CN108550143A - A kind of measurement method of the vehicle length, width and height size based on RGB-D cameras - Google Patents

Abstract

The invention discloses a method for measuring the length, width and height of a vehicle based on an RGB-D camera. The camera is installed at a close distance to obtain a clear vehicle image, and the camera is calibrated by a calibration method based on a vanishing point to obtain a camera model. Internal and external parameters, through the camera depth image scheme, realize the transformation of the 3D point cloud of the vehicle target in the world coordinate system, and obtain the 3D coordinate information of the vehicle's outer surface; use prior knowledge and image processing methods to obtain the 3D coordinates of the vehicle's outer surface, according to the vehicle movement process The sequence graphics in the vehicle image are spliced by the registration method to realize the three-dimensional measurement of the vehicle shape; the calibration method overcomes the high equipment requirements and cumbersome operation of the traditional method, and the calibration accuracy is high; through the matching relationship synthesis of the corresponding points of the same position of the vehicle in the image sequence The actual displacement of the vehicle is analyzed, and the registration accuracy is high; the accurate splicing of the side of the vehicle is realized, the measurement error of the vehicle length is reduced, and the accuracy of the splicing of the side of the vehicle is improved.

Description

A measurement method of vehicle length, width and height based on RGB-D camera

technical field

The invention belongs to the field of intelligent transportation, and in particular relates to a method for measuring vehicle length, width and height dimensions based on an RGB-D camera.

Background technique

The automatic recognition technology of vehicle length, width and height is one of the key technologies in the field of ITS (Intelligence Transportation System, Intelligent Transportation System). It mainly collects the original video of the vehicle, and uses image processing technology and image correlation algorithm to realize the vehicle outline size. Measurement. The three-dimensional size of the vehicle outline can effectively help realize the accurate identification of the vehicle, thereby promoting the development and improvement of the intelligent transportation system (ITS).

The over-length, over-width and over-height of vehicles are one of the important hidden dangers of unsafe accidents in traffic scenes. At present, the security inspection line, comprehensive inspection line and overrun monitoring station of the vehicle management station basically measure the shape of the vehicle by manual measurement. The length, width and height of the house are labor-intensive, and there are problems such as injustice and cheating caused by human factors.

Contents of the invention

In order to solve the problems existing in the prior art, the present invention provides an automatic measurement method of vehicle length, width and height based on RGB-D cameras, which can extract the three-dimensional structure information and real size data of the detected vehicle, which can be very accurate Determine the vehicle outline size information.

In order to achieve the above object, the technical solution adopted in the present invention is a method for measuring the length, width and height of a vehicle based on an RGB-D camera, comprising the following steps:

Step 1. Combined with the traffic application scene, the camera is calibrated using the calibration method based on the vanishing point, and the internal and external parameters of the camera model are obtained;

Step 2, by using the RGB-D camera depth image scheme, realize the 3D point cloud transformation of the vehicle target in the world coordinate system, and then measure its length, width and height information according to the reconstructed 3D point cloud of the vehicle, so as to obtain the outer surface of the vehicle 3D coordinate information;

Step 201, using an RGB-D camera to shoot a scene containing a simulated vehicle to obtain depth information of the simulated vehicle;

Step 202, transforming the depth information image into a three-dimensional point cloud of the camera coordinate system to obtain a one-to-one correspondence with a three-dimensional point in the camera coordinate system;

Step 203, converting the three-position point cloud of the vehicle into the RGB-D camera coordinate system in step 202 into the world coordinate system;

Step 204, unify the 3D point cloud data of the vehicle obtained by the left and right cameras into the world coordinate system, complete the 3D point cloud of the complete vehicle image, take the world coordinate system of the left camera as the standard, and then transform the world coordinate system of the right camera;

Step 3. For large and medium-sized vehicles whose length cannot be fully accommodated in the camera field of view, according to the sequence images in the process of vehicle movement, the optical flow method is used for registration to obtain the displacement of the vehicle between image sequences; and then the vehicle feature Points are matched and screened, the corresponding relationship of the vehicle image is obtained from the image sequence, and finally the complete mosaic of the vehicle side projection image is realized;

Step 4, complete the measurement of the vehicle shape according to the vehicle top view determined in step 3, determine the length and width of the vehicle through the boundary detection of the vehicle top view, and determine the height of the vehicle by using the pixel value statistics map of the vehicle top view.

In step 1, the method based on the vanishing point is used to calibrate the camera. First, the camera internal parameter matrix K and the rotation matrix R in the camera external parameters are determined, and then the positive direction of the world coordinate system is determined, and then the new world coordinate origin is determined. In the traffic scene, the road plane is defaulted to the Z=0 plane, and the orthocenter of the camera on the road plane is taken as the origin of the world coordinate system, and finally the internal and external parameters of the camera model are obtained.

In step 201, the depth information of the vehicle scene includes color RGB information of a point in the scene and a distance value from the point in the scene to the camera plane.

In step 202, according to the distance from the point in the actual scene to the vertical plane where the camera is located as the depth value at the position of the imaging point on the depth image, the transformation expression from the depth information image to the camera coordinate system to the 3D point cloud is obtained.

In step 203, the ground plane is taken as the Z=0 plane, and the positive direction of the Z-axis is taken vertically upward on the ground; three non-collinear points are taken on the ground, and the coordinates of the three points are used to calculate the ground plane at the camera coordinates The unit normal vector under the system, according to the pinhole camera model, make the unit normal vector in the same direction as the unit normal vector in the ground world coordinate system, and then rotate around the Z axis so that the X axis direction of the current 3D point cloud is in the world coordinate system The X-axis direction coincides, and the current 3D point cloud is translated to the world coordinate system, and finally the conversion from the camera coordinate system to the world coordinate system is completed.

In step 204, it is assumed that the coordinates of a group of corresponding points in the common field of view of the left and right cameras in the respective world coordinate systems are Then its corresponding relationship is:

Among them, x _t , y _t are the translations of the right camera world coordinate system to the left camera world coordinate system on the X-axis and Y-axis respectively.

When the optical flow method is used to complete the vehicle graphics registration in step 3, it is assumed that the small movement and brightness of the optical flow are consistent, and the brightness in the local area of a certain pixel is constant, and then the motion vector of the pixel is obtained, and then the strong corner point is used. The optical flow field calculation method, and then extract the strong corner points in the kth frame image, and find the corresponding relationship of the strong corner points on the k+1th frame image through the optical flow method.

In step 3, the constraint method based on the rigid body motion displacement consistency is used to screen the results of image registration using the optical flow method. Before adopting the rigid body motion displacement consistency, the obviously wrong matching results are eliminated through prior knowledge. The feature points on the body and satisfying the consistency of the rigid body are closely aggregated together, that is, the feature points corresponding to the displacement of the frequency peak position in the statistical experiment graph.

In step 3, the prior knowledge is specifically as follows: first, the vehicle is moving, and the feature points on the background can be deleted; second, there is only translational motion in the X direction during the movement of the vehicle, and theoretically set the image of the feature points on the vehicle body The longitudinal displacement is zero, and the displacement of the feature points of the vehicle image is counted, and then the matching point with the strongest displacement consistency is retained to eliminate the wrong match.

The method of mosaicing the vehicle side projection image in step 3 is as follows: Let the vehicle drive along the positive direction of the X-axis, and each element of the feature point set of the vehicle only includes the X-axis position of the feature point, and obtain the vehicle from the kth frame to the k+th The actual displacement of n frames and the average position of the feature point set in the X direction; the height of the back-projected image of the vehicle side is H, and the width is W, then the k-th frame image u: x _k → W, v: 0 → H is obtained The area is used as the first half of the back-projected image of the side of the vehicle, and the area of the k+nth frame image u: x _k +Δx→W, v:0→H is taken as the second half, and the three-dimensional feature points are obtained through steps 202 and 203 Coordinates, and then realize the stitching of the top view of the vehicle.

Compared with the prior art, the present invention has at least the following beneficial effects: the realization scheme of the depth image based on the RGB-D camera is proposed, and the clear vehicle image is obtained by using the method of installing the camera at a close distance, and then using prior knowledge and image processing The method obtains the three-dimensional coordinates of the vehicle's outer surface, and then according to the sequence graphics in the process of vehicle movement, a complete vehicle image is spliced through the method of image registration, and finally the three-dimensional measurement of the vehicle's shape is realized;

The calibration method based on the vanishing point effectively utilizes the information in the traffic scene, such as marking lines on the road, vehicles, etc., and provides good calibration conditions for the determination of the vanishing point, thereby realizing the solution of the camera parameters and avoiding the equipment requirements of the traditional method The shortcomings of high and cumbersome operation, and high calibration accuracy;

The displacement of the vehicle is calculated through the matching relationship between the corresponding points of the same position of the vehicle in the image sequence, and the L-K optical flow method is used to determine the matching relationship, that is, the feature points are extracted in the current vehicle image, and then the location of these feature points in the subsequent vehicle image is searched. position, and then count the displacement of each feature point, and comprehensively analyze the actual displacement of the vehicle. This tracking vehicle image registration method can truly solve the real-time application problem while the registration accuracy is high;

The constraint method based on the consistency of rigid body motion displacement effectively eliminates the feature points of matching errors, realizes the accurate splicing of the vehicle side, reduces the measurement error of the vehicle length, and improves the accuracy of the vehicle side splicing.

Description of drawings

Figure 1 is an effect diagram of the simulated experimental scene;

Fig. 2 is a schematic diagram of a scene containing parallel lines in X, Y and Z directions;

Fig. 3 is a schematic diagram of a depth image;

Figure 4 is a schematic diagram of a pinhole camera model;

Figure 5(a) is a schematic diagram of the point cloud in the camera coordinate system, and Figure 5(b) is a schematic diagram of the point cloud in the world coordinate system;

Figure 6(a) is a map of the unified world coordinate system of the left and right cameras; Figure 6(b) is a 3D point cloud of the unified world coordinate system;

Figure 7 (a), (b), (c), and (d) are schematic diagrams of the splicing process of the back projection of the vehicle side in turn;

Figure 8 simulates the vehicle model;

Figure 9 stitching complete vehicle top view;

Fig. 10 is the boundary detection of the top view of the vehicle;

Fig. 11 is a height statistical histogram of pixel information.

Detailed ways

The solutions of the present invention will be further explained and described in detail below in conjunction with the accompanying drawings and specific embodiments.

A method for measuring the length, width and height of a vehicle based on an RGB-D camera, comprising the following steps:

Step 1: Combined with the traffic application scene, the camera is calibrated using the calibration method based on the vanishing point, and the internal and external parameters of the camera model are obtained; FIG. 1 is a simulation effect diagram of the measurement scene of the present invention.

Step 101, select a certain test scene, see Figure 1, set up the camera on one side of the environment to be tested, set the height of the camera to 2.75 meters according to the height of the vehicle, and set the image resolution to 320*240, the ground plane of the scene The size is about 350*400cm;

Step 102, mark the parallel lines in the X, Y and Z directions in the test scene in step 101, see Figure 2, the vanishing points in the three orthogonal directions in the scene reflect the mapping relationship between the world coordinates and the image coordinates, through their Position constraints, solve the internal and external parameters of the camera;

Step 103, use the method based on the vanishing point to calibrate the camera, let (u _x , v _x ), (u _y , v _y ), (u _z , v _z ) be the X direction, Y direction and The vanishing point formed on the imaging plane in the Z direction, assuming that at the position of the vanishing point, the actual coordinate value in this direction is ∞, and the coordinate values in the other two directions perpendicular to the imaging plane are 0, the vanishing point in the X direction is (u _x ,v _x ) as an example, the normalized actual coordinates are:

Among them, λ _x represents the normalization factor of the world coordinates of the vanishing point in the X direction. Similarly, it can be obtained:

Using the above formula, the internal parameter matrix K can be obtained, and the rotation matrix R in the external parameters of the camera can be obtained at the same time, namely:

R＝K ^-1 (KR) (3)

When the rotation matrix is determined, the positive direction of the world coordinate system is determined, and the origin is the spatial position of the camera; determine the new origin of the world coordinates, and then translate the camera to this point, if point P is used to represent the coordinates of the current camera position, the coordinates are (0,0,0), P, is the origin of the new world coordinate system, and the coordinates are (X _cam , Y _cam , Z _cam ), then the process can be expressed as:

P'=R(P-T) (4)

In the traffic scene, the road plane is defaulted to the Z=0 plane, and the orthocenter of the camera on the road plane is taken as the origin of the world coordinate system, so

in, Get the internal and external parameters of the camera model for the height of the camera relative to the road plane.

Step 2: Use the camera depth image scheme to realize the transformation of the 3D point cloud of the vehicle target in the world coordinate system, and then measure its length, width and height information according to the reconstructed 3D point cloud of the vehicle, so as to obtain the 3D coordinate information of the outer surface of the vehicle;

Step 201, use RGB-D camera to shoot the scene containing the simulated vehicle, and obtain the depth information of the simulated vehicle, the depth information includes the color RGB information of the center point in the scene and the distance value from the center point in the scene to the camera plane.

Step 202, realize the transformation of the depth information image to the three-dimensional point cloud of the camera coordinate system, see Figure 3, the point P in the actual scene, the imaging point on the depth image is p, from point P to the vertical plane where the camera is located (ie X The distance of _c Y _c plane) is the depth value at the position of image point p; where, the three-dimensional coordinate system of the camera takes the camera center as the coordinate origin, the vertical direction of the camera plane is the Z _c axis, and the horizontal and vertical directions of the camera plane are respectively is the X _c axis and the Y _c axis, assuming that the coordinates of the point P in the actual scene in the camera coordinate system are (X _c , Y _c , Z _c ), and the coordinates of the projected point p of the point on the imaging plane are (u, v), the depth value is D(u,v), then the point cloud from the depth image to the camera coordinate system is transformed into:

Among them, u ₀ and v ₀ are the center coordinates of the depth image, and f is the focal length of the camera. According to the above formula, each point on the depth image corresponds to a three-dimensional point in the camera coordinate system. When measuring, the road plane is generally taken as the zero plane, and the vertical direction is taken as the height direction.

Step 203, transform the three-dimensional point cloud of the vehicle in the camera coordinate system into the world coordinate system defined above, that is, the ground plane is the Z=0 plane, and the positive direction of the Z axis is vertical to the ground. First, in Take three non-collinear points on the ground, and their coordinates in the camera coordinate system are P _c1 (X _c1 , Y _c1 , Z _c1 ), P _c2 (X _c2 , Y _c2 , Z _c2 ) and P _c3 ( X _c3 , Y _c3 , Z _c3 ), calculate the unit normal vector of the ground plane in the camera coordinate system but:

According to the pinhole camera model, see Figure 4, so that the unit normal vector and the unit normal vector in the ground world coordinate system In the same direction, then rotate around the Z axis so that the X axis direction of the current 3D point cloud coincides with the X axis direction defined in step 1. Finally, it is only necessary to translate the current 3D point cloud to the world coordinate system, thus completing the camera coordinate system to the world The conversion of the coordinate system, the conversion results are shown in Figure 5(a) and Figure 5(b);

Step 204, unify the 3D point cloud data of the vehicle obtained by the left and right RGB-D cameras into the world coordinate system to complete the 3D point cloud of the complete image of the vehicle, taking the world coordinate system of the left camera as the standard, and then transform the world coordinate system of the right camera , wherein, according to the definition of the world coordinate system in step 1, the coordinate axes of the left and right camera coordinate systems are collinear, that is, the corresponding coordinate axes are either in the same direction or in the opposite direction, assuming that a group of corresponding points in the common field of view of the left and right cameras is in The coordinates of the respective world coordinate systems are Then its corresponding relationship is:

Among them, x _t and y _t are the translations of the right camera world coordinate system to the left camera world coordinate system on the X-axis and Y-axis respectively, and the specific conversion is shown in Figure 6(a) and Figure 6(b);

Step 3: According to the sequence images in the process of vehicle movement, determine the registration method of large and medium-sized vehicle images with a long length, and realize the complete stitching of vehicle images,

In step 301, the optical flow method is used to complete the registration of the vehicle graphics. On the premise that the small movement and brightness of the optical flow are relatively consistent, it can be concluded that:

I(x,y,t)=I(x+dx,y+dy,t+dt) (9)

Where t is time, the first-order Taylor series expansion of this formula can be obtained:

That is to say: I _x dx+I _y dy+I _t dt=0, let Then there are:

I _x u + I _y v = -I _t (11)

If it is assumed that the brightness of the pixel point (u, v) in the local area is constant, it can be obtained:

which is: The calculation purpose of optical flow is to make: The minimum, so that the motion vector of the pixel is obtained:

In practical applications, there are two pictures of I _t (x, y) and I _t+1 (x, y). To calculate the motion of a pixel in I _t to I _t+1 , first, in I Find points with consistent pixels around the corresponding position _{of t} , and use the optical flow field calculation method based on strong corner points (LK method) through the above calculation, and then extract the Harrs corner points in the k-th frame image, and find these corner points through the optical flow method Correspondence on the k+1th frame image.

Step 302, for part of the matching error in step 301, use a constraint method based on the consistency of rigid body motion displacement to solve it. Before adopting the rigid body motion displacement consistency, delete the matching results that are obviously wrong by prior knowledge. The prior knowledge is : First, the vehicle is moving, and the feature points on the background can be deleted; second, there is only translational motion in the X direction during the vehicle movement, that is, the longitudinal displacement of the image of the feature points on the vehicle is theoretically zero; And the feature points that satisfy the rigid body consistency are closely aggregated together, that is, the feature points corresponding to the displacement of the frequency peak position in the statistical experiment graph.

In step 303, through the matching and screening of the vehicle feature points in step 302, the corresponding relationship of the vehicle images is obtained from the image sequence, and then the complete splicing of the vehicle images is realized. The form of the vehicle along the positive direction of the X-axis, let the matching feature point sets on the kth frame and the k+nth frame image be p _k and p _k+n respectively, where p _k ={p _k (i)| _{i=1, 2,... m} }, p _k+n = {p _k+n (i) | _{i=1, 2,... m} }. Because the vehicle is traveling along the positive direction of the X axis, each element of the feature point set only contains the X position of the feature point, then the actual displacement of the vehicle from the kth frame to the k+nth frame is:

Then on the kth frame image, the average position of the feature point set in the X direction is:

Assuming that the height of the designed vehicle side back-projection image is H and the width is W, then the area of the k-th frame image u: x _k → W, v: 0 → H is taken as the first half of the vehicle side back-projection image, and the k-th frame image is taken +n frames of image u:x _k +Δx→W, v:0→H are used as the second half, which can be put together directly. For the stitching process of the complete vehicle side back projection, see Figure 7(a), 7(b ), 7(c) and 7(d); the three-dimensional coordinates of the feature points are obtained through step 2, so as to realize the stitching of the top view of the vehicle, see Figure 9 (Figure 8 is a simulated vehicle).

Step 4 is to complete the vehicle shape measurement based on the top view of the vehicle. Obviously, the boundary of the top view of the vehicle directly reflects the contour of the vehicle's spatial position, and its pixel value reflects the height information of the vehicle. Therefore, the length and width of the vehicle can be determined by detecting the boundary of the top view of the vehicle, and the height of the vehicle can be determined by using the statistical map of the pixel values of the top view of the vehicle. See Figure 10 for the boundary detection of the top view of the vehicle, and Figure 11 for the height statistical histogram. The three-dimensional size of the known vehicle model is: length*width*height=310cm*240cm*189cm. In 10 repeated experiments, the actual measured average value is: 306.9cm*243.4cm*190.8cm has good measurement accuracy.

Claims (10)

1. A method for measuring the length, width and height of a vehicle based on an RGB-D camera is characterized by comprising the following steps:

step 1, calibrating a camera by adopting a calibration method based on a vanishing point in combination with a traffic application scene to obtain internal and external parameters of a camera model;

step 2, realizing three-dimensional point cloud conversion of the vehicle target under a world coordinate system by adopting an RGB-D camera depth image scheme, and then measuring length, width and height information of the vehicle target according to the reconstructed vehicle three-dimensional point cloud so as to obtain three-dimensional coordinate information of the outer surface of the vehicle;

step 201, shooting a scene containing a simulated vehicle by using an RGB-D camera to acquire depth information of the simulated vehicle;

step 202, converting a three-dimensional point cloud of a depth information image to a camera coordinate system, and enabling each point on the depth image to correspond to the camera coordinate system one by one;

step 203, converting the vehicle three-dimensional point cloud converted to the RGB-D camera coordinate system in the step 202 into a world coordinate system;

step 204, unifying the vehicle three-dimensional point cloud data obtained by the left camera and the right camera under a world coordinate system to complete the three-dimensional point cloud of the complete image of the vehicle, taking the world coordinate system of the left camera as a standard, and then converting the world coordinate system of the right camera;

step 3, registering large and medium vehicles with the length which cannot be completely accommodated in the camera vision field by adopting an optical flow method according to sequence images in the vehicle motion process to obtain the displacement of the vehicles among the image sequences; matching and screening the vehicle characteristic points, obtaining the corresponding relation of the vehicle images from the image sequence, and finally realizing the complete splicing of the projected images on the side surfaces of the vehicles;

and 4, finishing the measurement of the vehicle appearance according to the vehicle top view determined in the step 3, determining the length and the width of the vehicle through the boundary detection of the vehicle top view, and determining the height of the vehicle by using a pixel value statistical map of the vehicle top view.

2. The method for measuring the length, width and height of the vehicle based on the RGB-D camera as claimed in claim 1, wherein in step 1, the camera is calibrated by using a vanishing point-based method, the parameter matrix K in the camera and the rotation matrix R in the parameters outside the camera are firstly obtained, then the positive direction of the world coordinate system is determined, then a new world coordinate origin is determined, in a traffic scene, the road plane is defaulted to be the Z-0 plane, the vertical center of the camera on the road plane is used as the world coordinate origin, and finally the internal and external parameters of the camera model are obtained.

3. The method as claimed in claim 1, wherein the depth information of the vehicle scene includes color RGB information of points in the scene and a distance value from the points in the scene to the camera plane in step 201.

4. The method as claimed in claim 1, wherein in step 202, the conversion expression from the depth information image to the camera coordinate system to the three-dimensional point cloud is obtained by taking the distance from the point in the actual scene to the vertical plane where the camera is located as the depth value at the imaging point position on the depth image.

5. The RGB-D camera based vehicle length, width and height measurement method according to claim 1, wherein in step 203, the ground plane is the Z-0 plane, and the vertical ground direction is the positive direction of the Z axis; taking three non-collinear points on the ground, calculating a unit normal vector of a ground plane under a camera coordinate system through coordinates of the three points, enabling the unit normal vector to be in the same direction as the unit normal vector under a ground world coordinate system according to a pinhole camera model, rotating around a Z axis to enable the X axis direction of the current three-dimensional point cloud to be coincident with the X axis direction in the world coordinate system, translating the current three-dimensional point cloud to be under the world coordinate system, and finally completing the conversion from the camera coordinate system to the world coordinate system.

6. The method as claimed in claim 1, wherein the coordinates of a set of corresponding points in the common field of view of the left and right cameras in the world coordinate system are set as P in step 204_L(X_L,Y_L,h)，P_R(X_R,Y_RH), its corresponding relationship is:

wherein x is_t，y_tRespectively, the translation of the right camera world coordinate system to the left camera world coordinate system in the X-axis and Y-axis.

7. The method as claimed in claim 1, wherein when the optical flow method is used in step 3 to complete the alignment of the vehicle graphics, the minute motion and brightness of the optical flow are consistent, and the brightness in the local area of a certain pixel point is constant, so as to obtain the motion vector of the pixel, and then the method for calculating the optical flow field based on the strong angle point is used to extract the strong angle point from the k frame image, and the corresponding relationship of the strong angle point on the k +1 frame image is found by the optical flow method.

8. The method as claimed in claim 1, wherein the step 3 is performed by using a constraint method based on the consistency of rigid motion displacement to perform a sexual screening on the result of image registration by using an optical flow method, before the consistency of rigid motion displacement is used, the matching result with obvious error is removed by using priori knowledge, and then the feature points on the vehicle body and meeting the consistency of rigid motion are closely gathered, that is, the feature points of the frequency number peak position corresponding to the displacement in the experimental graph are counted.

9. The RGB-D camera-based vehicle length, width, height and size measuring method according to claim 8, wherein in step 3, the prior knowledge is specifically: firstly, the vehicle is moving and can delete the characteristic points on the background; and secondly, only translational motion in the X direction exists in the vehicle motion process, the longitudinal displacement of the image of the characteristic points on the vehicle body is theoretically set to be zero, the displacement of the characteristic points of the vehicle image is counted, then the matching points with the strongest consistency of the displacement are reserved, and the elimination of the error matching is realized.

10. The method as claimed in claim 1, wherein the step 3 of projecting the image of the side of the vehicle is performedThe splicing method comprises the following steps: setting a vehicle to run along the positive direction of an X axis, wherein each element of a characteristic point set of the vehicle only comprises the X axis position of a characteristic point, and obtaining the actual displacement of the vehicle from the k frame to the k + n frame and the average position of the characteristic point set in the X direction; if the height of the vehicle side reverse projection image is H and the width is W, the k frame image u: x is obtained_kThe area → W, v:0 → H is used as the first half of the vehicle side inverse projection image, and the k + n frame image u: x is taken_kThe area of + Δ x → W, v:0 → H is used as the latter half, and the three-dimensional coordinates of the feature points are obtained through steps 202 and 203, thereby realizing the top view stitching of the vehicle.