Air-ground target tracking performance evaluation method based on visual saliency

Abstract

The invention relates to the technical field of air-ground target tracking, in particular to an air-ground target tracking performance evaluation method based on visual saliency. Comprising the following steps: calculating a center positioning error between a target labeling frame and a prediction frame center point; calculating an adaptive center positioning error threshold based on the width and height of the target labeling box; calculating a precision index by using the threshold value; obtaining a saliency map of the target annotation box through a Bayesian framework, and calculating the weight of the target annotation box in combination with pixel saliency; the overlapping ratio under background suppression is calculated based on the overlapping region by using the significance weight, and pixel weight distribution of a real positive example, a false negative example and a false positive example is comprehensively considered in the overlapping ratio; and finally, drawing a success rate curve by setting different overlapping rate thresholds, and calculating an area under the curve as a model performance evaluation index. According to the method, the visual saliency and the adaptive threshold are combined, so that the accuracy and reliability of target tracking performance evaluation are effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of air-to-ground target tracking, and in particular to an air-to-ground target tracking performance evaluation method based on visual saliency. Background Art

[0002] Visual saliency refers to the mechanism by which humans automatically focus on areas of interest and ignore areas of no interest when facing a certain scene. The visual saliency mechanism is introduced in the field of computer vision. By imitating the human visual system to extract salient areas in images, it can tilt resources toward important information in the image, which is more in line with human visual cognitive needs. Therefore, visual saliency detection is widely used in target recognition and tracking, image segmentation, video compression, and image retrieval. Saliency detection based on Bayesian theory is a classic saliency detection method with the advantages of fast running speed and low computing power requirements.

[0003] In the field of air-to-ground target tracking, air-to-ground video sequences are prone to large differences in target size and axis-aligned annotation boxes containing a large amount of background due to the high dynamic changes in image distance and viewing angle. However, the diversity of target sizes in the test sequences and the large amount of background information contained in the axisymmetric annotation boxes seriously affect the adaptability and credibility of performance evaluation indicators based on center positioning error and overlap rate.

[0004] Since the target scale information is not taken into account, the accuracy index based on a fixed threshold cannot effectively adapt to the diversity of target sizes in air-ground video sequences, and thus cannot fully evaluate the performance of air-ground target tracking algorithms. Due to target irregularity and annotator cognitive bias, the axis-aligned target annotation box in the air-ground video sequence usually contains background information to varying degrees, which will lead to a decrease in the credibility of the AUC index based on the conventional overlap rate. Summary of the invention

[0005] The purpose of the present invention is to propose a method for evaluating the performance of air-to-ground target tracking based on visual saliency in response to the problems existing in the background technology.

[0006] The technical solution of the present invention is a method for evaluating the performance of air-to-ground target tracking based on visual saliency, comprising the following specific implementation steps:

[0007] S1, calculate the center positioning error;

[0008] S2, calculating the adaptive center positioning error threshold;

[0009] S3, calculating the accuracy index based on the adaptive threshold;

[0010] S4, obtaining the target annotation box weight based on visual saliency;

[0011] S5, calculating the overlap ratio under background suppression;

[0012] S6. Calculate the area under the curve to obtain the AUC index.

[0013] Preferably, the center positioning error is calculated by the Euclidean distance between the center point of the target annotation box and the predicted box to evaluate the target positioning deviation of the tracking algorithm:

[0014]

[0015] In the formula, (x 0 ,y 0 ) represents the target annotation box B in the open-ground video test sequence 0 Center point; (x 1 ,x 2 ) represents the target prediction box B of the air-ground target tracking algorithm to be evaluated 1 Center point; d CLE Indicates the center positioning error.

[0016] Preferably, the adaptive center positioning error threshold is dynamically determined according to the width and height of the target annotation box:

[0017]

[0018] Where h and w represent the height and width of the target annotation box respectively; θ a Represents the adaptive center positioning error threshold.

[0019] Preferably, based on the accuracy of the adaptive threshold, the adaptive center positioning error threshold θ is calculated by traversing each frame in the tracking sequence. a Tracking accuracy under:

[0020]

[0021] In the formula, θ a represents the adaptive center positioning error threshold; N represents the number of tracking sequence frames; i represents the tracking sequence image frame number; Pre represents the accuracy based on the adaptive center positioning error threshold; Represents the indicator function, which is used to determine whether the center positioning error e is less than the adaptive center positioning error threshold θ a ; e represents the center positioning error.

[0022] Preferably, the target annotation box weight uses a Bayesian model to combine the color distribution of the target and the background to generate a saliency map and assign a weight to the target annotation box:

[0023] S51, calculating a saliency map of the target annotation box based on a Bayesian framework;

[0024]

[0025] Where k is the color histogram partition; λ is the regularization parameter; Indicates the target annotation box area; Color histogram of the target label box area; Represents the background area around the target annotation box; Represents the color histogram of the background area around the target annotation box;

[0026] S52. Calculate the target annotation box weight Q∈R based on visual saliency w×h :

[0027] Q = P + 1;

[0028] Where P is the pixel saliency within the target annotation box Composition; R w×h Represents a matrix of size w×h, whose elements are natural numbers, R represents a natural number; h and w represent the height and width of the target annotation box respectively.

[0029] Preferably, the calculation process of the overlap ratio under background suppression is:

[0030]

[0031] Where TP represents the target annotation box B 0 and the target prediction box B 1 Overlapping part; FN indicates the part of the target annotation box that does not overlap with the target prediction box; FP indicates the part of the target prediction box that does not overlap with the target annotation box; q represents the overlap ratio; q ij (TP) and q ij (FN) represents the weight of the (i, j)th pixel in the target annotation box TP and FN area respectively; p ij (FP) represents the weight of the (i, j)th pixel in the FP area of ​​the target prediction box, and its value is 1.

[0032] Preferably, the calculation process of the AUC indicator is as follows:

[0033] S71. Calculate the success rate, that is, the percentage of frames in the test sequence in which the overlap rate between the target prediction box of the tracking algorithm and the target annotation box is greater than a given threshold:

[0034]

[0035] Where S represents the success rate; θ o is the overlap rate threshold; N is the number of tracking sequence frames; i is the tracking sequence image frame number; qi Indicates the overlap rate between the prediction box and the target annotation box of the i-th frame; Represents the indicator function, which is used to determine whether the overlap rate reaches the threshold θ o ;

[0036] S72. Draw a success rate curve by changing the set overlap rate threshold, and calculate the area under the curve to obtain the area under the curve, that is, the AUC indicator value.

[0037] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0038] This paper proposes a method for evaluating the performance of air-to-ground target tracking based on visual saliency. The performance of the air-to-ground target tracking algorithm is evaluated using the accuracy index based on adaptive threshold and the AUC index based on visual saliency:

[0039] (1) Improve the accuracy of target tracking performance evaluation: By combining the saliency weight of the target annotation box and fully considering the importance of the target area, the evaluation index can more accurately reflect the actual performance of the tracking algorithm;

[0040] (2) Enhance the robustness of the evaluation method: Adopt an adaptive center positioning error threshold and dynamically adjust the evaluation criteria to adapt to the scale changes of different targets and the complexity of tracking scenarios, thereby improving the robustness and universality of the evaluation;

[0041] (3) Comprehensive consideration of background interference: Through the overlap rate calculation method under background suppression, the interference effect of the background area is effectively eliminated, making the evaluation result closer to the actual tracking effect of the target;

[0042] (4) Multi-dimensional performance evaluation: Combining multiple indicators such as center positioning error (CLE), precision, overlap ratio, and area under the curve (AUC) to comprehensively measure the performance of the target tracking algorithm and ensure the comprehensiveness and scientificity of the evaluation results;

[0043] (5) Simplify target evaluation in complex scenes: The success rate calculation based on overlap rate and AUC curve analysis can intuitively reflect the performance of the tracking algorithm at different thresholds, making it easier to quantify the tracking effect in complex scenes;

[0044] (6) Improve practicality and applicability: By fully combining the characteristics of target saliency and tracking algorithms, the proposed method is applicable to a variety of practical scenarios of air-to-ground video target tracking, which helps promote the practical application of target tracking technology in air-to-ground scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A method flow chart of a method for evaluating air-to-ground target tracking performance based on visual saliency proposed by the present invention;

[0046] Figure 2 This is a schematic diagram of the center positioning error;

[0047] Figure 3 Schematic diagram of overlap ratio. DETAILED DESCRIPTION

[0048] Embodiment 1, as Figure 1 As shown, the present invention proposes a method for evaluating air-to-ground target tracking performance based on visual saliency, including the following specific implementation steps:

[0049] S1. Calculate the center location error (CLE);

[0050] S2, calculating the adaptive center positioning error threshold;

[0051] S3, calculating the precision (Precision, Pre) index based on the adaptive threshold;

[0052] S4, obtaining the target annotation box weight based on visual saliency;

[0053] S5, calculating the overlap ratio under background suppression;

[0054] S6. Calculate the area under the curve (AUC) indicator.

[0055] In an optional embodiment, the center location error (CLE) is used to evaluate the target location deviation of the tracking algorithm by calculating the Euclidean distance between the target annotation box and the center point of the prediction box:

[0056] like Figure 2 As shown, let P 0 (x 0 ,y 0 ) is the target annotation box B of the open space video test sequence 0 Center point, P 1 (x 1 ,x 2 ) is the target prediction box B of the air-ground target tracking algorithm to be evaluated 1 Center point, at this time the center positioning error d CLE The calculation is as follows:

[0057]

[0058] In an optional embodiment, an adaptive center positioning error threshold is used to dynamically determine the threshold according to the width and height of the target annotation box to enhance the adaptability of the evaluation:

[0059] Let the adaptive center positioning error threshold be θ a ,but:

[0060]

[0061] Among them, h and w represent the height and width of the target annotation box respectively.

[0062] In an optional embodiment, based on the precision (Precision, Pre) of the adaptive threshold, by traversing each frame in the tracking sequence, the adaptive center positioning error threshold θ is calculated. a Tracking accuracy under:

[0063]

[0064] In the formula, θ a represents the adaptive center positioning error threshold; N represents the number of tracking sequence frames; i represents the tracking sequence image frame number; Pre represents the accuracy based on the adaptive center positioning error threshold; Represents the indicator function, which is used to determine whether the center positioning error e is less than the adaptive center positioning error threshold θ a ; e represents the center positioning error.

[0065] In an optional embodiment, the target annotation box weight uses a Bayesian model to combine the color distribution of the target and the background to generate a saliency map and assign a dynamic weight to the target annotation box:

[0066] S41, calculating the saliency map of the target annotation box based on the Bayesian framework;

[0067] make represents the target annotation box area, The color histogram of the target label box area, Represents the background area around the target annotation box,

[0068] is the color histogram of the background area around the target annotation box. The significance of the pixel μ in the target annotation box is calculated using the Bayesian rule as follows:

[0069]

[0070] Where k is the color histogram partition and λ is the regularization parameter;

[0071] S42. Calculate the target annotation box weight Q∈R based on visual saliency w×h :

[0072] Q = P + 1;

[0073] Where P is the pixel saliency within the target annotation box Composition; R w×h Represents a matrix of size w×h, whose elements are natural numbers, R represents a natural number; h and w represent the height and width of the target annotation box respectively.

[0074] In an optional embodiment, if Figure 3 As shown in the figure, the overlap rate under background suppression integrates the weight information of the target and the predicted box area, evaluates their spatial overlap, and suppresses the background influence:

[0075]

[0076] In the formula, TP (True Positive) represents the target annotation box B 0 and the target prediction box B 1 Overlapping part; FN (False Negative) indicates the part of the target annotation box that does not overlap with the target prediction box; FP (False Positive) indicates the part of the target prediction box that does not overlap with the target annotation box; q represents the overlap ratio; q ij (TP) and q ij (FN) represents the weight of the (i, j)th pixel in the target annotation box TP and FN area respectively; p ij (FP) represents the weight of the (i, j)th pixel in the FP area of ​​the target prediction box, and its value is 1.

[0077] In an optional embodiment, the area under the curve (AUC) is calculated by plotting a curve graph of success rate versus threshold to quantify the overall performance of the tracking algorithm:

[0078] S61. The success rate is defined as the percentage of frames in the test sequence in which the overlap between the target prediction box of the tracking algorithm and the target annotation box is greater than a given threshold:

[0079]

[0080] Where S represents the success rate; θ o is the overlap rate threshold; N is the number of tracking sequence frames; i is the tracking sequence image frame number; qi Indicates the overlap rate between the prediction box and the target annotation box of the i-th frame; Represents the indicator function, which is used to determine whether the overlap rate reaches the threshold θ o :If O qi ≥θ o ,but It means that the i-th frame is successfully matched; if O qi <θ o ,but It means that the i-th frame has not been successfully matched;

[0081] S62. By setting the overlap rate threshold from 0 to 1, a success rate curve is drawn, and the area under the curve is calculated to obtain an AUC index value.

[0082] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for evaluating air-to-ground target tracking performance based on visual saliency, characterized in that: The specific implementation steps include the following: S1, calculate the center positioning error; S2, calculating the adaptive center positioning error threshold; S3, calculating the accuracy index based on the adaptive threshold; S4, obtaining the target annotation box weight based on visual saliency; S5, calculating the overlap ratio under background suppression; S6. Calculate the area under the curve to obtain the AUC index.

2. The method for evaluating air-to-ground target tracking performance based on visual saliency according to claim 1, characterized in that: The center positioning error evaluates the target positioning deviation of the tracking algorithm through the Euclidean distance between the target annotation box and the center point of the prediction box: Where (x0, y0) represents the center point of the target annotation box B0 of the open-ground video test sequence; (x1, x2) represents the center point of the target prediction box B1 of the open-ground target tracking algorithm to be evaluated; d CLE Indicates the center positioning error.

3. The method for evaluating air-to-ground target tracking performance based on visual saliency according to claim 1, characterized in that: The adaptive center positioning error threshold is dynamically determined according to the width and height of the target annotation box: Where h and w represent the height and width of the target annotation box respectively; θ a Represents the adaptive center positioning error threshold.

4. The method for evaluating air-to-ground target tracking performance based on visual saliency according to claim 1, characterized in that: Based on the accuracy of the adaptive threshold, the adaptive center positioning error threshold θ is calculated by traversing each frame in the tracking sequence. a Tracking accuracy under: In the formula, θ a represents the adaptive center positioning error threshold; N represents the number of tracking sequence frames; i represents the tracking sequence image frame number; Pre represents the accuracy based on the adaptive center positioning error threshold; Represents the indicator function, which is used to determine whether the center positioning error e is less than the adaptive center positioning error threshold θ a ; e represents the center positioning error.

5. The method for evaluating air-to-ground target tracking performance based on visual saliency according to claim 1, characterized in that: The target annotation box weight uses the Bayesian model to combine the color distribution of the target and background to generate a saliency map and assign a weight to the target annotation box: S51, calculating a saliency map of the target annotation box based on a Bayesian framework; Where k is the color histogram partition; λ is the regularization parameter; Indicates the target annotation box area; Color histogram of the target label box area; Represents the background area around the target annotation box; Represents the color histogram of the background area around the target annotation box; S52. Calculate the target annotation box weight Q∈R based on visual saliency w×h : Q = P + 1; Where P is the pixel saliency within the target annotation box Composition; R w×h Represents a matrix of size w×h, whose elements are natural numbers, R represents a natural number; h and w represent the height and width of the target annotation box respectively.

6. The method for evaluating air-to-ground target tracking performance based on visual saliency according to claim 1, characterized in that: The calculation process of the overlap rate under background suppression is: Where TP represents the overlapped part of the target annotation box B0 and the target prediction box B1; FN represents the part of the target annotation box that does not overlap with the target prediction box; FP represents the part of the target prediction box that does not overlap with the target annotation box; q represents the overlap ratio; q ij (TP) and q ij (FN) represents the weight of the (i, j)th pixel in the target annotation box TP and FN area respectively; p ij (FP) represents the weight of the (i, j)th pixel in the FP area of ​​the target prediction box, and its value is 1.

7. The method for evaluating air-to-ground target tracking performance based on visual saliency according to claim 1, characterized in that: The calculation process of the AUC indicator is as follows: S71. Calculate the success rate, that is, the percentage of frames in the test sequence in which the overlap rate between the target prediction box of the tracking algorithm and the target annotation box is greater than a given threshold: Where S represents the success rate; θ o is the overlap rate threshold; N is the number of tracking sequence frames; i is the tracking sequence image frame number; qi Indicates the overlap rate between the prediction box and the target annotation box of the i-th frame; Represents the indicator function, which is used to determine whether the overlap rate reaches the threshold θ o ; S72. Draw a success rate curve by changing the set overlap rate threshold, and calculate the area under the curve to obtain the area under the curve, that is, the AUC indicator value.

Images