CN106023270A - Video vehicle detection method based on locally symmetric features - Google Patents

Abstract

The invention discloses a video vehicle detection method based on locally symmetric features. The method comprises the step of describing a vehicle target through the adoption of the set of locally unchanged features so as to effectively avoid partition problem; compared with the prior art, the symmetric features of the vehicle are used as important clues about local feature clustering, the positioning of the vehicle central position is realized, the problem of high algorithm complexity produced through the adoption of the conventional clustering algorithm is effectively avoided; and the method is high in detection precision, simple in operation process, strong in real time, and capable of effectively detecting and identifying the vehicle target in a static picture and the video image; therefore, the method has wide application prospect.

Description

Video vehicle detection method based on local symmetric characteristics

Technical Field

The invention belongs to the technical field of video detection, and particularly relates to a video vehicle detection method based on local symmetric characteristics.

Background

The rapid development of economy and the rapid increase of the number of vehicles lead the developed countries and the developing countries to be fully troubled by traffic problems, how to acquire traffic information, effectively dredge traffic and relieve traffic jam, so that the occurrence of traffic accidents is more and more important, and the video-based vehicle detection technology is used as important research content in the field of computer vision and image processing, and is more and more concerned by people due to the advantages of convenience and quickness.

In order to detect and identify a vehicle target, the vehicle target usually needs to be extracted, and a commonly used vehicle target extraction method is to separate the target from a traffic background by using the motion characteristics of a vehicle so as to detect the vehicle target, however, the method is easily affected by the change of external conditions such as light, weather, vehicle shielding and the like, and it is very difficult to separate the motion target from a complex background; another method is to use low-level features (such as edge features, color features, shape features, etc.) of the vehicle to perform segmentation, but this method is difficult to obtain ideal segmentation effect, and the above two methods mainly use global characteristics of the vehicle, inevitably segment suspected vehicle targets, and the accuracy of segmentation will directly affect the result of vehicle recognition. However, in an actual traffic scene, moving vehicles are easily affected by factors such as weather and illumination changes, vehicle congestion, occlusion, and shadows, and thus the target vehicle is more difficult to segment.

Disclosure of Invention

Aiming at the defects and shortcomings of the prior art, the invention aims to provide a video vehicle detection method based on local symmetric features.

In order to achieve the above object, the present invention adopts the following technical solution, a video vehicle detection method based on local symmetric features, comprising the following steps:

manually setting an ROI (region of interest) along a lane line on an initial frame, and recording coordinates of each pixel point on a boundary line of the ROI;

step two, for the interior of the ROI, calculating a characteristic corner point p of a video image in the current frame_iAnd recording the characteristic corner point p_iThe position coordinates of (a);

step (ii) ofThirdly, taking each characteristic angular point obtained in the second step as a center, respectively constructing a square area, and adopting a characteristic description operator to construct a characteristic vector of each characteristic angular point

Fourthly, constructing a horizontal symmetrical angular point q of each characteristic angular point_iFeature vector of

Step five, calculating any characteristic angular point p_iHorizontal symmetrical angular points q corresponding to all other characteristic angular points respectively_iIs given by the minimum distance denoted Min (p)_i) The next smallest distance is denoted Min^Sec(p_i)，Min(p_i) The calculation formula of (a) is as follows:

$Min\left(p_i\right)=Min\left(D\right(p_i,q_j\left)\right)=\underset{i=1}{\overset{M}{\Σ}}\underset{j\≠i}{\overset{M}{\Σ}}{(V_p^i-V_q^j)}^2---\left(1.1\right)$

wherein,a feature vector representing the ith feature corner,representing the jth horizontally symmetric corner point q_jCharacteristic vector of (q)_jAs a characteristic corner point p_jA horizontal symmetry corner point;

step six, if the characteristic angular point p_iThe characteristic corner point p is satisfied_iAnd a characteristic corner point p_jFor symmetrical corner pairs, characteristic corner p_jAs a characteristic corner point p_iHorizontal symmetrical corner point q with minimum distance_jThe corresponding characteristic angular point;

$\frac{Min\left(p_i\right)}{{\mathrm{Min}}^{Sec}\left(p_i\right)}<0.65---\left(1.2\right)$

step seven, traversing all the characteristic angular points p_iRepeating the fifth step and the sixth step until all the symmetrical angle point pairs are found;

step eight, assume useRepresenting each symmetric angle pair, respectively obtaining the x coordinate of each symmetric angle pair center position asCalculating to obtain the central positionThe peak point of the statistical histogram is used as the initial position x of the candidate vehicle center line_vehicleAnd calculating the variance of the statistical histogram

Step nine, judging symmetrical angle point pairsWhether the angle points belong to symmetrical angle point pairs on the same vehicle or not, if the angle point pairs belong to symmetrical angle point pairsIf the x coordinate of the central position of the point pair satisfies the formula (1.3), the symmetrical angle point pair is reserved, otherwise, the point pair is deleted;

$|C_{p\&LeftRightArrow;q}^i.x-x_{vehicle}|<3{\mathrm{\&sigma;}}_{vehicle}^2---\left(1.3\right)$

step ten, counting the average value mu of all the symmetrical corner points retained in the step nine to the central position_vehicle，μ_vehicleThe center line position of the candidate vehicle is obtained;

eleven, selecting a plurality of frames in the current video image, manually selecting a shadow area at the bottom of the vehicle, modeling by using the geometric characteristics, brightness and color information of the shadow, and training the mean value mu of the shadow sample_shadowSum variance σ_shadowThe number l of x-direction pixel points and the number h of y-direction pixel points in the shadow area;

step twelve, testing the suspected vehicle area pixel points on the two sides of the central line by using a Gaussian mixture model, wherein the suspected vehicle area is x ∈ [ mu ]_vehicle-l，μ_vehicle+l]As shown in formula (1.4), wherein p is_iRepresenting the measured pixel point, T, in the image_shadowG for shadow sample set_shadow(p_i) The mean of the function; if the measured pixel point satisfies the formula (1.5), the pixel point is judged to be a shadow point, and when the shadow point is continuous and satisfies the | N |_x-l | < 0.2 x l and | N_y-h | < 0.1 × h, N_x、N_yIf the number of the continuous shadow points in the x direction and the y direction is respectively, the central line corresponding to the shadow area can be judged to be the central line of the vehicle, otherwise, the central line is not the central line of the vehicle, and the detection is finished;

$G_{shadow}\left(p_i\right)=\exp (-\frac{{(p_i-{\mathrm{\&mu;}}_{shadow})}^2}{{\mathrm{\&sigma;}}_{shadow}^2})---\left(1.4\right)$

G_shadow(p_i)＞T_shadow(1.5)

step thirteen, assume that the vehicle target region may be represented as R ═ l_R,r_R,u_R,b_R)，l_RAnd r_RIs respectively measured in mu_vehicleLeft and right boundary values of the centered target region, u_RAnd b_RBoundary values representing the upper and lower sides of the target region, determining l_R、r_R、u_RAnd b_RThe value taking method comprises the following steps:

b_Rthe calculating method of (2): counting the number of pixel values of shadow areas at the bottom of the vehicle line by line, and taking the maximum value as b_RWidth of (2) with b_RBy width of (2) as a limit, using a droopA straight Sobel operator for extracting the edge of the image in the ROI above the shadow region at the bottom of the vehicle, and the central line between the rear bumper of the vehicle and the shadow region at the bottom of the vehicle is determined as b_RThe position of (a);

r_Rand l_RThe calculating method of (2): r is_RAnd l_RIs determined according to the position of the center line of the vehicle and b_RIs determined as shown in formula (1.6);

l_R＝μ_vehicle-width(b_R)/2r_R＝μ_vehicle+width(b_R)/2 (1.6)

u_Rthe calculating method of (2): assuming that the vehicle height is h, the height and width satisfy the proportional relationship h ═ γ (r)_R-l_R) Wherein γ is obtained from a set of vehicle data training; image [ l ] extraction using vertical Sobel operator_R,r_R]Horizontal edges within the range and counting the horizontal projection of the gray level image, assuming that P (y) represents the distribution of the horizontal projection histogram, then y ∈ [ b_R+0.5h,b_R+1.5h]The maximum value of P (y) in the range is u_RA value of (d);

by applying the above method, we can obtain b sequentially_R、l_R、r_RAnd u_RAnd taking values, thereby determining the target vehicle.

Further, the method for extracting the feature corner in the second step is a Harris corner extraction algorithm.

Further, the feature vector V of each feature corner point in the third step_iThe construction method comprises the steps of utilizing a Haar wavelet template to calculate response values of a square area image, respectively obtaining Haar wavelet response values along the X direction and the Y direction, counting response values of ∑ dx, ∑ dy, ∑ | dx | and ∑ | dy |, and obtaining the response value of each characteristic angular point p_iFeature vector of

Further, the square area size constructed in the third step is 5 x 5.

Further, the size of the Haar wavelet template in the third step is 2 x 2.

Further, the horizontal symmetrical corner point q of each characteristic corner point in the fourth step_iFeature vector ofThe construction method comprises the following steps:

compared with the prior art, the method has the advantages that the vehicle target is described by using the local invariant feature set, so that the segmentation problem can be effectively avoided.

Drawings

Fig. 1 is a view for manually setting the ROI along the lane line.

Fig. 2 is a schematic structural diagram of an image region and a Haar wavelet template for constructing a Harris corner feature vector, where (a) is a 5 × 5 image region centered on a Harris corner, and (b) and (c) are Haar wavelet response templates in the X direction and the Y direction, respectively.

Fig. 3 is a schematic diagram of parameters of a vehicle target area, where (a) is an original image of a vehicle, and (b) is a horizontal edge extraction and horizontal projection image of the vehicle.

Fig. 4 shows a result of detection of a vehicle target.

Detailed Description

The invention is further described with reference to the following figures and detailed description.

A video vehicle detection method based on local symmetric features comprises the following steps:

step one, as shown in fig. 1, manually setting an ROI, i.e., a square frame (red) in fig. 1, along a lane line on an initial frame, and recording coordinates of each pixel point on a boundary line of the ROI;

step two, for the interior of the ROI, calculating a characteristic corner p of a video image in the current frame by adopting a Harris corner extraction algorithm_iAnd recording the characteristic corner point p_iThe position coordinates of (a);

step three, as shown in fig. 2, with each characteristic corner point obtained in step two as the center, each square region (fig. 2a) with the size of 5 × 5 is constructed, a Haar wavelet template is used to calculate the response value of the image of the square region, wherein the size of the Haar wavelet template is 2 × 2, Haar wavelet response values along the X direction (fig. 2b) and the Y direction (fig. 2c) are obtained, and the response values of ∑ dx, ∑ dy, ∑ | dx | and ∑ | dy | are counted, so as to obtain the response value of each characteristic corner point p_iFeature vector ofFeature vector for constructing each feature angular point by adopting feature description operator

Fourthly, constructing a horizontal symmetrical angular point q of each characteristic angular point_iFeature vector of

Step five, calculating any characteristic angular point p_iHorizontal symmetrical angular points q corresponding to all other characteristic angular points respectively_iIs given by the minimum distance denoted Min (p)_i) The next smallest distance is denoted Min^Sec(p_i)，Min(p_i) The calculation formula of (a) is as follows:

$Min\left(p_i\right)=Min\left(D\right(p_i,q_j\left)\right)=\underset{i=1}{\overset{M}{\&Sigma;}}\underset{j\&NotEqual;i}{\overset{M}{\&Sigma;}}{(V_p^i-V_q^j)}^2---\left(1.1\right)$

wherein,a feature vector representing the ith feature corner,representing the jth horizontally symmetric corner point q_jCharacteristic vector of (q)_jAs a characteristic corner point p_jA horizontal symmetry corner point;

step six, if the characteristic angular point p_iThe characteristic corner point p is satisfied_iAnd a characteristic corner point p_jFor symmetrical corner pairs, characteristic corner p_jAs a characteristic corner point p_iHorizontal symmetrical corner point q with minimum distance_jThe corresponding characteristic angular point;

$\frac{Min\left(p_i\right)}{{\mathrm{Min}}^{Sec}\left(p_i\right)}<0.65---\left(1.2\right)$

step seven, traversing all the characteristic angular points p_iRepeating the fifth step and the sixth step until all the symmetrical angle point pairs are found;

step eight, assume useRepresenting each symmetric angle pair, respectively obtaining the x coordinate of each symmetric angle pair center position asCalculating to obtain the central positionThe peak point of the statistical histogram is used as the initial position x of the candidate vehicle center line_vehicleAnd calculating the variance of the statistical histogram

Step nine, judging symmetrical angle point pairsWhether the angle points belong to symmetrical angle point pairs on the same vehicle or not, if the angle point pairs belong to symmetrical angle point pairsIf the x coordinate of the central position of the point pair satisfies the formula (1.3), the symmetrical angle point pair is reserved, otherwise, the point pair is deleted;

$|C_{p\&LeftRightArrow;q}^i.x-x_{vehicle}|<3{\mathrm{\&sigma;}}_{vehicle}^2---\left(1.3\right)$

step ten, counting the average value mu of all the symmetrical corner points retained in the step nine to the central position_vehicle，μ_vehicleThe center line position of the candidate vehicle is obtained;

eleven, selecting a plurality of frames in the current video image, manually selecting a shadow area at the bottom of the vehicle, modeling by using the geometric characteristics, brightness and color information of the shadow, and training the mean value mu of the shadow sample_shadowSum variance σ_shadowThe number l of x-direction pixel points and the number h of y-direction pixel points in the shadow area;

step twelve, testing the suspected vehicle area pixel points on the two sides of the central line by using a Gaussian mixture model, wherein the suspected vehicle area is x ∈ [ mu ]_vehicle-l，μ_vehicle+l]As shown in formula (1.4), wherein p is_iRepresenting the measured pixel point, T, in the image_shadowG for shadow sample set_shadow(p_i) The mean of the function; if the measured pixel point satisfies the formula (1.5), the pixel point is judged to be a shadow point, and when the shadow point is continuous and satisfies the | N |_x-l | < 0.2 x l and | N_y-h | < 0.1 × h, N_x、N_yIf the number of the continuous shadow points in the x direction and the y direction is respectively, the central line corresponding to the shadow area can be judged to be the central line of the vehicle, otherwise, the central line is not the central line of the vehicle, and the detection is finished;

$G_{shadow}\left(p_i\right)=\exp (-\frac{{(p_i-{\mathrm{\&mu;}}_{shadow})}^2}{{\mathrm{\&sigma;}}_{shadow}^2})---\left(1.4\right)$

G_shadow(p_i)＞T_shadow(1.5)

step thirteen, as shown in fig. 3, assume that the vehicle target region can be expressed as R ═ (l)_R,r_R,u_R,b_R)，l_RAnd r_RIs respectively measured in mu_vehicleLeft and right boundary values of the centered target region, u_RAnd b_RBoundary values representing the upper and lower sides of the target region, determining l_R、r_R、u_RAnd b_RThe value taking method comprises the following steps:

b_Rthe calculating method of (2): counting the number of pixel values of shadow areas at the bottom of the vehicle line by line, and taking the maximum value as b_RWidth of (2) with b_RIs used as a boundary, the edge of the image in the ROI area above the shadow area at the bottom of the vehicle is extracted by using a vertical Sobel operator, and the central line between the rear bumper of the vehicle and the shadow area at the bottom of the vehicle is determined as b_RThe position of (a);

r_Rand l_RThe calculating method of (2): r is_RAnd l_RIs determined according to the position of the center line of the vehicle and b_RIs determined as shown in formula (1.6);

l_R＝μ_vehicle-width(b_R)/2r_R＝μ_vehicle+width(b_R)/2 (1.6)

u_Rthe calculating method of (2): as shown in fig. 3, assuming that the vehicle height is h, the height and width satisfy the proportional relationship h ═ γ (r)_R-l_R) Wherein γ is obtained from a set of vehicle data training; image [ l ] extraction using vertical Sobel operator_R,r_R]Horizontal edges within the range (as shown in fig. 3 b) and counting the horizontal projection of its gray scale image, assuming that p (y) represents its horizontal projection histogram distribution, then y ∈ [ b_R+0.5h,b_R+1.5h]The maximum value of P (y) in the range is u_RA value of (d);

by applying the above method, we can obtain b in turn, as shown in FIG. 4_R、l_R、r_RAnd u_RAnd taking values, thereby determining the target vehicle.

Claims (6)

1. A video vehicle detection method based on local symmetric features is characterized by comprising the following steps:

manually setting an ROI (region of interest) along a lane line on an initial frame, and recording coordinates of each pixel point on a boundary line of the ROI;

step two, for the interior of the ROI, calculating a characteristic corner point p of a video image in the current frame_iAnd recording the characteristic corner point p_iThe position coordinates of (a);

step three, taking each characteristic angular point obtained in the step two as a center, and constructing a first structureConstructing feature vectors of feature angular points by using feature description operators in each square region

Fourthly, constructing a horizontal symmetrical angular point q of each characteristic angular point_iFeature vector of

Step five, calculating any characteristic angular point p_iHorizontal symmetrical angular points q corresponding to all other characteristic angular points respectively_iIs given by the minimum distance denoted Min (p)_i) The next smallest distance is denoted Min^Sec(p_i)，Min(p_i) The calculation formula of (a) is as follows:

$Min\left(p_i\right)=Min\left(D\right(p_i,q_j\left)\right)=\underset{i=1}{\overset{M}{\&Sigma;}}\underset{j\&NotEqual;i}{\overset{M}{\&Sigma;}}{(V_p^i-V_q^j)}^2---\left(1.1\right)$

wherein,a feature vector representing the ith feature corner,representing the jth horizontally symmetric corner point q_jCharacteristic vector of (q)_jAs a characteristic corner point p_jA horizontal symmetry corner point;

step six, if the characteristic angular point p_iThe characteristic corner point p is satisfied_iAnd a characteristic corner point p_jFor symmetrical corner pairs, characteristic corner p_jAs a characteristic corner point p_iHorizontal symmetrical corner point q with minimum distance_jThe corresponding characteristic angular point;

$\frac{Min\left(p_i\right)}{{\mathrm{Min}}^{Sec}\left(p_i\right)}<0.65---\left(1.2\right)$

step seven, traversing all the characteristic angular points p_iRepeating the fifth step and the sixth step until all the symmetrical angle point pairs are found;

step eight, assume useRepresenting each symmetric angle pair, respectively obtaining the x coordinate of each symmetric angle pair center position asCalculating to obtain the central positionThe peak point of the statistical histogram is used as the initial position x of the candidate vehicle center line_vehicleAnd calculating the variance of the statistical histogram

Step nine, judging symmetrical angle point pairsWhether the angle points belong to symmetrical angle point pairs on the same vehicle or not, if the angle point pairs belong to symmetrical angle point pairsIf the x coordinate of the central position of the point pair satisfies the formula (1.3), the symmetrical angle point pair is reserved, otherwise, the point pair is deleted;

$|C_{p\&LeftRightArrow;q}^i.x-x_{vehicle}|<3{\mathrm{\&sigma;}}_{vehicle}^2---\left(1.3\right)$

step ten, counting the average value mu of all the symmetrical corner points retained in the step nine to the central position_vehicle，μ_vehicleThe center line position of the candidate vehicle is obtained;

eleven, selecting a plurality of frames in the current video image, manually selecting a shadow area at the bottom of the vehicle, modeling by using the geometric characteristics, brightness and color information of the shadow, and training the mean value mu of the shadow sample_shadowSum variance σ_shadowThe number l of x-direction pixel points and the number h of y-direction pixel points in the shadow area;

step twelve, testing the suspected vehicle area pixel points on the two sides of the central line by using a Gaussian mixture model, wherein the suspected vehicle area is x ∈ [ mu ]_vehicle-l，μ_vehicle+l]As shown in formula (1.4), wherein p is_iRepresenting the measured pixel point, T, in the image_shadowG for shadow sample set_shadow(p_i) The mean of the function; if the measured pixel point satisfies the formula (1.5), the pixel point is judged to be a shadow point, and when the shadow point is continuous and satisfies the | N |_x-l | < 0.2 x l and | N_y-h | < 0.1 × h, N_x、N_yIf the number of the continuous shadow points in the x direction and the y direction is respectively, the central line corresponding to the shadow area can be judged to be the central line of the vehicle, otherwise, the central line is not the central line of the vehicle, and the detection is finished;

$G_{shadow}\left(p_i\right)=\exp (-\frac{p_i-{\mathrm{\&mu;}}_{shadow}{)}^2}{{\mathrm{\&sigma;}}_{shadow}^2})---\left(1.4\right)$

G_shadow(p_i)＞T_shadow(1.5)

step thirteen, assume that the vehicle target region may be represented as R ═ l_R,r_R,u_R,b_R)，l_RAnd r_RIs respectively measured in mu_vehicleLeft and right boundary values of the centered target region, u_RAnd b_RBoundary values representing the upper and lower sides of the target region, determining l_R、r_R、u_RAnd b_RThe value taking method comprises the following steps:

b_Rthe calculating method of (2): counting the number of pixel values of shadow areas at the bottom of the vehicle line by line, and taking the maximum value as b_RWidth of (2) with b_RIs used as a boundary, the edge of the image in the ROI area above the shadow area at the bottom of the vehicle is extracted by using a vertical Sobel operator, and the central line between the rear bumper of the vehicle and the shadow area at the bottom of the vehicle is determined as b_RThe position of (a);

r_Rand l_RThe calculating method of (2): r is_RAnd l_RIs determined according to the position of the center line of the vehicle and b_RIs determined as shown in equation (1.6)；

l_R＝μ_vehicle-width(b_R)/2 r_R＝μ_vehicle+width(b_R)/2 (1.6)

u_RThe calculating method of (2): assuming that the vehicle height is h, the height and width satisfy the proportional relationship h ═ γ (r)_R-l_R) Wherein γ is obtained from a set of vehicle data training; image [ l ] extraction using vertical Sobel operator_R,r_R]Horizontal edges within the range and counting the horizontal projection of the gray level image, assuming that P (y) represents the distribution of the horizontal projection histogram, then y ∈ [ b_R+0.5h,b_R+1.5h]The maximum value of P (y) in the range is u_RA value of (d);

by applying the above method, we can obtain b sequentially_R、l_R、r_RAnd u_RAnd taking values, thereby determining the target vehicle.

2. The method for detecting vehicles based on local symmetric features according to claim 1, wherein the feature corner point extraction method in the second step is a Harris corner point extraction algorithm.

3. The method for detecting video vehicles based on local symmetric features of claim 1, wherein the feature vector V of each feature corner point in the three steps_iThe construction method comprises the steps of utilizing a Haar wavelet template to calculate response values of a square area image, respectively obtaining Haar wavelet response values along the X direction and the Y direction, counting response values of ∑ dx, ∑ dy, ∑ | dx | and ∑ | dy |, and obtaining the response value of each characteristic angular point p_iFeature vector of

4. The method for video vehicle detection based on local symmetry features of claim 1, wherein the size of the square region constructed in step three is 5 x 5.

5. The method as claimed in claim 1, wherein the size of the Haar wavelet templates in the third step is 2 x 2.

6. The method for detecting video vehicles based on local symmetric features of claim 1, wherein the horizontal symmetric corner point q of each feature corner point in the fourth step_iFeature vector ofThe construction method comprises the following steps: