Target detection result suppression method based on probability density function

Abstract

The application provides a target detection result suppression method based on a probability density function, comprising: obtaining at least two target detection results; converting the target detection results into probability density functions based on Gaussian distribution; calculating the relative distances of the target detection results according to the probability density functions; determining that the relative distances of two target detection results are less than a set threshold value, and then performing non-maximum suppression based on confidence for an overlapping result set composed of at least two target detection results with the relative distances less than the set threshold value to obtain the target detection result with the maximum confidence as a target detection conclusion. Compared with a traditional intersection over union method, the application effectively reduces the complexity of a calculation model, converts various possible geometric problems into algebraic problems with a general solution, can quickly determine whether multiple candidate frames are the same detection target, and thus effectively improves the efficiency of target detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and in particular to a target detection result suppression method based on a probability density function. BACKGROUND

[0002] The task of object detection is to find all objects of interest in an image (or point cloud data), determine their categories and positions, and is one of the core problems in the field of deep learning. Due to different appearances, shapes and poses of various objects, and the interference of factors such as light, occlusion (point cloud noise) during imaging, object detection has always been the most challenging problem in the field of deep learning.

[0003] Object detection produces a large number of candidate boxes at the same target position, and the candidate boxes overlap with each other. How to determine whether multiple candidate boxes are the same detection target is a problem that must be solved in the task of object detection.

[0004] The prior art often uses the intersection over union method to make the determination, which requires calculating the overlapping area of the candidate boxes. Considering the differences in size, shape, angle and overlapping degree of the candidate boxes, the complexity of this calculation process is high, which is not conducive to improving the efficiency of object detection.

[0005] Therefore, how to provide a method for quickly determining whether multiple candidate boxes are the same detection target has become a problem that needs to be solved in the industry. SUMMARY

[0006] The present application provides a target detection result suppression method based on a probability density function, which solves the defect that the complexity of the calculation process in the prior art is high and is not conducive to improving the efficiency of object detection, and realizes the quick determination of whether multiple candidate boxes are the same detection target.

[0007] The present application provides a target detection result suppression method based on a probability density function, which includes:

[0008] Obtaining at least two target detection results;

[0009] Converting the target detection results into a probability density function based on a Gaussian distribution;

[0010] According to the probability density function, the relative distance of the target detection results is calculated;

[0011] If it is determined that the relative distance of two target detection results is less than a set threshold, then for the overlapping result set composed of at least two target detection results whose relative distance is less than the set threshold, non-maximum suppression based on confidence is performed, and the target detection result with the maximum confidence is obtained as the target detection conclusion.

[0012] The application provides a target detection result suppression method based on a probability density function, and the target detection result is a two-dimensional rectangular detection frame.

[0013] The step of converting the target detection result into the probability density function based on the Gaussian distribution comprises the following steps:

[0014] The center point of the two-dimensional rectangular detection frame is taken as an origin, the height of the two-dimensional rectangular detection frame is taken as a long axis length, and the width of the two-dimensional rectangular detection frame is taken as a short axis length, so that a target detection ellipse is constructed.

[0015] The probability density function Σ is constructed based on a two-dimensional Gaussian distribution for a set of points in the target detection ellipse. 1 / 2 :

[0016]

[0017] In the formula, θ is a rotation angle of the two-dimensional rectangular detection frame, w is the width of the two-dimensional rectangular detection frame, and h is the height of the two-dimensional rectangular detection frame.

[0018] The application provides a target detection result suppression method based on a probability density function, and the target detection result is a three-dimensional cubic detection frame.

[0019] The step of converting the target detection result into the probability density function based on the Gaussian distribution comprises the following steps:

[0020] The center point of the three-dimensional cubic detection frame is taken as an origin, the length of a bottom surface of the three-dimensional cubic detection frame, the width of the bottom surface and the height of the three-dimensional cubic detection frame are taken as three-axis lengths, so that a target detection ellipsoid is constructed.

[0021] The probability density function is constructed based on a three-dimensional Gaussian distribution for a set of points in the target detection ellipsoid.

[0022] The application provides a target detection result suppression method based on a probability density function, and the step of calculating a relative distance of the target detection result according to the probability density function comprises the following steps:

[0023] The relative entropy D of the target detection result is calculated according to the probability density function.

[0024] The relative entropy D is normalized to obtain the relative distance of the target detection result

[0025]

[0026] In the formula, τ is a set offset constant.

[0027] According to the target detection result suppression method based on the probability density function, the step of calculating the relative distance of the target detection result according to the probability density function comprises:

[0028] According to the probability density function Σ 1 / 2 a and the probability density function Σ 1 / 2 b , the relative entropy D(a||b) of the first target detection result and the second target detection result is calculated:

[0029]

[0030] In the formula, w a is the width of the two-dimensional rectangular detection frame in the first target detection result; w b is the width of the two-dimensional rectangular detection frame in the second target detection result; h a is the height of the two-dimensional rectangular detection frame in the first target detection result; h b is the height of the two-dimensional rectangular detection frame in the second target detection result; Δx is the difference between the x coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; and Δy is the difference between the y coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result.

[0031] The relative entropy D(a||b) is normalized to obtain the relative distance D of the first target detection result and the second target detection result

[0032]

[0033] In the formula, τ is a set offset constant.

[0034] According to the target detection result suppression method based on the probability density function, the step of calculating the relative distance of the target detection result according to the probability density function comprises:

[0035] According to the probability density function Σ 1 / 2 a and the probability density function Σ 1 / 2 b , the relative entropy D(a||b) of the first target detection result and the second target detection result is calculated:

[0036]

[0037] In the formula, wa w is the width of the two-dimensional rectangular detection frame in the first target detection result; b w is the width of the two-dimensional rectangular detection frame in the second target detection result; a h is the height of the two-dimensional rectangular detection frame in the first target detection result; b h is the height of the two-dimensional rectangular detection frame in the second target detection result; Δx is the difference between the x coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; Δy is the difference between the y coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; ε is a constant set;

[0038] normalizing the relative entropy D(a||b) to obtain the relative distance of the first target detection result and the second target detection result

[0039]

[0040] In the formula, τ is a set offset constant.

[0041] The application further provides a target detection result suppression system based on a probability density function, comprising:

[0042] an acquisition module configured to acquire at least two target detection results;

[0043] a probability module configured to convert the target detection results into probability density functions based on Gaussian distribution;

[0044] a distance module configured to calculate the relative distance of the target detection results according to the probability density functions;

[0045] a suppression module configured to determine whether the relative distance of the two target detection results is less than a set threshold value, and if so, perform confidence-based non-maximum suppression on an overlapping result set composed of the at least two target detection results whose relative distance is less than the set threshold value to obtain the target detection result with the maximum confidence as a target detection conclusion.

[0046] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the target detection result suppression method based on a probability density function according to any of the above embodiments when executing the program.

[0047] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the steps of the target detection result suppression method based on a probability density function according to any of the above embodiments.

[0048] The application further provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the target detection result suppression method based on a probability density function according to any one of the above.

[0049] The target detection result suppression method based on a probability density function provided by the application converts target detection results into probability distributions, and then obtains an overlapping set of detection targets by calculating the probability distribution distance of the target detection results, and performs non-maximum suppression based on the overlapping set. Compared with the traditional intersection over union method, the method effectively reduces the complexity of the calculation model, converts the geometric problem with multiple possibilities into an algebraic problem with a general solution, and can quickly determine whether multiple candidate boxes are the same detection target, thereby effectively improving the efficiency of target detection. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0051] Figure 1 is a flowchart of the target detection result suppression method based on a probability density function provided by the application;

[0052] Figure 2 is a structural schematic diagram of an electronic device provided by the application. DETAILED DESCRIPTION

[0053] In order to make the objects, technical solutions and advantages of the application clearer, the technical solutions in the application will be described clearly and completely below with reference to the drawings in the application. Obviously, the described embodiments are some embodiments of the application, but not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0054] The following will describe the target detection result suppression method based on a probability density function of the application. Figure 1 The target detection result suppression method based on a probability density function of the application is described as follows.

[0055] As shown in Figure 1 , the embodiment of the application provides a target detection result suppression method based on a probability density function, comprising:

[0056] Step 101, obtaining at least two target detection results;

[0057] Step 103, converting the target detection result into a probability density function based on a Gaussian distribution;

[0058] Step 105, calculating the relative distance of the target detection result according to the probability density function;

[0059] Step 107, determining that the relative distance of two target detection results is less than a set threshold, and performing confidence-based non-maximum suppression on an overlapping result set composed of at least two target detection results with a relative distance less than the set threshold to obtain the target detection result with the maximum confidence as the target detection conclusion.

[0060] In this embodiment, the target detection result can be a two-dimensional detection box based on the image to be detected, or a three-dimensional detection box based on the point cloud data to be detected.

[0061] It is worth noting that in this embodiment, the relative distance calculated according to the probability density function is a relative quantity based on two target detection results. The overlapping result set can have the following three element relationships according to different application scenarios:

[0062] 1. In the overlapping result set, the relative distance of any two target detection results is less than the set threshold;

[0063] 2. In the overlapping result set, for any target detection result (denoted as A), there is at least one other target detection result (denoted as B) different from the target detection result such that the relative distance between A and B is less than the set threshold;

[0064] 3. The overlapping result set includes a basic target detection result, and the relative distance between the basic target detection result and all other target detection results is less than the set threshold.

[0065] In addition, the naming of the overlapping result set cannot be understood as a limitation of this embodiment. The overlap in this embodiment refers to the overlapping of the value range of the relative distance in the set and the allowed range of the threshold, rather than the overlap between the detection boxes of two target detection results.

[0066] In one preferred embodiment of the present embodiment, the step of calculating the relative distance of the target detection result according to the probability density function comprises:

[0067] According to the probability density function, the relative entropy D of the target detection result is calculated;

[0068] The relative entropy D is normalized to obtain the relative distance of the target detection result

[0069]

[0070] wherein τ is a set offset constant.

[0071] That is, the relative distance is a relative entropy KL of a normalized probability distribution.

[0072] The embodiment has the following beneficial effects:

[0073] The target detection result is converted into a probability distribution, and then the probability distribution distance of the target detection result is calculated to obtain an overlapping set of the same detection target, and non-maximum suppression is performed based on the overlapping set; compared with the traditional intersection over union method, the complexity of the calculation model is effectively reduced, the geometric problem with multiple possibilities is converted into an algebraic problem with a general solution, and it can be quickly judged whether multiple candidate boxes are the same detection target, thereby effectively improving the efficiency of target detection.

[0074] According to the above embodiment, in the embodiment, the two-dimensional and three-dimensional target detection results are further described respectively.

[0075] For the two-dimensional target detection result:

[0076] The target detection result is a two-dimensional rectangular detection box;

[0077] The step of converting the target detection result into a probability density function based on the Gaussian distribution comprises:

[0078] The center point of the two-dimensional rectangular detection box is taken as the origin, the height of the two-dimensional rectangular detection box is taken as the length of the long axis, and the width of the two-dimensional rectangular detection box is taken as the length of the short axis to construct a target detection ellipse.

[0079] A probability density function Σ 1 / 2 is constructed based on the two-dimensional Gaussian distribution for the set of points in the target detection ellipse.

[0080]

[0081] wherein θ is the rotation angle of the two-dimensional rectangular detection box; w is the width of the two-dimensional rectangular detection box; and h is the height of the two-dimensional rectangular detection box.

[0082] The step of calculating the relative distance of the target detection result according to the probability density function comprises:

[0083] The probability density function Σ 1 / 2 a of the first target detection result and the probability density function Σ 1 / 2 b of the second target detection result are calculated to obtain the relative entropy D(a||b) of the first target detection result and the second target detection result.

[0084]

[0085] wherein w a is the width of the two-dimensional rectangular detection frame in the first target detection result; w b is the width of the two-dimensional rectangular detection frame in the second target detection result; h a is the height of the two-dimensional rectangular detection frame in the first target detection result; h b is the height of the two-dimensional rectangular detection frame in the second target detection result; Δx is the difference between the x coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; and Δy is the difference between the y coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result;

[0086] normalizing the relative entropy D(a||b) to obtain the relative distance between the first target detection result and the second target detection result

[0087]

[0088] wherein τ is a set offset constant, which can be selected as 1 or 2 in the embodiment.

[0089] The above is one embodiment of the two-dimensional target detection frame. In order to prevent the case that the intermediate quantity in the operation exceeds the range allowed by the processor when the denominator quantity is small, and further causes operation error, the following preferred solution can be used:

[0090] The step of calculating the relative distance of the target detection result according to the probability density function comprises:

[0091] calculating the probability density function Σ 1 / 2 a of the first target detection result and the probability density function Σ 1 / 2 b of the second target detection result to obtain the relative entropy D(a||b) of the first target detection result and the second target detection result:

[0092]

[0093] wherein w a is the width of the two-dimensional rectangular detection frame in the first target detection result; w b is the width of the two-dimensional rectangular detection frame in the second target detection result; h a is the height of the two-dimensional rectangular detection frame in the first target detection result; h bThe height of the two-dimensional rectangular detection frame in the second target detection result; Δx is the difference between the x-coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x-coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; Δy is the difference between the y-coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y-coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; ε is a constant set in this embodiment, which can be selected as 0.000001;

[0094] Normalizing the relative entropy D(a||b) to obtain the relative distance of the first target detection result and the second target detection result

[0095]

[0096] In the formula, τ is a set offset constant.

[0097] For a three-dimensional target detection result:

[0098] The target detection result is a three-dimensional cubic detection frame;

[0099] The step of converting the target detection result into a probability density function based on a Gaussian distribution comprises:

[0100] Taking the center point of the three-dimensional cubic detection frame as the origin, the length of the bottom surface, the width of the bottom surface, and the height of the three-dimensional cubic detection frame as the lengths of three axes, a target detection ellipsoid is constructed.

[0101] A probability density function is constructed based on a three-dimensional Gaussian distribution for the set of points inside the target detection ellipsoid.

[0102] The three-dimensional Gaussian distribution and the relative entropy calculation can be extended on the basis of the above-mentioned two-dimensional scheme, and this embodiment will not make more description.

[0103] The following will describe the embodiment in the automatic driving scene, for target detection of other vehicles other than the vehicle.

[0104] For a two-dimensional target detection result, the aforementioned two-dimensional scheme can be referred to.

[0105] For a three-dimensional target detection result based on automobile point cloud data, i.e., a three-dimensional cubic detection frame, since automobile movement is based on the ground surface and vehicle rollover is a small probability event, the process of obtaining the probability density function corresponding to the three-dimensional cubic detection frame of the automobile point cloud can be further simplified, the influence of the pitch angle and the roll angle is ignored, and only the yaw angle θ is considered for the transformation and construction of the probability density function, so that the calculation process of the three-dimensional cubic detection frame is simplified to the calculation of a two-dimensional detection frame based on a top view.

[0106] The beneficial effects of the embodiment are that:

[0107] 1. The complex process of calculating the overlapping area of two rotating detection boxes is omitted.

[0108] 2. The KL distance is used to measure the overlapping degree of the rotating detection box, the calculation is fast, and the whole process is differentiable, which can be trained as a loss function.

[0109] The probability density function-based target detection result suppression device provided by the present application is described below, and the probability density function-based target detection result suppression device described below can be correspondingly referred to the probability density function-based target detection result suppression method described above.

[0110] The embodiment of the present application also provides a probability density function-based target detection result suppression system, characterized in that it comprises:

[0111] The acquisition module is configured to acquire at least two target detection results.

[0112] The probability module is configured to convert the target detection results into probability density functions based on Gaussian distribution.

[0113] The distance module is configured to calculate the relative distance of the target detection results based on the probability density functions.

[0114] The suppression module is configured to determine that the relative distance of two target detection results is less than a set threshold value, and then perform non-maximum suppression based on confidence for an overlapping result set composed of at least two target detection results whose relative distance is less than the set threshold value, to obtain the target detection result with the maximum confidence as the target detection conclusion.

[0115] Further, the distance module comprises:

[0116] The relative entropy unit is configured to calculate the relative entropy D of the target detection results based on the probability density functions.

[0117] The normalization unit is configured to normalize the relative entropy D to obtain the relative distance of the target detection results.

[0118]

[0119] In the formula, τ is a set offset constant.

[0120] Further, for two-dimensional target detection results:

[0121] The target detection result is a two-dimensional rectangular detection box.

[0122] The probability module comprises:

[0123] a two-dimensional ellipse unit, configured to construct a target detection ellipse with a center point of the two-dimensional rectangular detection frame as an origin, a height of the two-dimensional rectangular detection frame as a length of a major axis, and a width of the two-dimensional rectangular detection frame as a length of a minor axis;

[0124] a two-dimensional probability density function unit, configured to construct a probability density function Σ 1 / 2 :

[0125]

[0126] wherein θ is a rotation angle of the two-dimensional rectangular detection frame, w is a width of the two-dimensional rectangular detection frame, and h is a height of the two-dimensional rectangular detection frame.

[0127] The relative entropy unit can be further configured to:

[0128] a probability density function Σ 1 / 2 a of the first target detection result, and a probability density function Σ 1 / 2 b of the second target detection result, to obtain a relative entropy D(a||b) of the first target detection result and the second target detection result:

[0129]

[0130] wherein w a is a width of a two-dimensional rectangular detection frame in the first target detection result, w b is a width of a two-dimensional rectangular detection frame in the second target detection result, h a is a height of a two-dimensional rectangular detection frame in the first target detection result, h b is a height of a two-dimensional rectangular detection frame in the second target detection result, Δx is a difference between an x coordinate of a center point of a two-dimensional rectangular detection frame in the first target detection result and an x coordinate of a center point of a two-dimensional rectangular detection frame in the second target detection result, and Δy is a difference between a y coordinate of a center point of a two-dimensional rectangular detection frame in the first target detection result and a y coordinate of a center point of a two-dimensional rectangular detection frame in the second target detection result.

[0131] Alternatively, the relative entropy unit can be configured to:

[0132] a probability density function Σ 1 / 2 a of the first target detection result, and a probability density function Σ 1 / 2 b of the second target detection result, to obtain a relative entropy D(a||b) of the first target detection result and the second target detection result:

[0133]

[0134] In the formula, w a is the width of the two-dimensional rectangular detection frame in the first target detection result; w b is the width of the two-dimensional rectangular detection frame in the second target detection result; h a is the height of the two-dimensional rectangular detection frame in the first target detection result; h b is the height of the two-dimensional rectangular detection frame in the second target detection result; Δx is the difference between the x coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; Δy is the difference between the y coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; ε is a constant set;

[0135] The normalization unit is further used to:

[0136] normalize the relative entropy D(a||b) to obtain the relative distance of the first target detection result and the second target detection result

[0137]

[0138] In the formula, τ is a set offset constant.

[0139] The embodiment has the following beneficial effects:

[0140] The target detection result is converted into a probability distribution, and then the probability distribution distance of the target detection result is calculated to obtain an overlapping set of the same detection target, and non-maximum suppression is performed based on the overlapping set; compared with the traditional intersection-over-union method, the complexity of the calculation model is effectively reduced, the geometric problem with multiple possibilities is converted into an algebraic problem with a general solution, whether multiple candidate frames are the same detection target can be quickly judged, and thus the efficiency of target detection is effectively improved.

[0141] Figure 2 An example of an entity structure diagram of an electronic device is shown in FIG. 1. Figure 2As shown, the electronic device can include a processor 210, a communications interface 220, a memory 230, and a communications bus 240, wherein the processor 210, the communications interface 220, and the memory 230 complete mutual communication through the communications bus 240. The processor 210 can invoke a logic instruction in the memory 230 to execute a target detection result suppression method based on a probability density function, which includes: obtaining at least two target detection results; converting the target detection results into a probability density function based on a Gaussian distribution; calculating a relative distance of the target detection results according to the probability density function; determining that the relative distances of two target detection results are less than a set threshold, and then performing non-maximum suppression based on confidence for an overlapping result set composed of at least two target detection results whose relative distances are less than the set threshold to obtain the target detection result with the maximum confidence as a target detection conclusion.

[0142] In addition, the logic instruction in the memory 230 described above can be implemented in the form of a software functional unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the part of the technical solutions that essentially contribute to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0143] In another aspect, the present application also provides a computer program product comprising a computer program, which can be stored on a non-transitory computer readable storage medium, and the computer program is executable by a processor to enable a computer to perform the target detection result suppression method based on a probability density function provided by the above method, which comprises: obtaining at least two target detection results; converting the target detection results into a probability density function based on a Gaussian distribution; calculating the relative distance of the target detection results according to the probability density function; determining that the relative distance of two target detection results is less than a set threshold value, and then performing confidence-based non-maximum suppression on an overlapping result set composed of at least two target detection results with a relative distance less than the set threshold value to obtain the target detection result with the maximum confidence as the target detection conclusion.

[0144] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, and the computer program is executable by a processor to implement the target detection result suppression method based on a probability density function provided by the above method, which comprises: obtaining at least two target detection results; converting the target detection results into a probability density function based on a Gaussian distribution; calculating the relative distance of the target detection results according to the probability density function; determining that the relative distance of two target detection results is less than a set threshold value, and then performing confidence-based non-maximum suppression on an overlapping result set composed of at least two target detection results with a relative distance less than the set threshold value to obtain the target detection result with the maximum confidence as the target detection conclusion.

[0145] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separate, and the components illustrated as units can or can not be physical units, i.e., can be located in one place or distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement without creative labor.

[0146] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0147] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for suppressing a target detection result based on a probability density function, characterized in that, The method comprises the following steps: obtaining at least two target detection results; converting the target detection results into a probability density function based on a Gaussian distribution; According to the probability density function, a relative distance of the target detection result is calculated, and the step comprises: according to the probability density function, a relative entropy of the target detection result is calculated ; and the relative entropy is normalized to obtain a relative distance of the target detection result . ​ ； In the formula, is a set offset constant; determining whether the relative distance between two target detection results is less than a set threshold value, and if so, performing confidence-based non-maximum suppression on an overlapping result set composed of at least two target detection results with a relative distance less than the set threshold value, and obtaining the target detection result with the maximum confidence value as the target detection conclusion; the target detection result is a two-dimensional rectangular detection box; the step of converting the target detection result into a probability density function based on a Gaussian distribution comprises: constructing a target detection ellipse with the center point of the two-dimensional rectangular detection box as the origin, the height of the two-dimensional rectangular detection box as the length of the major axis, and the width of the two-dimensional rectangular detection box as the length of the minor axis; A set of in-ellipse points is detected for the target, and a probability density function is constructed based on a two-dimensional Gaussian distribution : ； In the formula, is a rotation angle of the two-dimensional rectangular detection frame; is a width of the two-dimensional rectangular detection frame; is a height of the two-dimensional rectangular detection frame; alternatively, the target detection result is a three-dimensional cubic detection box; a probability density function of the first target detection result and a probability density function of the second target detection result , and the relative entropy of the first target detection result and the second target detection result is calculated : ； In the formula, is the width of the two-dimensional rectangular detection frame in the first target detection result; is the width of the two-dimensional rectangular detection frame in the second target detection result; is the height of the two-dimensional rectangular detection frame in the first target detection result; is the height of the two-dimensional rectangular detection frame in the second target detection result; is the difference between the x-coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x-coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; is the difference between the y-coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y-coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; is a constant set. normalizing the relative entropy to obtain a relative distance between the first target detection result and the second target detection result : ； the step of converting the target detection result into a probability density function based on a Gaussian distribution comprises: constructing a target detection ellipsoid with the center point of the three-dimensional cubic detection box as the origin, the length of the bottom surface, the width of the bottom surface, and the height of the three-dimensional cubic detection box as the lengths of the three axes; and constructing a probability density function based on a three-dimensional Gaussian distribution for a set of points inside the target detection ellipsoid. The method comprises the following steps:

2. A probability density function based target detection result suppression system, characterized by, an obtaining module for obtaining at least two target detection results; a probability module for converting the target detection results into a probability density function based on a Gaussian distribution; the target detection result is a two-dimensional rectangular detection box; the step of converting the target detection result into a probability density function based on a Gaussian distribution comprises: constructing a target detection ellipse with the center point of the two-dimensional rectangular detection box as the origin, the height of the two-dimensional rectangular detection box as the length of the major axis, and the width of the two-dimensional rectangular detection box as the length of the minor axis; alternatively, the target detection result is a three-dimensional cubic detection box; A set of in-ellipse points is detected for the target, and a probability density function is constructed based on a two-dimensional Gaussian distribution : ； In the formula, is a rotation angle of the two-dimensional rectangular detection frame; is a width of the two-dimensional rectangular detection frame; is a height of the two-dimensional rectangular detection frame; the step of converting the target detection result into a probability density function based on a Gaussian distribution comprises: constructing a target detection ellipsoid with the center point of the three-dimensional cubic detection box as the origin, the length of the bottom surface, the width of the bottom surface, and the height of the three-dimensional cubic detection box as the lengths of the three axes; and a distance module configured to calculate a relative distance of the target detection result according to the probability density function, the steps comprising: calculating a relative entropy of the target detection result according to the probability density function ; and normalizing the relative entropy to obtain the relative distance of the target detection result . ​ ； In the formula, is a set offset constant; constructing a probability density function based on a three-dimensional Gaussian distribution for a set of points inside the target detection ellipsoid; a probability density function of the first target detection result and a probability density function of the second target detection result , and the relative entropy of the first target detection result and the second target detection result is calculated : ； In the formula, is the width of the two-dimensional rectangular detection frame in the first target detection result; is the width of the two-dimensional rectangular detection frame in the second target detection result; is the height of the two-dimensional rectangular detection frame in the first target detection result; is the height of the two-dimensional rectangular detection frame in the second target detection result; is the difference between the x-coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the x-coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; is the difference between the y-coordinate of the center point of the two-dimensional rectangular detection frame in the first target detection result and the y-coordinate of the center point of the two-dimensional rectangular detection frame in the second target detection result; is a constant set. normalizing the relative entropy to obtain a relative distance between the first target detection result and the second target detection result : ； the step of calculating the relative distance of the target detection result based on the probability density function comprises:

3. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, a suppression module for determining whether the relative distance between two target detection results is less than a set threshold value, and if so, performing confidence-based non-maximum suppression on an overlapping result set composed of at least two target detection results with a relative distance less than the set threshold value, and obtaining the target detection result with the maximum confidence value as the target detection conclusion.

4. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The processor executes the program to implement the steps of the target detection result suppression method based on a probability density function according to claim 1. The computer program is executed by the processor to implement the steps of the target detection result suppression method based on a probability density function according to claim 1.

5. A computer program product comprising a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the target detection result suppression method based on a probability density function according to claim 1.

Images