A fog environment intelligent automobile camera perception test model and test method

Abstract

The present application relates to a kind of fog environment intelligent automobile camera perception test model and test method, pixel of no fog image is classified using semantic segmentation, obtain sky area, road area and traffic participant area;And considering the feature of corresponding area respectively carries out atmospheric light estimation. Using the expected simulation of fog visibility, in combination with the depth of field of test scene, the transmissivity of fog is estimated, and the scene object depth of field distribution of image plane is determined by camera geometric model to ensure accurate fog simulation, finally the fog image is output through image blur model.In the process of camera fog test, CON 50 Index is used to evaluate image recognition algorithm, describe the function boundary of image recognition algorithm under fog environment, and guide the scene building of simulation or real vehicle test, according to CON 50 Arrange target object, test the intelligent automobile camera perception function of this visibility, improve test efficiency and reduce test cost.

Description

TECHNICAL FIELD

[0001] The application relates to an intelligent automobile camera perception test model and a test method, in particular to a fog environment intelligent automobile camera perception test model and a test method. BACKGROUND

[0002] Scene-based testing is an effective method to ensure the actual landing of intelligent automobiles. However, due to the complex and changeable actual scene of intelligent automobiles, traditional mileage-based real vehicle testing has defects such as high cost, and therefore, in recent years, a simulation-based test scene construction method has become a hot spot. Meteorological environment can have a significant impact on the perception system of an intelligent automobile. A camera is one of the most commonly used sensors in the perception system, and the camera is a passive sensor, which is more affected by meteorological environmental factors in the scene than an active sensor. Fog, as one of the most common complex meteorological phenomena in the real world, often appears in the driving environment and can seriously affect camera perception. Therefore, fog environment testing is an essential link before the actual landing of an intelligent automobile, but in the actual test scene construction process, there is a severe lack of image natural driving data sets for fog environment, and image style transfer is an effective solution to this problem. However, the current fog image style transfer method applied to the construction of intelligent automobile test scenes still has some problems.

[0003] Firstly, for the application of intelligent automobile camera testing, unlike other application scenarios, the driving track of most working conditions of the vehicle is parallel to the road surface, and the sky is mostly located in the upper half of the image. For road factors such as lane lines, pedestrian information, road conditions and other information, they are mostly located in the lower half of the image. The current image style transfer method does not consider the influence of the image location area characteristics.

[0004] Secondly, the current style transfer method lacks an image blurring algorithm that directly controls the fog concentration. For the construction of intelligent automobile test scenes, a large number of scene test camera function boundaries with different fog visibilities are needed. If the original images with different fog concentrations are completely trained, a large amount of manpower and resources will be consumed. Moreover, the current fog simulation method uses direct changes in the atmospheric scattering coefficient for simulation, which can only perform global uniform fog simulation, and therefore, the simulation method cannot be applied to actual camera perception testing.

[0005] Therefore, there is an urgent need for an image style transfer method that can precisely control the fog visibility of the generated image, so as to construct test scenes with different visibilities and apply them to intelligent automobile camera perception testing. SUMMARY

[0006] To solve the above technical problems, the application provides a fog environment intelligent automobile camera perception test model, which is constructed by the following steps:

[0007] (1) Determine the depth of field distribution of different scene elements

[0008] The input of the camera geometry model is the test scene element geometric position information, and the output is the position of the scene element in the camera imaging plane, so as to determine the distribution of the depth of field of different scene elements in the image plane;

[0009] The scene elements include a traffic participant area, a sky area and a road area;

[0010] For the traffic participant area, the centroid position of the target object is obtained by the test scene element geometric position information, the axis-aligned bounding box (AABB) of the traffic participant is determined by the length, width and height of the target object, and the intersection relationship between the AABB and the camera imaging plane is determined by the reverse calculation of the line connecting the pixel point and the camera optical center, and then the position of the traffic participant area in the camera imaging plane is determined. The extension line of the line connecting the pixel point and the optical center does not intersect with the AABB of the target object in the scene, which is an invalid pixel; the extension line of the line connecting the pixel point and the optical center intersects with the AABB of the target object in the scene, which is a valid pixel. The depth of field distribution of the traffic participant is determined by using this method. For the sky area, the depth of field of this part is set to a larger value compared with the traffic participant area, which is approximately infinite. For the road area, the depth of field of the road area is determined by using the camera parameters, and the depth of field distribution of the road area is established by the installation parameters and built-in parameters of the camera.

[0011] After determining the depth of field distribution of different scene elements, the fog transmittance at different positions can be determined to ensure the fidelity of the fog image simulation.

[0012] (2) Fog image blur model construction

[0013] In a foggy environment, due to the scattering effect of atmospheric particles, the contrast and color change of the image are mainly composed of two parts. On the one hand, the reflected light from the surface of the object is lost due to scattering, which reduces the light intensity reaching the camera end. The degree of influence will decrease exponentially with the increase of distance. On the other hand, due to the effect of environmental light and particle scattering, atmospheric light is generated, which makes the medium surface show light source characteristics. Therefore, the whole image blur model is composed of two models, namely the incident light attenuation model and the atmospheric light imaging model, and its expression is:

[0014] I(x) = E0(λ)t(x) + A(1-t(x)) (1)

[0015] t(x) = e -β(λ)d(x) (2)

[0016] In the formula, I(x) is a fog image; x is a pixel position coordinate; E0(λ) is a target object reflected light intensity when the depth of field is 0, also represents a fog-free image; λ is a wavelength; t(x) is a fog transmittance; A is an atmospheric light; β(λ is an atmospheric scattering coefficient; d(x) is a scene depth.

[0017] Further determine the fog transmittance and atmospheric light estimation:

[0018] (2.1) Fog transmittance calculation

[0019] In order to construct the fog image blur model with controllable visibility, it is necessary to establish the relationship between visibility V and fog transmittance. Based on the atmospheric attenuation model, the luminance contrast attenuation expression under the fog environment condition is:

[0020] H=H0e -γD (3)

[0021] In the formula, H is the luminance contrast; the formula is explained as the target object can be perceived by human vision when the contrast reaches H0 when the atmospheric extinction coefficient of the fog is γ and the distance from the camera D.

[0022] In the actual test scene application, the scene fog intensity information is defined by the visibility V; the International Commission on Illumination defines the visibility as: the maximum distance corresponding to the specified size of the black target object (H0=1) that can be observed by the naked eye of human under the condition that the contrast threshold value H / H0 is 0.05. According to the definition, the relationship between V and γ can be derived:

[0023]

[0024] The present application assumes that the fog model is a pure scattering fog, the main effect on the light is scattering, the absorption effect of the fog is relatively small, so the light mainly occurs scattering phenomenon in the propagation process, and the absorption effect of the fog is ignored in the present application, and the relationship expression between the atmospheric visibility and the transmittance is obtained:

[0025]

[0026] In the formula, d(x) is the scene depth; t(x) is the transmittance; when the input visibility V needs to be simulated, the fog transmittance t(x) is obtained through formula (5).

[0027] (2.2) Atmospheric light value calculation

[0028] The global atmospheric light value is obtained by globally calculating the image, first calculating the dark features of the image, and the calculation method expression is:

[0029]

[0030] In the formula, I dark To obtain the dark channel feature image; I C Let x be the original input image; C represents the R, G, and B channels of the image; Ω represents the neighborhood centered at pixel x; y represents the set of all pixels within the neighborhood Ω; I is defined as follows. dark The average value of the top 0.1% of pixels is the global atmospheric light value (Air).

[0031]

[0032]

[0033] In the formula, x air For I dark The position of the top 0.1% of pixels; I(x) air ) represents the corresponding x air The pixel intensity at a given location, and its average value, is the global atmospheric light value;

[0034] Semantic segmentation is used to divide image information into sky region, traffic participant region and road region, and the corresponding atmospheric light is obtained to simulate fog image;

[0035] For the sky region, compensation is performed using the bright channel. The method for obtaining prior information of the bright channel image is the opposite of that for the dark channel; the largest pixel value among the three channels of the image is used as the bright feature, expressed as:

[0036]

[0037] Ω represents the neighborhood centered at pixel position x; select I ligh t The average value of the top 0.1% of pixels is used as the bright feature output, expressed as:

[0038]

[0039]

[0040] In the formula, x air-sky For I ligh t The position of the top 0.1% of pixels; I(x) air-sky ) represents the corresponding x air-sky The pixel intensity at a given location; then the atmospheric light A in the sky region. sky The value can be:

[0041] A sky =μAir + (1-μ)Air sky (12)

[0042] In the formula, μ is the proportion of the sky region to all pixels, ranging between 0 and 1, which is determined by semantic segmentation; the local air light of the road region and the traffic participant region is determined according to the dark feature determination mode, and the global air light value Air is modified according to the pixel proportion output by the semantic segmentation;

[0043] (3) The image blurring model is input by combining the air light value of the corresponding region with the fog transmittance, and a fog-added image is output.

[0044] The application further provides a fog environment-oriented intelligent automobile camera perception test method, wherein the fog environment intelligent automobile camera perception test model is used for intelligent automobile camera perception test, and the test task is crucial for determining the functional boundary of the image perception algorithm, i.e., finding specific scene parameter combinations that cause the tested algorithm to fail; the test procedure is as follows:

[0045] 1) Determine the image processing algorithm to be tested;

[0046] 2) Determine the logical scene, i.e., determine the value range of the scene parameters, including the initial distance Dis between the target vehicle and the host vehicle = 1-150 m, the fog visibility V to be simulated = 15-80 m, and the target object uniform motion speed

[0047] 3) Build a test scene by using a virtual simulation scene, collect image data of a no-fog scene parameter, perform fog addition processing on the image by using the model construction method proposed in the application to obtain a specific visibility, and test the image perception algorithm to be tested by using the image after fog addition;

[0048] 4) Record the confidence of the target recognition result corresponding to different visibilities and the distance relationship between the two vehicles, and take the intersection of the straight line with a confidence of 0.5, wherein the Euclidean distance between the two vehicles corresponding to the first intersection point is recorded as CON 50 The application takes CON 50 as the evaluation index of the image recognition algorithm under the fog visibility scene, and the smaller the CON 50 value, the easier the image recognition algorithm is to fail under the fog environment; and the physical meaning of CON 50 is that the distance boundary of the tested algorithm for the target object under the fog visibility.

[0049] The application has the following beneficial effects:

[0050] The present application classifies the pixels of the haze-free image by semantic segmentation to obtain the sky region, the road region and the traffic participant region, and estimates the atmospheric light according to the features of the corresponding regions. In addition, the transmittance of the fog is estimated by combining the test scene depth with the expected simulated fog visibility of the test personnel, and the scene object depth distribution of the image plane is determined through the camera geometric model to ensure accurate fog simulation, and finally the fog image is output through the image blur model. In the actual camera fog test application process, the present application uses CON 50 index to evaluate the image recognition algorithm, CON 50 can describe the functional boundaries of image recognition algorithms in foggy environments and guide the construction of simulation or real vehicle test scenes. When testing in real vehicles, the target objects can be arranged according to CON 50 to test the intelligent vehicle camera perception function at this visibility, which can improve test efficiency and reduce test cost. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 A flowchart of the present application fog environment intelligent vehicle camera perception test model construction;

[0052] Figure 2 AABB for the vehicle target of the present application;

[0053] Figure 3 A schematic diagram of the present application road region depth determination method;

[0054] Figure 4 A schematic diagram of the relationship between the road pixel distance and the actual distance below the image;

[0055] Figure 5 A schematic diagram of the pixel position geometric relationship;

[0056] Figure 6 A schematic diagram of image blur in fog environment;

[0057] Figure 7 A schematic diagram of the real vehicle driving environment;

[0058] Figure 8 A schematic diagram of the semantic segmentation result of the present application;

[0059] Figure 9 A schematic diagram of the fog image simulation result of different visibility of the present application;

[0060] Figure 10 A schematic diagram of the test scene of the present application;

[0061] Figure 11 A schematic diagram of the fog environment intelligent vehicle camera perception algorithm test result of the present application. DETAILED DESCRIPTION

[0062] The present application provides a kind of fog environment intelligent automobile camera perception test model, refer to Figure 1 As shown in the following steps are constructed:

[0063] 1, determine the depth of field distribution of different scene elements

[0064] The input of camera geometric model is the element geometric position information of test scene, and the position of scene element in camera imaging plane is output.The role of this model is to determine the distribution of depth of field of different scene elements in the image plane, and after determining the distribution, the fog transmittance of different positions can be determined to ensure the fidelity of fog image simulation.For traffic participant area, sky area and road area, there are different depth of field distribution characteristics.This application takes traffic participants as an example, and the depth of field distribution characteristics of the other two areas can be obtained in the same way.

[0065] For traffic participant area, the centroid position of target object can be obtained from the geometric position information of test scene element, the axis-aligned bounding box (AABB) of vehicle is determined by the length, width and height of target object, and the intersection relationship between the reverse calculation of the line from pixel point to camera optical center and the AABB is used to determine the position of traffic participant area in camera imaging plane.First, the AABB of target object is extended from the centroid point, as shown in Figure 2 Taking vehicle target object as an example.

[0066] As shown in Figure 5 Starting from the pixel point of camera plane, the intersection relationship between the light ray and the AABB is used to determine the pixel position in the image plane.

[0067] Invalid pixel refers to the pixel and the extension line of the optical center line do not intersect with the AABB of the target object in the scene; valid pixel refers to the pixel and the extension line of the optical center line intersect with the AABB of the target object in the scene. The depth of field distribution of traffic participants is determined by this method.

[0068] For sky area, the present application sets the depth of field of this part to a larger value compared with traffic participant area, which is approximately infinite.

[0069] For road area, the present application determines the depth of field of road area by camera parameters, and establishes the depth of field distribution of road area by camera installation parameters and built-in parameters. The installation parameters of camera include installation angle δ I , installation height H I and horizontal distance Dis between camera and the front end of the vehicle, δ I is the angle between camera optical axis and vertical plane in Figure 3 , H Iis the vertical height of the camera's center of mass from the ground; and the horizontal distance between the camera and the front end of the vehicle, denoted as Dis. For the camera intrinsic parameters, the camera intrinsic parameters considered by the present application include the vertical field of view fov, the camera focal length f, as shown in Figure 3 real represents the actual distance between the road and the camera (in meters), d imgage represents the distance between the road region pixel and the lower edge of the image (in pixels), δ O represents the angle between the camera optical axis and the line connecting the projection of the target object on the imaging plane. According to the principle of similar triangles and the principle of camera transmission projection, we can get:

[0070]

[0071] From the above formula, we can get the relationship between d image and D real , as shown in Figure 4 , which is used to determine the depth of field distribution of the road region.

[0072] 2. Fog image blur model construction

[0073] Fog is a complex meteorological environment caused by the condensation of water vapor into small raindrops due to the decrease in air temperature when the air humidity reaches 100%. In a foggy environment, the scattering effect of water droplets causes the reflected light of the target object to weaken, resulting in a decrease in the object's visibility, which has a serious impact on the camera perception of intelligent vehicles.

[0074] As shown in Figure 6 , in a foggy environment, due to the scattering effect of atmospheric particles, the contrast and color change of the image are mainly composed of two aspects. On the one hand, the reflected light from the object surface is lost due to scattering, reducing the light intensity reaching the camera end. This degree of influence decreases exponentially with distance. On the other hand, due to the scattering effect of environmental light and particles, atmospheric light is produced, making the medium surface exhibit light source characteristics. Therefore, the entire image blur model is composed of two models: an incident light attenuation model and an atmospheric light imaging model, and its expression is:

[0075] I(x) = E0(λ)t(x) + A(1-t(x)) (2)

[0076] t(x) = e -β(λ)d(x) (3)

[0077] In the formula, E0(λ) is the reflected light intensity of the target object when the depth of field is 0, which also represents the fog-free image; β(λ) is the atmospheric scattering coefficient; d(x) is the scene depth; λ is the wavelength; A is the atmospheric light; t(x) is the fog transmittance; I(x) is the fog image; and x is the pixel position coordinate. ​

[0078] The present application further determines the fog transmittance and the atmospheric light estimation:

[0079] 2.1, Fog transmittance calculation

[0080] In order to construct the fog image blur model with controllable visibility, it is necessary to establish the relationship between visibility V and fog transmittance. Based on the atmospheric attenuation model, previous studies have proposed a luminance contrast attenuation expression under fog environment conditions:

[0081] H = H0e -γD (4)

[0082] In the formula, H is the luminance contrast. This formula can be explained as follows: when the atmospheric extinction coefficient of the fog is γ and the distance from the camera D, the target object can be perceived by human vision when the contrast reaches H0. In actual test scene applications, the fog intensity information of the scene is often defined by the visibility V. The International Commission on Illumination defines the visibility as follows: the maximum distance corresponding to the specified size of the black target object (H0 = 1) that can be observed by the naked eye of a human being when the contrast threshold value H / H0 is 0.05. According to this definition, the relationship between V and γ can be derived as follows:

[0083]

[0084] Since the fog model assumed in the present application is a pure scattering fog, the main effect on the light is scattering, and the absorption effect of the fog is relatively small, so the light mainly undergoes scattering phenomena during propagation, and the absorption effect is relatively small. In the present application, the absorption effect of the fog is ignored, so the relationship expression between the atmospheric visibility and the transmittance can be derived as follows:

[0085]

[0086] In the formula, d(x) is the scene depth; t(x) is the transmittance. Therefore, when the input visibility V to be simulated is input, the fog transmittance t(x) can be derived through formula (6).

[0087] 2.2, Atmospheric light value calculation

[0088] The traditional atmospheric light estimation method is to obtain the dark feature of the image globally, and the expression for obtaining the dark channel feature image is:

[0089]

[0090] In the formula, I dark is the dark channel feature image obtained; I Cis the input original image; C is the R, G, B three channels where the image is located; Omega is a neighborhood with the x point pixel position as the center, and the present application takes a range of 9*9; y is a set of all pixel points in the neighborhood Omega. dark The average value corresponding to the first 0.1% of pixel points in I is the global airlight value Air:

[0091]

[0092]

[0093] In the formula, x air is the position of the first 0.1% of pixel points in I dark ; I(x air ) represents the pixel intensity corresponding to the x air position, and the average value is the global airlight.

[0094] However, there is about 30% of the sky area in the forward image environment of the intelligent automobile driving, as shown in FIG. 1. Figure 7 Through actual experiments, it is found that due to the existence of a large area of sky area, the airlight value estimated by directly passing through the dark channel is inaccurate, and the fog image generated by using the traditional method is prone to distortion. The main reason for the distortion is that the airlight of each area in the driving environment is not estimated accurately, so that an accurate blurred image cannot be generated. Therefore, the present application uses semantic segmentation to divide the image information into three areas of the sky area, the traffic participant area and the road surface area, respectively calculates the corresponding airlight to simulate the fog image, as shown in FIG. 2. Figure 8

[0095] The light channel is used for compensation for the sky area. The light channel image prior information is obtained in the opposite way to the dark channel. The maximum channel pixel value in the three channels of the image is used as the light feature, and the expression is as follows:

[0096]

[0097] Omega is a neighborhood with the x point pixel position as the center, and the present application takes a range of 9*9. Secondly, the same as the dark feature, the average value of the first 0.1% of pixel values in I ligh t is taken as the light feature output, and the expression is as follows:

[0098]

[0099]

[0100] In the formula, x air-sky is the position of the first 0.1% of pixel points in I ligh t ; I(x air-sky ) represents the pixel intensity corresponding to the x air-sky ​The pixel intensity of the position. Then the atmospheric light A of the sky region sky is taken as

[0101] A sky = mu * Air + (1-mu) * Air sky (13)

[0102] In the formula, mu is the proportion of the sky region in all pixels, ranging from 0 to 1, which is determined by semantic segmentation. For the road region and the traffic participant region, the atmospheric light of the corresponding region is calculated according to the dark feature calculation method, and the global atmospheric light value Air is modified according to the pixel proportion output by semantic segmentation.

[0103] 3. The image is output after fogging by inputting the image into the fogging model using the atmospheric light value of the corresponding region and the fog transmittance, as shown in Figure 9 .

[0104] Test application of the fog model

[0105] The present application provides a kind of intelligent car camera perception test method for fog environment, and the intelligent car camera perception test of the above fog environment intelligent car camera perception test model is carried out, and the test task is critical to determine the function boundary of image perception algorithm, i.e. find out under which specific scene parameter combination will cause the failure of the algorithm to be tested. The model proposed in the present application can quickly test the image perception algorithm in the fog environment. The specific test process is as follows:

[0106] 1) determine the image processing algorithm to be tested, and the present application selects YOLO-V5 image recognition algorithm for testing;

[0107] 2) determine the logic scene, i.e. determine the value range of scene parameters, the present application includes the initial distance Dis between the target vehicle and the vehicle, the fog visibility V to be simulated, and the range of the target object uniform motion speed , the scene schematic diagram is shown in Figure 10 .

[0108] 3) build a test scene using virtual simulation scene, collect image data of no-fog scene parameters, and use the method proposed in the present application to add fog to the image with a specific visibility, and use the image after fogging to test the image perception algorithm to be tested;

[0109] 4) record the confidence of the target recognition result corresponding to different visibility and the distance relationship between the two vehicles, and take the intersection with the straight line with confidence of 0.5, and the two-car Euclidean distance corresponding to the first intersection point is recorded as CON 50 , and the present application takes CON 50As an evaluation index of the image recognition algorithm in the fog visibility scene, CON 50 The smaller the value represents the easier the image recognition algorithm fails in the fog environment. 50 The physical meaning of CON

[0110] The test results of the fog visibility by the above steps are shown in Table 1. Figure 11 As shown in Table 1, the test results of CON 50 are 39.9167, 56.6147 and 84.2147 meters respectively under the fog conditions with the simulated visibility of 44, 50 and 100 meters by YOLO-V5. It can be concluded that the recognition accuracy of YOLO-V5 decreases with the decrease of the simulated fog visibility, which verifies the effectiveness of the method proposed in the present application.

Claims

1. A smart car camera perception test model for foggy environments, characterized in that: The construction is carried out through the following steps: (1) Determine the depth of field distribution of different scene elements The input of the camera geometry model is the geometric position information of the test scene element, and the output is the position of the scene element in the camera imaging plane, so as to determine the distribution of the depth of field of different scene elements in the image plane; The scene elements include the traffic participant area, the sky area and the road area; (2) Fog image blur model construction The fog environment image blur model is composed of an incident light attenuation model and an atmospheric light imaging model, and its expression is: wherein, is a foggy image; is a pixel position coordinate; is a target object reflected light intensity when the depth of field is 0, also represents a non-foggy image; is a wavelength; is a fog transmittance; is an atmospheric light; is an atmospheric scattering coefficient; is a scene depth; (2.1) Fog transmittance calculation Based on the atmospheric attenuation model, the brightness contrast attenuation expression under the fog environment condition is: In the formula, L* is the luminance contrast; Visibility Relationship with atmospheric extinction coefficient of fog : Assuming that the fog model is a pure scattering fog, the absorption effect of the fog is ignored, and the relationship expression between the atmospheric visibility and the transmittance is obtained: In the formula, is the scene depth; is the transmittance; when the input needs to be simulated , the fog transmittance is obtained by formula (5) ; (2.2) Atmospheric light value calculation The global atmospheric light value is obtained by globally calculating the image, and the dark feature of the image is first calculated, and the calculation method expression is: In the formula, is the dark channel feature image to be obtained; is the input original image; is the R, G, B three channels where the image is located; is the average value of the first 0.1% of the pixel points in the image as the global atmospheric light value is the neighborhood centered on the pixel position of the point; is the neighborhood is the set of all pixel points in the neighborhood is the average value of the first 0.1% of the pixel points in the image as the global atmospheric light value : In the formula, is The position of the middle 0.1% of the pixel points; representing the corresponding The average value of the pixel intensity at the position is the global airlight value. The image information is divided into sky area, traffic participant area and road area by using semantic segmentation, and the corresponding atmospheric light is calculated to simulate the fog image; For the sky area, the bright channel is used for compensation, and the prior information of the bright channel image is obtained in the opposite way to the dark channel, that is, the maximum channel pixel value in the image three channels is used as the bright feature, and the expression is: To this end A neighborhood centered at the pixel position; select The average value of the top 0.1% of pixel values as the light feature output, expressed as: In the formula, is The position of the pixel point in the middle 0.1% of the front; representing the corresponding The pixel intensity of the position; the atmospheric light of the sky region is: In the formula, is the proportion of the total pixels occupied by the sky region, ranging between 0 and 1, which is determined by semantic segmentation; for the road region and the traffic participant region, the corresponding regional ambient light is calculated according to the dark feature calculation method, and the global ambient light value is calculated according to the pixel proportion output by the semantic segmentation is corrected. (3) Input the image blur model by using the atmospheric light value of the corresponding area combined with the fog transmittance, and output the fog-added image.

2. The foggy environment intelligent automobile camera perception test model according to claim 1, characterized in that: The depth of field distribution determination method of the traffic participant area, the sky area and the road area is as follows: For the traffic participant area, the centroid position of the target object is obtained through the geometric position information of the test scene element, the axis-aligned bounding box (AABB) of the traffic participant is determined by using the length, width and height of the target object, and the intersection relationship between the extension line of the pixel point and the camera optical center and the AABB of the target object in the scene is calculated in reverse, so as to determine the position of the traffic participant area in the camera imaging plane; If the extension line of the pixel point and the camera optical center does not intersect with the AABB of the target object in the scene, it is an invalid pixel; If the extension line of the pixel point and the camera optical center intersects with the AABB of the target object in the scene, it is a valid pixel; For the sky area, the depth of field of this part is set to a larger value compared with the traffic participant area, which is approximately infinite; For the road region, the depth of field of the road region is determined by using camera parameters, and the depth of field distribution of the road region is established by the installation parameters and built-in parameters of the camera; wherein the installation parameters of the camera include an installation angle , an installation height , and a horizontal distance between the camera and the front end of the vehicle , is an angle between the optical axis of the camera and a vertical plane; is a vertical height of the centroid of the camera from the ground; and the horizontal distance between the camera and the front end of the vehicle is denoted as ; for the built-in parameters of the camera, including a vertical field of view angle , and a camera focal length , according to the principle of triangle similarity and the camera transmission projection principle, the following is obtained: wherein represents the distance between the actual road and the camera, denotes the distance between the road region pixel and the lower image edge, represents the angle between the camera optical axis and the line connecting the projection of the target object in the imaging plane; From the above equation, the relationship between the depth of field of a road region and the expression is determined. and the depth of field of a road region.

3. A fog environment-oriented intelligent vehicle camera perception test method, which uses the fog environment intelligent vehicle camera perception test model of claim 1 or 2 for intelligent vehicle camera perception test, and the test task is to determine the functional boundary of the image perception algorithm, that is, to find the specific scene parameter combination that will cause the tested algorithm to fail; The test process is as follows: 1) Determine the image processing algorithm to be tested; 2) determine the logical scenario, i.e. determine the value range of the scene parameters, including the initial distance between the target vehicle and the host vehicle , the fog visibility to be simulated , and the uniform motion speed of the target object ; 3) Build a test scene by using a virtual simulation scene, collect image data of a no-fog scene parameter, and use the image to perform fog processing with a specific visibility by using the model construction method, and use the fog-added image to test the image perception algorithm to be tested; 4) record the confidence of the target recognition result corresponding to different visibility and the distance between the two vehicles, intersect it with the straight line with a confidence of 0.5, and record the Euclidean distance between the two vehicles corresponding to the first intersection point as , and As an evaluation index of the image recognition algorithm in the fog visibility scene, The smaller the value, the easier the image recognition algorithm is to fail in the fog environment. The physical meaning of is the distance between the recognition function of the to-be-tested algorithm for the target object and the boundary in the fog visibility.

Images