolutional neural networks in the detection rectangular area D of step 1) A={ A₁,A₂,...,A_i}；

The vehicle set A that step 3) obtains according to step 2), filter out the vehicle set B={ B intersected with lane line L₁, B₂,...,B_i, wherein,Matched again with tracking vehicle list TL, and update tracking vehicle list TL={ T₁, T₂,...,T_j}；

Step 4) judges the tracking vehicle T in the tracking vehicle list TL after renewal_jWhether it is lane change vehicle violating the regulations；If tracking Vehicle T_jFor lane change vehicle violating the regulations, then tracking vehicle T is marked out_jPosition, and deleted from the tracking vehicle list TL after renewal Lane change vehicle T violating the regulations_jInformation.

2. detection method according to claim 1, it is characterised in that the step 1) specifically includes：

Step 1-1) demarcation lane line L={ L₁,L₂,...,L_j, every lane line L_jPosition include：Starting point pb_j={ x_bj, y_bjAnd terminal pe_j={ x_ej,y_ej}；Wherein, x_bj,y_bjThe respectively horizontal stroke of starting point, ordinate, x_ej,y_ejRespectively the horizontal stroke of terminal, Ordinate；

Step 1-2) according to step 1-1) track line position, calculate detection zone rectangle D={ x, y, w, h }, wherein,

X=min (x_b1,x_e1,x_b2,x_e2,...,x_bj,x_ej)

Y=min (y_b1,y_e1,y_b2,y_e2,...,y_bj,y_ej)

W=max (x_b1,x_e1,x_b2,x_e2,...,x_bj,x_ej)-x

H=max (y_b1,y_e1,y_b2,y_e2,...,y_bj,y_ej)-y。

3. detection method according to claim 1, it is characterised in that step 2) specifically includes：

According to the detection zone D and image I of step 1), the detection zone D in image I is normalized to 576 × 576 as defeated Enter image, the vehicle set A={ A in image I are detected based on deep layer convolutional neural networks₁,A₂,...,A_i, A_iPosition note For PA_i={ x_i,y_i,w_i,h_i}；

Wherein, the convolutional neural networks include：7 convolutional layers, 4 down-sampling layers, 2 full articulamentums and an output layer；Volume Scanning boundary filling 0 automatically, enters line activating using Leaky-ReLu function pair neurons in lamination；Used in down-sampling layer Maximum pond；

Convolutional layer C1 convolution kernel size is 9 × 9,32 convolution kernels, and step-length 2, generation characteristic pattern size is 288 × 288；Under Sample level S1 window sizes are 4 × 4, step-length 4, and generation characteristic pattern size is 72 × 72；

Convolutional layer C2 convolution kernels size is 1 × Isosorbide-5-Nitrae convolution kernel, and step-length 1, generation characteristic pattern size is 72 × 72；

Convolutional layer C3 convolution kernels size is 3 × 3,8 convolution kernels, and step-length 1, generation characteristic pattern size is 72 × 72；Down-sampling Layer S2 window sizes are 2 × 2, step-length 2, and generation characteristic pattern size is 36 × 36；

Convolutional layer C4 convolution kernels size is 1 × 1,8 convolution kernels, and step-length 1, generation characteristic pattern size is 36 × 36；

Convolutional layer C5 convolution kernels size is 3 × 3,16 convolution kernels, and step-length 1, generation characteristic pattern size is 36 × 36；Down-sampling Layer S3 window sizes are 2 × 2, step-length 2, and generation characteristic pattern size is 18 × 18；

Convolutional layer C6 convolution kernels size is 3 × 3,32 convolution kernels, and step-length 1, generation characteristic pattern size is 18 × 18；Down-sampling Layer S4 window sizes are 2 × 2, step-length 2, and generation characteristic pattern size is 9 × 9；

Convolutional layer C7 convolution kernels size is 1 × 1,64 convolution kernels, and step-length 1, generation characteristic pattern size is 9 × 9；

Full articulamentum F1 is made up of 256 neurons, enters line activating using Relu function pair neurons；Full articulamentum F2 is by 4096 Individual neuron is formed, and enters line activating using Leaky-ReLu function pair neurons；Output layer is made up of 891 neurons, is used Relu function pair neurons enter line activating.

4. detection method according to claim 1, it is characterised in that the step 3) specifically includes：

Step 3-1) the vehicle set A={ A that are obtained according to step 2)₁,A₂,...,A_i, therefrom filter out and intersect with lane line L Vehicle set B={ B₁,B₂,...,B_i, wherein,

Step 3-2) according to step 3-1) obtained vehicle set B, traversal tracking vehicle list TL={ T₁,T₂,...,T_j, choosing Go out registration U_ij＞ 0.2 vehicle set C,C={ C₁,C₂,...,C_m}；Calculate Hog features hog_jAnd hog_k, then count Calculate its COS distance Cos_jk；Tracking vehicle T is selected again_jWith COS distance Cos maximum in vehicle set C_jmAnd corresponding vehicle C_mIf Cos_jm＞ 0.75, then T_jAnd C_mIt is mutually matched, and enters step 3-3)；If Cos_jm≤ 0.75, then in vehicle set C not In the presence of and tracking vehicle T_jThe vehicle to match, and enter step 3-4)；

Step 3-3) according to step 3-2) obtained vehicle C_m, renewal tracking vehicle T_jVehicle location be current vehicle location, i.e., TP_j=CP_m；Renewal tracking vehicle is Current vehicle picture, i.e. TM_j=CM_m；Continuously with losing number TT_j=0, trace point and track Line L horizontal directed distance TD_jIt is updated to the horizontal directed distance of current tracking point and lane line L, i.e. TD_j=CD_m；Now C_m It has been be matched that, whether Current vehicle is matched mark CF_i=true；

Step 3-4) renewal tracking vehicle T_jIt is continuous with losing number TT_j', wherein, TT_j'=TT_j+ 1, if TT_j' it is continuous with losing Number is more than 3 times, then deletes vehicle T from tracking vehicle list_j；

Step 3-5) traversal vehicle set B, the vehicle B that will be matched_u={ BP_u,BM_u,BF_u,BQ_u,BD_uAs new tracking Vehicle T_u={ TP_u,TM_u,TT_u,TOD_u,TD_uBe added in tracking vehicle list TL, vehicle tracking list TL is updated, wherein,

TP_u=BP_u

TM_u=BM_u

TT_u=0

TOD_u=BD_u

TD_u=BD_u。

5. detection method according to claim 4, it is characterised in that the step 3-1) specifically include：

Step 3-1-1) the lane line L and vehicle A that are obtained according to step 1) and step 2)_iPosition rectangle PA_i, calculate lane line L_j With vehicle A_iPosition rectangle PA_iThe intersection point on four sides, and lane line L_jWith rectangle left edge, right side edge, upper edge, downside The intersection point at edge is respectively pl_ij={ x_lij,y_lij}、pr_ij={ x_rij,y_rij}、pu_ij={ x_uij,y_uij}、pd_ij={ x_dij,y_dij}；

Wherein, { x_lij,y_lijIt is respectively that left side edge intersection point is horizontal, ordinate, { x_rij,y_rijBe respectively right side edge intersection point it is horizontal, Ordinate, { x_uij,y_uijIt is respectively upper edge intersection point horizontal stroke, ordinate, respectively lower edge intersection point horizontal stroke, ordinate；

Step 3-1-2) according to step 3-1-1) intersection point of obtained lane line and rectangle, if intersection point on rectangle top and the left side and Meet following preparatory condition, by vehicle rectangle A_iIt is added in vehicle set B；

x_i+w_i/ 4 ＜ x_uij＜ x_i+w_i

y_i+h_i/ 2 ＜ y_lij＜ y_i+h_i；

Step 3-1-3) according to step 3-1-1) intersection point of obtained lane line and rectangle, if intersection point on rectangle top and the right and Meet following preparatory condition, by vehicle rectangle A_iIt is added in vehicle set B；

x_i＜ x_uij＜ x_i+3w_i/4

y_i+h_i/ 2 ＜ y_rij＜ y_i+h_i；

Step 3-1-4) according to step 3-1-1) intersection point of obtained lane line and rectangle, if intersection point in rectangle bottom and upper segment and Meet following preparatory condition, by vehicle rectangle A_iIt is added in vehicle set B；

x_i+w_i/ 4 ＜ x_uij＜ x_i+3w_i/4

x_i+w_i/ 4 ＜ x_dij＜ x_i+3w_i/4。

6. detection method according to claim 4, it is characterised in that the step 3-2) specifically include：

Step 3-2-1) according to step 3-1) obtained vehicle set B, information of vehicles is initialized, including：Vehicle location BP_i ={ x_i,y_i,w_i,h_i, vehicle pictures BM_i, Current vehicle whether be matched mark BF_i=false, trace point BQ_i=(x_i+w_i/ 2,y_i+4h_i/ 5) and trace point BQ_iWith every lane line L_jHorizontal directed distance Bd_ij, that is, track vehicle B_i={ BP_i,BM_i, BF_i,BQ_i,BD_i, wherein, BD_i={ Bd_i1,Bd_i2,...,Bd_il}；

Step 3-2-2) traversal tracking vehicle list TL={ T₁,T₂,...,T_j, T_j={ TP_j,TM_j,TT_j,TOD_j,TD_j, T_j∈ TL, wherein, TP_jRepresent vehicle location, TM_jRepresent vehicle pictures, TT_jRepresent continuously with losing number, TOD_jRepresent starting trace point With lane line L horizontal directed distance, TD_jRepresent the horizontal directed distance of current tracking point and lane line L；TOD_j={ TOd_j1, TOd_j2,...,TOd_jl, TD_j={ Td_j1,Td_j2,...,Td_jl}；Calculate tracking vehicle T_jVehicle location TP_jWith vehicle set B In each truck position BP for not being matched_iRegistration U_ij：

U_ij=S_cij/S_zij

Wherein, S_zijFor BP_iAnd TP_jThe union of rectangular area, S_cijFor BP_iAnd TP_jThe common factor of rectangular area；

Select registration U_ij＞ 0.2 vehicle set C,C={ C₁,C₂,...,C_m}；If C is sky, in vehicle set B In the absence of with tracking vehicle T_jThe vehicle to match, and enter step 3-4)；If C is not sky, into step 3-2-3)；

Step 3-2-3) according to step 3-2-2) obtained vehicle set C={ C₁,C₂,...,C_m, vehicle T will be tracked_jVehicle Picture TM_jWith the picture CM of each car in vehicle set C_k64 × 64 sizes are normalized to, and carry out gray processing, calculate Hog features hog_jAnd hog_k；Wherein, in the calculating, the window size used is 64 × 64, and block size is 16 × 16, and it is big that block slides increment Small is 8 × 8, and cell element size is 8 × 8, and the quantity of histogram of gradients is 9 in each born of the same parents' unit；

Step 3-2-4) according to step 3-2-3) obtained Hog features hog_jAnd hog_k, calculate its COS distance Cos_jk；

Step 3-2-5) according to step 3-2-4) obtained COS distance Cos_jk, select tracking vehicle T_jWith in vehicle set C most Big COS distance Cos_jmAnd its corresponding vehicle C_mIf Cos_jm＞ 0.75, then T_jAnd C_mIt is mutually matched, and enters step 3- 3)；If Cos_jm≤ 0.75, then it is not present and tracks vehicle T in vehicle set C_jThe vehicle to match, and enter step 3-4).

7. detection method according to claim 1, it is characterised in that the step 4) specifically includes：

Step 4-1) the tracking vehicle list TL after renewal is traveled through, calculate T_jStarting trace point and lane line L_jIt is horizontal oriented Distance TOd '_jjWith current tracking point and lane line L_jHorizontal directed distance Td '_jjProduct, if TOd '_jj*Td′_jj＜ 0, represent Vehicle is successively located at lane line L_jBoth sides, T_jFor lane change vehicle of breaking rules and regulations；

Step 4-2) according to step 4-1) obtained lane change vehicle T violating the regulations_j, then tracking vehicle T is marked out_jPosition, and from renewal Lane change vehicle T violating the regulations is deleted in tracking vehicle list TL afterwards_jInformation.

8. a kind of real-time detecting system based on lane line demarcation vehicle peccancy lane change, the detecting system is intelligent traffic administration system System, including memory, processor and storage on a memory and the computer program that can run on a processor, its feature It is, the step of claim 1~7 detection method is realized during the computing device described program.