Target detection performance adaptive measurement method in new scene

Abstract

The invention discloses a target detection performance adaptive measurement method in a new scene. The method comprises the following steps: fusing illumination, color cast and divergence estimation to measure background difference (S1), and calculating statistics of background difference indexes of a training scene and an application scene (S1-1); on this basis, the training scene is used as a normal scene, and the application scene is classified into a normal scene, a high dynamic scene, a strong light scene, a low illumination scene, a color cast scene and a low visibility scene, which are used as measures of background differences between the scenes (S1-3). Measuring a target difference by integrating the tag offset and the feature offset (S2); wherein a label offset result is calculated by means of a KS test (S2-1); using the target feature entropy of the image detection results in the training scene and the application scene to calculate a feature offset result (S2-2); and taking as a measurement result of the target difference. Different from an existing method, the method not only can judge whether domain differences exist among the scenes, but also can analyze the types of the domain differences of the scenes in a fine-grained manner. The performance of a target detection algorithm in a new scene can be improved.

Description

Technical Field

[0001] The invention relates to an adaptive measurement method for target detection performance in a new scene, and belongs to the technical field of digital image processing. Background Art

[0002] With the widespread application of video surveillance, a huge amount of video data is generated every day. Video surveillance systems are gradually developing towards intelligence. For example, high-level computer vision tasks such as target detection are playing an increasingly important role as artificial assistance. The target detection algorithm is an algorithm for obtaining the size, position and category of the target in the image. Most of the existing target detection algorithms use deep learning algorithms. However, deep learning algorithms have strong scene dependence. When the trained deep learning model is applied to a new scene, the performance often degrades. This phenomenon is called domain migration. Domain migration is particularly prominent in video surveillance application scenarios, because the complexity and diversity of monitoring scenes make the data distribution of training scenes and application scenes have a large difference. This limits the effective application scenarios of intelligent algorithms. Therefore, people are committed to improving the generalization ability of the model through technical means such as domain adaptation and transfer learning, thereby improving the performance of deep learning algorithms in new scenarios.

[0003] Existing domain adaptation methods include domain adaptation based on adversarial feature learning, domain adaptation based on pseudo-label self-training, and domain adaptation based on image translation. Domain adaptation based on adversarial feature learning uses a domain discriminator to perform adversarial training on the target detection model, so that the model can learn domain invariant features. Domain adaptation based on pseudo-label self-training uses the high-confidence prediction results generated in the target domain by the detector model trained in the source domain to train the model of the target domain. Domain adaptation based on image translation uses an unpaired image conversion model to map the target image to an image space similar to the source image, reducing the distribution shift of the image style. Since the detector is more likely to perform well on target images similar to the source domain, the performance of the model in the target domain is improved.

[0004] It can be seen that the existing domain adaptation methods mainly adjust the deep learning algorithm model to adapt the model to the new scenario, but the measurement of the domain difference between the training scenario and the application scenario is not considered enough. When there is no domain difference between the training scenario and the application scenario, there is no need to use the domain adaptation method to adjust the model. In addition, reducing the domain difference should be an optimization goal of model adjustment or parameter adaptive adjustment, so the accurate measurement of domain differences between scenarios is an important issue. Summary of the invention

[0005] Under the above background of the prior art, the present invention considers the problem of accurate measurement of domain differences between scenes, takes monitoring scenes as the research object, and faces the target detection visual task, decouples the domain differences between scenes into background differences and target differences, and measures the background differences and target differences between scenes respectively, and then proposes a method for measuring domain differences between scenes for target detection, which realizes the fine and accurate measurement of scene domain differences, and plays a guiding role for subsequent tasks such as camera parameter adaptive adjustment or domain adaptation. The method of the present invention is of great significance for intelligent monitoring systems.

[0006] According to one aspect of the present invention, a method for measuring domain differences between scenes for target detection is proposed, which decouples domain differences between scenes into background differences and target differences. Based on the differences in background distribution between different scenes, a background domain difference measurement model is established by integrating illumination, color cast, and scattering estimation. Based on the differences in target distribution between different scenes, a target domain difference measurement model is established by integrating label offset and feature offset. The above two models are combined to achieve accurate measurement and analysis of domain differences between scenes.

[0007] According to a further aspect of the present invention, a method for adaptively measuring target detection performance in a new scenario is provided, which is characterized by comprising:

[0008] A) In order to measure the background differences between scenes, the images in the training scene A are estimated for illumination, color shift and scattering offline to obtain the statistics of the proportion of bright and dark pixels, the color shift level and the scattering level. Then, the images in the application scene B are estimated for illumination, color shift and scattering online to obtain the proportion of bright and dark pixels, the color shift level and the scattering level. The statistics of the indicators of the application scene B are compared with those of the indicators of the training scene A. The training scene is regarded as a normal scene, and the application scenes are classified into normal scenes, high dynamic scenes, strong light scenes, low illumination scenes, color shift scenes and low visibility scenes. The classification results of the application scenes are used as a measure of the background differences between scenes, including:

[0009] A1) Perform illumination estimation, color cast estimation and scatter estimation on all images in training scene A, and obtain statistics of illumination estimation (including mean dark pixel ratio μ s，d , Variance of dark pixel proportion σ s，d , mean value of bright pixel ratio μ s，l , Bright pixel ratio variance σ s，l ), statistics of color cast estimation (including the color cast level mean μ s，ccb ) and statistics of scatter estimates (including the mean of the scatter level μ s，s and the variance of the scattering level σ s，s ).

[0010] A2) Perform illumination estimation, color shift estimation and scatter estimation on multiple frames of images acquired in real time in application scenario B, and obtain statistics of illumination estimation (including mean value of dark pixel ratio μ t，d and the mean value of bright pixel ratio μ t，l ), statistics of color cast estimation (mean value of color cast level μ t，ccb ) and the scatter estimate statistic (the mean of the scatter level μ t，s ).

[0011] A3) Comprehensively consider the illumination estimation statistics, color cast estimation statistics and scattering estimation statistics of the training scene A and the application scene B, and refer to Raida's law to classify the application scenes into normal scenes, high dynamic scenes, strong light scenes, low illumination scenes, color cast scenes and low visibility scenes. The classification results of the application scenes are used as a measure of the background differences between scenes.

[0012] B) In order to measure the target difference between scenarios, the label shift result is calculated based on the performance label distribution of training scenario A and application scenario B. If label shift occurs in application scenario B, the feature shift result is calculated based on the statistical results of the target feature entropy of training scenario A and application scenario B. The measurement result of the target difference between training scenario A and application scenario B is obtained by combining the label shift and feature shift results, including:

[0013] B1) The mean confidence level of the image detection results is used as the image label, and the distribution of labels for the training scenario and the application scenario is calculated respectively. The Kolmogorov-Smirnov test is used to determine whether label shift occurs.

[0014] B2) If label shift occurs, calculate the mean feature entropy μ of the target in training scene A offline s，ie , and online calculate the mean value μ of the target feature entropy of application scenario B t，ie ,Based on the outlier processing method in statistics, it is judged whether the information of the target obtained in the application scenario is insufficient, which serves as the basis for feature shift.

[0015] B3) Integrate the results of label shift and feature shift to determine whether the information of the target in the current scene is sufficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 FIG. 4 is a schematic diagram of a background difference measurement method according to an embodiment of the present invention.

[0017] Figure 2 is a schematic diagram of target difference measurement according to an embodiment of the present invention.

[0018] Figure 3It is a schematic diagram showing that the comparison method (weather adaptive judgment) according to an embodiment of the present invention misclassifies low visibility weather as low illumination weather in a test image.

[0019] Figure 4 (a) to Figure 4 (f) is the experimental result of the domain difference measurement method for target detection according to an embodiment of the present invention on the test data set, indicating that the method can accurately measure the domain difference and subdivide the scene into low illumination scene, strong light scene, high dynamic scene, color cast scene, low visibility scene and scene with insufficient target information. DETAILED DESCRIPTION

[0020] According to one aspect of the present invention, a method for adaptively measuring target detection performance in a new scenario is provided. The method of the present invention realizes accurate measurement of domain differences between scenes, and the domain difference measurement results between scenes obtained by the method are more refined, which can not only determine whether there are domain differences, but also further subdivide them into background differences and target differences. The background difference measurement results will subdivide the scene, and the target difference measurement results include label offset and feature offset results.

[0021] In order to achieve accurate measurement of domain differences between scenes, the present invention proposes an adaptive measurement method for target detection performance in a new scene, which is characterized by comprising:

[0022] A) Figure 1 The background difference measurement model constructed by the present invention is used to measure the background difference between scenes. The illumination estimation, color deviation estimation and scattering estimation are performed offline on the image in the training scene A to obtain the statistics of the proportion of bright and dark pixels, the color deviation level and the scattering level. Then, the illumination estimation, color deviation estimation and scattering estimation are performed online on the image in the application scene B to obtain the proportion of bright and dark pixels, the color deviation level and the scattering level. The statistics of the indicators of the application scene B and the indicators of the training scene A are compared. The training scene is taken as a normal scene, and the application scenes are classified into normal scenes, high dynamic scenes, strong light scenes, low illumination scenes, color deviation scenes and low visibility scenes. The classification results of the application scenes are used as the measurement of the background difference between the scenes, including:

[0023] A1) Perform illumination estimation, color cast estimation and scatter estimation on all images in training scene A, and obtain statistics of illumination estimation (including mean dark pixel ratio μ s，d , Variance of dark pixel proportion σ s，d , mean value of bright pixel ratio μ s，l , Bright pixel ratio variance σ s，l ), statistics of color cast estimation (including the color cast level mean μ s，ccb ) and statistics of scatter estimates (including the mean of the scatter level μ s，s and the variance of the scattering level σs，s ).

[0024] A2) Perform illumination estimation, color shift estimation and scatter estimation on multiple frames of images acquired in real time in application scenario B, and obtain statistics of illumination estimation (including mean value of dark pixel ratio μ t，d and the mean value of bright pixel ratio μ t，l ), statistics of color cast estimation (mean value of color cast level μ t，ccb ) and the scatter estimate statistic (the mean of the scatter level μ t，s ).

[0025] A3) Comprehensively consider the illumination estimation statistics, color cast estimation statistics and scattering estimation statistics of the training scene A and the application scene B, and refer to Raida's law to classify the application scenes into normal scenes, high dynamic scenes, strong light scenes, low illumination scenes, color cast scenes and low visibility scenes. The classification results of the application scenes are used as a measure of the background differences between scenes.

[0026] B) Figure 2 It is a target difference measurement model constructed by the present invention. In order to measure the target difference between scenarios, the label shift result is calculated based on the performance label distribution of the training scenario A and the application scenario B. If the application scenario B has a label shift, the feature shift result is calculated based on the statistical results of the target feature entropy of the training scenario A and the application scenario B. The measurement result of the target difference between the training scenario A and the application scenario B is obtained by combining the label shift and feature shift results, including:

[0027] B1) The mean confidence level of the image detection results is used as the image label, and the distribution of labels for the training scenario and the application scenario is calculated respectively. The Kolmogorov-Smirnov test is used to determine whether label shift occurs.

[0028] B2) If label shift occurs, calculate the mean feature entropy μ of the target in training scene A offline s，ie , and online calculate the mean value μ of the target feature entropy of application scenario B t，ie ,Based on the outlier processing method in statistics, it is judged whether the information of the target obtained in the application scenario is insufficient, which serves as the basis for feature shift.

[0029] B3) Integrate the results of label shift and feature shift to determine whether the information of the target in the current scene is sufficient.

[0030] According to a further embodiment of the present invention, in the above step A), the illumination estimation method adopted is illumination estimation based on the Retinex model, wherein the maximum brightness value in a larger neighborhood is first taken as the initial result, and then similar to the Retinex method, constraints including pixel values ​​and gradients are introduced as the objective function, and the final illumination feature map is solved in an iterative manner, and finally the proportion of bright and dark pixels is counted as an indicator of image illumination estimation, that is:

[0031] L (0) =max(I),E(L)=αE pix (L)+βE grad (L)

[0032] (1),

[0033]

[0034]

[0035] Where L (0) represents the initial illumination map, N represents the total number of pixels, L (k) represents the illumination feature map after the kth iteration, represents the total number of pixels, I represents the image in the scene, α and β represent the hyperparameters set by the user, E(L), E pix (L) and E grad (L) represent the overall objective function, pixel value objective function and gradient objective function respectively, L (k+1) represents the illumination feature map after the k+1th iteration, represents the brightness value at position i, θ dark and θ light represents the set threshold, and M dark and M light Represents the ratio of dark and bright pixels, which is used as an indicator of illumination estimation.

[0036] According to a further embodiment of the present invention, in the above step A), the color cast estimation method adopted is a color cast estimation based on a channel color cast index, and a channel color cast index proposed based on the grayscale world hypothesis is used to estimate the color cast of an image in a scene. The channel color cast index extracts the brightness information of the image pixel by pixel and calculates its channel color cast level.

[0037] C=E i -E j , i, j∈{R, G, B} (2),

[0038] M ccb =||C||1

[0039] Among them, ||||1 represents the vector norm, i and j represent the channel with the highest mean and the channel with the lowest mean in the RGB channel of the image respectively, and E i and E j Indicates the pixel value of the corresponding channel, the channel color cast level M ccb Represents the sum of the differences between the channel with the largest mean and the channel with the smallest mean,

[0040] According to a further embodiment of the present invention, in the background difference measurement method in the above step A), the definition of a strong light scene is: compared with a normal scene, a strong light scene means that the brightness of the acquired image is generally larger.

[0041] μ t，l >μ s，l +σ s，l (3),

[0042] μ t,l Represents the average proportion of bright pixels in the application scenario, μ s,l represents the mean of the proportion of bright pixels in the training scene, σ s,l Represents the variance of the proportion of bright pixels in the training scene.

[0043] According to a further embodiment of the present invention, in the background difference measurement method in the above step A), the definition of a low-illuminance scene is: compared with a normal scene, a low-illuminance scene means that the brightness of the acquired image is overall low and dark pixels account for a large proportion.

[0044] μ t，d >μ s，d +σ s，d

[0045] (4)

[0046] μ t，d Represents the average proportion of dark pixels in the application scene, μ s，d represents the mean of the proportion of dark pixels in the training scene, σ s，d Represents the variance of the proportion of dark pixels in the training scene.

[0047] According to a further embodiment of the present invention, in the background difference measurement method in the above step A), the definition of a high dynamic scene is: compared with a normal scene, a high dynamic scene means that the brightness of the acquired image is distributed on the bright side and the dark side.

[0048] μ t，l >μ s，l +σ s，l

[0049] μ t，d >μ s，d +σ s，d

[0050] (5)

[0051] μ t，d Represents the average proportion of dark pixels in the application scene, μ s，d represents the mean of the proportion of dark pixels in the training scene, σ s，d represents the variance of the proportion of dark pixels in the training scene, σ s，l Represents the variance of the proportion of bright pixels in the training scene.

[0052] According to a further embodiment of the present invention, in the background difference measurement method in the above step A), the definition of a color cast scene is: a color cast scene means that the color cast level of the acquired image is generally high.

[0053] μ t，ccb >μ s，ccb

[0054] (6)

[0055] μ t，ccb Represents the average color cast level in the application scenario, μ s，ccb Represents the mean color cast level in the training scenes.

[0056] According to a further embodiment of the present invention, in the background difference measurement method in the above step A), the definition of a low-visibility scene is: compared with a normal scene, a low-visibility scene means that the scattering level of the acquired image is generally higher.

[0057] μ t，s >μ s，s +σ s，s (7)

[0058] μ t，s Represents the mean value of the scattering level in the application scenario, μ s，s represents the mean of the scattering level in the training scene, σ s，s Represents the variance of the scattering level in the training scenes.

[0059] According to a further embodiment of the present invention, in the above step A), the color shift estimation method adopted is a scattering estimation based on a CEP ellipsoid prior, and the image transmittance feature map is estimated by the CEP ellipsoid prior, 12 scattering-related features are extracted from the image transmittance feature map, and these features are fitted into a multivariate Gaussian (MVG) model, and then, the Mahalanobis distance method is used to calculate the deviation between the features of the test image and the features of 500 fog-free images and another 500 foggy images, and each corresponding distance is defined as a foggy scattering level and a fog-free scattering level, respectively. Finally, the scattering level D of the calculated image is expressed as the ratio of the foggy scattering level to the fog-free scattering level.

[0060]

[0061] Where t(x) represents the transmittance calculated pixel by pixel, Ω(x) represents the size of the divided area, and I c (y) represents the pixel value of a specific channel, A c represents the global atmospheric composition of a particular channel, D represents the scattering level, and D f Expressed as fog scattering level, D ff Indicates the haze-free scattering level,

[0062] According to a further embodiment of the present invention, the model is first applied to the training scene and the target scene respectively, and the target confidence mean in the image detection result is used as the image label to obtain the label distribution of the source domain and the target domain. Then, the KS test is performed to determine whether there is a label offset between the source domain and the target domain.

[0063]

[0064]

[0065]

[0066] Where F represents the object detection model trained in the source domain, represents the mean of the target confidence of the inference result of F for the i-th image in the training scene, represents the mean value of the target confidence of the inference result of F for the i-th image in the application scenario, D l represents the statistic of the KS test after sampling l images from the training scene and the application scene, p represents the threshold for judging whether the two distributions are consistent,

[0067] According to a further embodiment of the present invention, in the above step B), the method for calculating feature offset is as follows: first, the model is applied to the training scene and the target scene respectively, the features in the target backbone network in the image detection result are calculated, and then the information entropy of the target features is calculated to obtain the feature distribution of the source domain and the target domain, and then the statistics (mean and variance) of the feature distribution in the source domain are calculated. For the feature distribution in the target domain, the processing of outliers in statistics is used to determine whether the feature information of the target in the scene is sufficient, which serves as the basis for subsequent triggering of camera parameter adjustment.

[0068] μ t，ie ≤μ s，ie (1O),

[0069] μ t,ie represents the mean value of the feature information entropy distribution in the target scene, μ s,ie Represents the mean of the information entropy distribution of the training scene features.

[0070] Therefore, the present invention proposes an adaptive measurement method for target detection performance in new scenarios. The domain difference measurement method decouples domain differences into background differences and target differences. Based on the background distribution differences between different scenes, the background domain difference measurement model is established by integrating illumination, color cast and scattering estimation; based on the distribution differences of targets between different scenes, the target domain difference measurement model is established by integrating label offset and feature offset. Combining the above two models, accurate measurement and analysis of domain differences between scenes are achieved. Therefore, the method of the present invention can be used as a front module of camera parameter adaptive adjustment or domain adaptive method, provide guidance for camera parameter adaptive adjustment or domain adaptive, and thus improve the performance of target detection algorithm in new scenarios.

[0071] The adaptive measurement method for target detection performance in new scenarios provided by the present invention constitutes a substantial and significant improvement over the existing domain difference measurement method. Compared with the prior art, the beneficial effect of the present invention is reflected in the accurate and fine measurement of domain differences between scenes. As a pre-module for camera parameter adaptive adjustment or domain adaptive adjustment, the two work together to improve the performance of target detection algorithms in new scenarios.

[0072] Figure 3 It is a schematic diagram showing that the comparison method (weather adaptive judgment) according to an embodiment of the present invention misclassifies low visibility weather as low illumination weather in a test image. Figure 4 The experimental results of the domain difference measurement method of the present invention are shown in Figure 4 (a) to Figure 4 (f) The original images shown are all in color). Adaptive weather judgment uses manually set thresholds to judge whether the current scene is a normal scene, a low-illumination scene, or a low-visibility scene. Some foggy images will be judged as low-illumination images. In addition, the use of manually set thresholds relies on human experience and cannot achieve adaptive judgment of any scene. In contrast, the domain difference measurement method proposed in the present invention decouples the measurement of domain differences from the background and target perspectives, and refers to the statistical measurement indicators of the source domain. It can well judge whether there is a domain offset between the source scene and the target scene, and divides the domain differences in a fine-grained manner, thereby achieving accurate measurement of domain differences.

[0073] It should be noted that the above disclosure is only a specific implementation example of the present invention. According to the technical ideas provided by the present invention, changes that can be conceived by ordinary technicians in the field should fall within the protection scope of the present invention.

Claims

1. A method for adaptively measuring target detection performance in a new scenario, characterized in that include: A) In order to measure the background differences between scenes, the images in the training scene A are estimated for illumination, color shift and scattering offline to obtain the statistics of the proportion of bright and dark pixels, the color shift level and the scattering level. Then, the images in the application scene B are estimated for illumination, color shift and scattering online to obtain the proportion of bright and dark pixels, the color shift level and the scattering level. The statistics of the indicators of the application scene B are compared with those of the indicators of the training scene A. The training scene is regarded as a normal scene, and the application scenes are classified into normal scenes, high dynamic scenes, strong light scenes, low illumination scenes, color shift scenes and low visibility scenes. The classification results of the application scenes are used as a measure of the background differences between the scenes, including: A1) Perform illumination estimation, color cast estimation and scatter estimation on all images in training scene A, and obtain: Statistics of the first illumination estimate, including the mean of the dark pixel ratio μ s,d , Variance of dark pixel proportion v s,d , the average value of bright pixel ratio v s,l , Bright pixel ratio variance σ s,l , The first color shift estimation statistic, including the color shift level mean μ s,ccb ,and The first scatter estimation statistic, including the mean of the scatter levels μ s,s and the variance of the scatter level v s,s , A2) Illumination estimation, color shift estimation and scattering estimation are performed on multiple frames of images acquired in real time in application scenario B to obtain: The second illumination estimation statistic includes the mean of the dark pixel ratio μ t，d and the mean value of bright pixel ratio μ t，l , The second color shift estimation statistic includes the color shift level mean μ t,ccb ,and Second scatter estimation statistics, including the mean of the scatter levels μ t,s , A3) combining the first and second illumination estimation statistics, the first and second color shift estimation statistics, and the first and second scattering estimation statistics, and referring to Raida's law, classifying the application scenes into normal scenes, high dynamic scenes, strong light scenes, low illumination scenes, color shift scenes, and low visibility scenes, and using the classification results of the application scenes as a measure of the background differences between the scenes, B) In order to measure the target difference between scenarios, the label shift is calculated based on the performance label distribution of training scenario A and application scenario B. When label shift occurs in application scenario B, the result of feature shift is calculated based on the statistical results of the target feature entropy of training scenario A and application scenario B. The results of label shift and feature shift are combined to obtain the measurement result of the target difference between training scenario A and application scenario B, including: B1) Applying the target detection algorithm to the image obtained in the scene to obtain the location of the target in the scene and the confidence of the target type as the image detection result, The mean confidence level of the image detection results is used as the image label, and the distribution of labels for the training scenario and the application scenario is calculated respectively. The Kolmogorov-Smirnov test is used to determine whether label shift occurs. B2) When label shift occurs, the mean value μ of the feature entropy of the target in training scene A is calculated offline s，ie , and online calculate the mean value μ of the target feature entropy of application scenario B t，ie , based on the outlier processing method in statistics, it is judged whether the target information obtained in the application scenario is insufficient, which serves as the basis for feature shift. B3) Comprehensively analyze the results of label shift and feature shift to determine whether the information of the target in the current scene is sufficient. in: The above illumination estimation includes illumination estimation based on the Retinex model, including: firstly taking the maximum brightness in a larger neighborhood as the initial result, and then similar to the Retinex method, introducing constraints including pixel values ​​and gradients as the objective function, solving the final illumination feature map in an iterative manner, and finally counting the proportion of bright and dark pixels as an indicator of image illumination estimation, namely: Where L (0) represents the initial illumination map, N represents the total number of pixels, L (k) represents the illumination feature map after the kth iteration, represents the total number of pixels, I represents the image in the scene, α and β represent the hyperparameters set by the user, E(l), E pix (L) and E grad (L) represent the overall objective function, pixel value objective function and gradient objective function respectively, L (k+1) represents the illumination feature map after the k+1th iteration, represents the brightness value at position i, θ dark and θ lig represents the set threshold, and M dark and M light Represents the ratio of dark and bright pixels, which is used as an indicator of illumination estimation. Color cast estimation is based on channel color cast indicators, including using channel color cast indicators based on the grayscale world assumption to estimate the color cast of images in the scene. The channel color cast indicator extracts the brightness information of the image pixel by pixel and calculates its channel color cast level. Among them, ||||1 represents the vector norm, i and j represent the channel with the highest mean and the channel with the lowest mean in the RGB channel of the image respectively, and E i and E j Indicates the pixel value of the corresponding channel, the channel color cast level M ccb Represents the sum of the differences between the channel with the largest mean and the channel with the smallest mean, In the above-mentioned measurement of background differences between scenes, the definition of a bright light scene is: compared with a normal scene, a bright light scene means that the brightness of the acquired image is generally larger. m t,l >m s,l +s s,l (3), μ t,l Represents the average proportion of bright pixels in the application scenario, μ s,l represents the mean of the proportion of bright pixels in the training scene, σ s,l represents the variance of the proportion of bright pixels in the training scene, In the measurement of background differences between the above scenes, the definition of low-light scenes is: compared with normal scenes, low-light scenes mean that the brightness of the acquired image is lower overall, and the proportion of dark pixels is larger. m t,d >m s,d +s s,d (4), μ t,d Represents the average proportion of dark pixels in the application scene, μ s,d represents the mean of the proportion of dark pixels in the training scene, σ s,d represents the variance of the proportion of dark pixels in the training scene, In the above-mentioned measurement of background differences between scenes, the definition of a high dynamic scene is: compared with a normal scene, a high dynamic scene means that the brightness of the acquired image is distributed on the bright side and the dark side. m t,l >m s,l +s s,l m t,d >m s,d +s s,d (5), μ t,d Represents the average proportion of dark pixels in the application scene, μ s,d represents the mean of the proportion of dark pixels in the training scene, σ s,d represents the variance of the proportion of dark pixels in the training scene, σ s,l represents the variance of the proportion of bright pixels in the training scene, In the above-mentioned measurement of background differences between scenes, the definition of color cast scene is: color cast scene means that the color cast level of the acquired image is generally high. m t,ccb >m s,ccb (6), μ t,ccb Represents the average color cast level in the application scenario, μ s,ccb represents the mean color cast level in the training scene, In the above-mentioned measurement of background differences between scenes, the definition of low-visibility scenes is: compared with normal scenes, low-visibility scenes mean that the scattering level of the acquired images is generally higher. m t,s >m s,s +s s,s (7), μ t,s Represents the mean value of the scattering level in the application scenario, μ s,s represents the mean of the scattering level in the training scene, σ s,s Represents the variance of the scattering level in the training scenes.

2. The method for adaptively measuring target detection performance in a new scenario according to claim 1, characterized in that The color shift estimation is based on the scattering estimation of the CEP ellipsoid prior, including: The image transmittance feature map is estimated by CEP ellipsoid prior, 12 scattering-related features are extracted from the image transmittance feature map, and these features are fitted into a multivariate Gaussian model. The Mahalanobis distance method is used to calculate the deviation between the features of the test image and the features of 500 fog-free images and another 500 foggy images. Each corresponding distance is defined as the foggy scattering level and the fog-free scattering level, respectively. Then, the scattering level D of the calculated image is expressed as the ratio of the scattering level with fog to the scattering level without fog: Where t(x) represents the transmittance calculated pixel by pixel, Ω(x) represents the size of the divided area, and I c (y) represents the pixel value of a specific channel, A c represents the global atmospheric composition of a particular channel, D represents the scattering level, and D f Expressed as fog scattering level, D ff Indicates the fog-free scattering level.

3. The method for adaptively measuring target detection performance in a new scenario according to claim 1, characterized in that: The calculated label offsets used in step B) include: First, the object detection model obtained in the training scene is applied to the training scene and the target scene respectively. The target confidence mean in the image detection result is used as the image label to obtain the label distribution of the source domain and the target domain. Then the KS test is used to determine whether there is a label offset between the source domain and the target domain: Where F represents the object detection model trained in the source domain, represents the mean of the target confidence of the inference result of F for the i-th image in the training scene, represents the mean value of the target confidence of the inference result of F for the i-th image in the application scenario, D l It represents the statistic of KS test after sampling l images from the training scene and the application scene, and p represents the threshold for judging the consistency of the two distributions.

4. The method for adaptively measuring target detection performance in a new scenario according to claim 1, characterized in that The above process of calculating feature offset includes: First, the model is applied to the training scene and the target scene respectively, and the features in the target backbone network in the image detection results are calculated. Then the information entropy of the target feature is calculated to obtain the feature distribution of the source domain and the target domain. Then, the statistics of the feature distribution in the source domain are calculated. For the feature distribution in the target domain, the outlier processing in statistics is used to determine whether the feature information of the target in the scene is sufficient, which serves as the basis for the subsequent triggering of camera parameter adjustment. μ t,ie ≤μ s,ie (10), μ t,ie represents the mean value of the feature information entropy distribution in the target scene, μ s,ie It represents the mean of the information entropy distribution of the training scene features, and the statistics include the mean and variance.

Images