Non-maximum suppression method and device, electronic equipment and storage medium

Abstract

The method comprises: determining a feasible region corresponding to a second feature map based on a first target detection box of a preset 3D target in a first feature map; determining a plurality of reference detection boxes according to the feasible region; determining a rotation reference vector according to the plurality of reference detection boxes and the first target detection box; determining a similarity compensation value of a non-maximum value reference detection box based on the rotation reference vector and a direction vector of the non-maximum value reference detection box; and determining a similarity between the non-maximum value reference detection box and a target reference detection box based on an intersection over union and the similarity compensation value. The method solves the problem that the traditional NMS method has poor robustness for 3D target detection results and is prone to false detection, and improves the 3D target detection accuracy.

Description

Technical Field

[0001] This disclosure relates to the field of target detection technology, and in particular to a nonmaximum suppression method, apparatus, electronic device, and storage medium. Background Technology

[0002] Non-Maximum Suppression (NMS) is a widely used technique in computer vision for detecting and removing redundant object bounding boxes or keypoints from images or point clouds. NMS is commonly used in object detection tasks to remove duplicate detection results, thereby improving the accuracy and efficiency of the algorithm.

[0003] In traditional object detection methods, object detectors typically output a series of candidate boxes and use a threshold filter to select the candidate boxes with high confidence. However, these candidate boxes may overlap or be close to each other, leading to duplicate detections or false detections. Non-Maximum Search (NMS) addresses these issues by comparing the intersection-union ratio (IUU) of different candidate boxes and retaining the candidate box with the highest confidence.

[0004] Unfortunately, the aforementioned techniques for resolving duplicate or false detections using Non-Maximum Detection (NMS) are suitable for 2D object detection because 2D object bounding boxes are generally axially aligned, making it relatively easy to design the Intersection over Union (IOU) threshold. However, for 3D object detection, the bounding boxes are rotated, and this rotation causes significant changes in the IOU threshold. Even if two bounding boxes are close, their IOU may be low due to rotation, resulting in poor performance of NMS in resolving duplicate or false detections in 3D object detection.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] To address the aforementioned technical problems, or at least partially address them, this disclosure provides a non-maximum suppression method, apparatus, electronic device, and storage medium. These methods solve the problems of poor robustness and susceptibility to false detection in traditional NMS methods for 3D target detection. By considering historical frame detection results and the unique rotational properties of 3D target detection, the method can effectively handle target overlap issues within a rotated frame, thereby improving the accuracy of 3D target detection.

[0007] In a first aspect, embodiments of this disclosure provide a non-maximum suppression method, the method comprising:

[0008] Based on the first target detection box of the preset 3D target in the first feature map, a feasible region corresponding to the second feature map is determined. The first feature map is a historical feature map before the second feature map.

[0009] Based on the feasible region, multiple reference detection boxes are determined from the candidate detection boxes of the preset 3D target in the second feature map;

[0010] A rotation reference vector is determined based on the plurality of reference detection boxes and the first target detection box;

[0011] For each non-maximum reference detection box, the intersection-union ratio (IUGR) between the non-maximum reference detection box and the target reference detection box is determined; the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection box is the reference detection box among the plurality of reference detection boxes excluding the target reference detection box.

[0012] The similarity compensation value of the non-maximum reference detection box is determined based on the rotation reference vector and the orientation vector of the non-maximum reference detection box.

[0013] The similarity between the non-maximum reference detection box and the target reference detection box is determined based on the intersection-union ratio and the similarity compensation value.

[0014] Secondly, embodiments of this disclosure also provide a non-maximum suppression device, the device comprising:

[0015] The first determining module is used to determine the feasible region corresponding to the second feature map based on the first target detection box of the preset 3D target in the first feature map, wherein the first feature map is a historical feature map preceding the second feature map;

[0016] The second determining module is used to determine multiple reference detection boxes from the candidate detection boxes of the preset 3D target in the second feature map based on the feasible region;

[0017] The third determining module is used to determine a rotation reference vector based on the plurality of reference detection boxes and the first target detection box;

[0018] The fourth determining module is used to determine the intersection-union ratio (IUR) between the non-maximum reference detection box and the target reference detection box for each non-maximum reference detection box; the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection box is the reference detection box among the plurality of reference detection boxes excluding the target reference detection box;

[0019] The fifth determining module is used to determine the similarity compensation value of the non-maximum reference detection box based on the rotation reference vector and the direction vector of the non-maximum reference detection box;

[0020] The sixth determining module is used to determine the similarity between the non-maximum reference detection box and the target reference detection box based on the intersection-union ratio and the similarity compensation value.

[0021] Thirdly, embodiments of this disclosure also provide an electronic device, the electronic device comprising: one or more processors; a storage device for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the nonmaximum suppression method as described above.

[0022] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the nonmaximum suppression method as described above.

[0023] This disclosure provides a non-maximum suppression method, which determines a feasible region corresponding to a second feature map based on a first target detection box of a preset 3D target in a first feature map. The first feature map is a historical feature map preceding the second feature map. Multiple reference detection boxes are determined from the candidate detection boxes of the preset 3D target in the second feature map according to the feasible region. A rotation reference vector is determined based on the multiple reference detection boxes and the first target detection box. For each non-maximum reference detection box, the intersection-union ratio (IUR) between the non-maximum reference detection box and the target reference detection box is determined. The target reference detection box is the reference detection box with the highest confidence among the multiple reference detection boxes. The maximum reference detection box is the reference detection box other than the target reference detection box among the plurality of reference detection boxes; the similarity compensation value of the non-maximum reference detection box is determined based on the rotation reference vector and the direction vector of the non-maximum reference detection box; the similarity between the non-maximum reference detection box and the target reference detection box is determined based on the intersection-union ratio and the similarity compensation value. This technical means solves the problem that the traditional NMS method has poor robustness to 3D target detection results and is prone to false detection. By considering the detection results of historical frames and the rotation attributes unique to 3D target detection, the target overlap problem under the rotating box can be effectively handled, thereby improving the accuracy of 3D target detection. Attached Figure Description

[0024] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0025] Figure 1 This is a flowchart of a nonmaximum suppression method according to an embodiment of the present disclosure;

[0026] Figure 2 This is a schematic diagram of a rotating target detection box in an embodiment of this disclosure;

[0027] Figure 3 This is a schematic diagram of a feasible domain in an embodiment of this disclosure;

[0028] Figure 4 This is a schematic diagram of a rotation reference vector in an embodiment of this disclosure;

[0029] Figure 5 This is a schematic diagram of a nonmaximum suppression device according to an embodiment of the present disclosure;

[0030] Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this disclosure. Detailed Implementation

[0031] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0032] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0033] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0034] Figure 1 This is a flowchart of a nonmaximum suppression method according to an embodiment of this disclosure. The method is used for 3D target detection and can be executed by a nonmaximum suppression device, which can be implemented in software and / or hardware and can be configured in an electronic device. Figure 1 As shown, the method may specifically include the following steps:

[0035] S110. Based on the first target detection box of the preset 3D target in the first feature map, determine the feasible region corresponding to the second feature map. The first feature map is a historical feature map before the second feature map.

[0036] The preset 3D target can be any object in the physical world, such as a moving car or a pedestrian.

[0037] When acquiring point cloud frames or image frames, the detector typically outputs multiple candidate detection boxes for the preset 3D target. Each candidate detection box can be represented as [cls, x, y, z, l, w, h, theta, vx, vy, score]. cls represents the category of the 3D target (cls is the same and unique among multiple candidate detection boxes for the same 3D target), x, y, and z represent the coordinates of the candidate detection box, l, w, and h represent the length, width, and height of the candidate detection box, respectively, theta represents the rotation angle of the candidate detection box, vx and vy represent the velocity of the candidate detection box, and score represents the confidence level. Then, a threshold filter is used to select candidate detection boxes with high confidence. However, these candidate detection boxes may still overlap or be close to each other, resulting in multiple redundant detection boxes for the same target. It is necessary to compare the similarity between different detection boxes using a non-maximum suppression method. Based on the similarity, it is determined whether the current detection box and the detection box with the highest confidence represent the same target. If they do, the current detection box is deleted; otherwise, the current detection box is retained. This can solve the problem of multiple redundant detection boxes for the same target.

[0038] The traditional nonmaximum suppression method follows this process:

[0039] For each category, multiple candidate detection boxes are selected, with the highest confidence score, as the target detection box. The overlap (e.g., intersection-union ratio) between the remaining candidate detection boxes and the target detection box is calculated, and candidate detection boxes with an overlap greater than a threshold are deleted. However, for 3D object detection, the candidate detection boxes are rotated, and this rotation causes significant changes in the Interchange of Units (IOU). Even if two candidate detection boxes are very close, their IOU may be small due to rotation, leading to poor non-maximum suppression in 3D object detection. Figure 2 As shown, although the gray target 210 and the blue target 220 are very close, the rotation results in a small IOU between them. Therefore, both candidate detection boxes are retained, even though they are actually detection boxes for the same target, leading to false detections. To address this problem, the solution proposed in this application is presented.

[0040] The first target detection box is the detection box with the highest confidence for the preset 3D target in the first feature map. If the first target detection box is the first frame in the entire sequence, it can be selected from multiple candidate detection boxes using a traditional non-maximum suppression method.

[0041] The second feature map is a feature map that follows the first feature map. For example, if the first feature map is the first frame, the second feature map could be the second or third frame. In other words, the first feature map is a historical feature map that precedes the second feature map, and typically the feature map of the frame preceding the second feature map is selected as the first feature map.

[0042] Specifically, the determination of the feasible region corresponding to the second feature map based on the first target detection box with a preset 3D target in the first feature map includes:

[0043] The feasible region is determined based on the prediction speed of the preset 3D target indicated by the first target detection box.

[0044] Specifically, in a preset coordinate system, with the center point of the first target detection box as the origin, the image region defined by the following function is determined as the feasible region:

[0045]

[0046] Among them, v x v represents the predicted velocity of the preset 3D target indicated by the first target detection box along the x-axis. y a represents the predicted velocity of the preset 3D target indicated by the first target detection box along the y-axis. x a represents the acceleration of the preset 3D target along the x-axis. y dt represents the acceleration of the preset 3D target along the y-axis, and dt represents the acquisition time difference between the original data frames corresponding to the first feature map and the second feature map, respectively. x a y The target detection category cls can be set based on human experience or dynamically calculated based on the acceleration formula; this application does not impose any restrictions.

[0047] For example, such as Figure 3 As shown, the dashed box 310 represents the first target detection box, and the quarter-circle 320 represents the feasible region. A rectangular box that falls entirely within the feasible region 320 or a rectangular box whose most area falls within the feasible region 320 is a detection box within the feasible region.

[0048] S120. Based on the feasible region, determine a plurality of reference detection boxes from the candidate detection boxes of the preset 3D target in the second feature map.

[0049] Specifically, candidate detection boxes whose areas fall within the feasible region and reach a preset proportion are determined as reference detection boxes, such as... Figure 3 As shown.

[0050] S130. Determine a rotation reference vector based on the plurality of reference detection boxes and the first target detection box.

[0051] Specifically, the average value of the center point coordinates of the plurality of reference detection boxes is determined; the rotation reference vector is constructed with the center point of the first target detection box as the starting point and the point corresponding to the average value as the ending point. For example... Figure 4 The diagram shows a rotating reference vector, where point 410 represents the center point of the first target detection box 310, point 420 represents the point corresponding to the average coordinate of the center points of the plurality of reference detection boxes, and arrow 430 represents the rotating reference vector.

[0052] S140. For each non-maximum reference detection box, determine the intersection-union ratio (IUGR) between the non-maximum reference detection box and the target reference detection box; the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection box is the reference detection box among the plurality of reference detection boxes excluding the target reference detection box.

[0053] S150. Determine the similarity compensation value of the non-maximum reference detection box based on the rotation reference vector and the direction vector of the non-maximum reference detection box.

[0054] Specifically, the distance between the rotation reference vector and the direction vector of the non-maximum reference detection box is determined. This distance can be cosine distance, Euclidean distance, or other forms that can represent the distance between vectors. Taking cosine distance as an example, the distance d between the rotation reference vector and the direction vector of the non-maximum reference detection box is determined by the following expression: Where u represents the rotation reference vector, v represents the direction vector of the non-maximum reference detection box, ||u|| represents the magnitude of vector u, and ||v|| represents the magnitude of vector v.

[0055] The rotation similarity penalty term for the non-maximum reference detection box is determined based on the distance using the following formula:

[0056] cost = 1 - d

[0057] Where cost represents the rotation similarity penalty term, and d represents the distance;

[0058] The similarity compensation value is determined based on the rotation similarity penalty term according to the following formula:

[0059] T = k × cost, where T represents the similarity compensation value and k represents the preset weight.

[0060] S160. Determine the similarity between the non-maximum reference detection box and the target reference detection box based on the intersection-union ratio and the similarity compensation value.

[0061] Specifically, the similarity is determined based on the following formula:

[0062] Metric = IOU + T

[0063] Wherein, Metric represents the similarity, IOU represents the intersection-union ratio, and T represents the similarity compensation value.

[0064] After determining the similarity between the non-maximum reference detection box and the target reference detection box, the similarity is compared with a threshold. If the similarity is greater than the threshold, the non-maximum reference detection box is deleted.

[0065] This embodiment provides a non-maximum suppression method for 3D object detection that combines historical frames and rotation information. It introduces a rotation similarity penalty term for 3D object detection to suppress false detections of targets with small Intersection over Union (IoU) but close actual distances, thus achieving better suppression and improving the accuracy of 3D object detection. Specifically, regarding the combination of historical frame detection results, this invention proposes the concept of a feasible region to filter out the range of target detection results belonging to the current frame. Regarding rotation similarity, this invention constructs a rotation reference vector using the mean center of the detection results from historical frames and the redundant detection results from the current frame, and together with the rotation inherent in the detection box, constructs a rotation similarity penalty term. This effectively suppresses target detection boxes with small IoU but large rotation angle differences, thus effectively alleviating the false detection problem caused by insufficient IoU measurement due to rotation in 3D object detection.

[0066] Figure 5 This is a schematic diagram of a nonmaximum suppression device according to an embodiment of this disclosure. Figure 5 As shown: The device includes: a first determining module 510, a second determining module 520, a third determining module 530, a fourth determining module 540, a fifth determining module 550 and a sixth determining module 560.

[0067] The first determining module 510 is used to determine a feasible region corresponding to a second feature map based on a first target detection box of a preset 3D target in the first feature map, wherein the first feature map is a historical feature map preceding the second feature map; the second determining module 520 is used to determine multiple reference detection boxes from the candidate detection boxes of the preset 3D target in the second feature map according to the feasible region; the third determining module 530 is used to determine a rotation reference vector according to the multiple reference detection boxes and the first target detection box; and the fourth determining module 540 is used to determine the non-maximum parameter for each non-maximum reference detection box. The intersection-union ratio (IUR) of the reference detection box and the target reference detection box; the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection box is the reference detection box other than the target reference detection box among the plurality of reference detection boxes; the fifth determining module 550 is used to determine the similarity compensation value of the non-maximum reference detection box based on the rotation reference vector and the direction vector of the non-maximum reference detection box; the sixth determining module 560 is used to determine the similarity between the non-maximum reference detection box and the target reference detection box based on the IUR and the similarity compensation value.

[0068] Furthermore, the first determining module 510 includes a first determining unit, used to determine the feasible region based on the prediction speed of the preset 3D target indicated by the first target detection box.

[0069] Furthermore, the first determining unit is specifically used to: determine the image region that performs the following function as the feasible region in a preset coordinate system, with the center point of the first target detection box as the origin:

[0070]

[0071] Among them, v x v represents the predicted velocity of the preset 3D target indicated by the first target detection box along the x-axis. y a represents the predicted velocity of the preset 3D target indicated by the first target detection box along the y-axis. x a represents the acceleration of the preset 3D target along the x-axis. y dt represents the acceleration of the preset 3D target along the y-axis, and dt represents the acquisition time difference between the original data frames corresponding to the first feature map and the second feature map, respectively.

[0072] Furthermore, the third determining module 530 is specifically used to: determine the average value of the center point coordinates of the plurality of reference detection boxes; and construct the rotation reference vector with the center point of the first target detection box as the starting point and the point corresponding to the average value as the ending point.

[0073] Furthermore, the fifth determining module 550 is specifically used to: determine the distance between the rotation reference vector and the direction vector of the non-maximum reference detection box;

[0074] The rotation similarity penalty term for the non-maximum reference detection box is determined based on the distance using the following formula:

[0075] cost = 1 - d

[0076] Where cost represents the rotation similarity penalty term, and d represents the distance;

[0077] The similarity compensation value is determined based on the rotation similarity penalty term according to the following formula:

[0078] T = k × cost, where T represents the similarity compensation value and k represents the preset weight.

[0079] Furthermore, the sixth determining module 560 is specifically used to: determine the similarity based on the following formula:

[0080] Metric = IOU + T

[0081] Wherein, Metric represents the similarity, IOU represents the intersection-union ratio, and T represents the similarity compensation value.

[0082] Furthermore, the second determining module 520 is specifically used to: determine the candidate detection box whose area falls within the feasible region and reaches a preset ratio as the reference detection box.

[0083] The nonmaximum suppression device provided in this embodiment can execute the steps in the nonmaximum suppression method provided in this embodiment, and has the execution steps and beneficial effects, which will not be repeated here.

[0084] Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this disclosure. See below for details. Figure 6 It shows a schematic diagram of a structure suitable for implementing the electronic device 500 in the embodiments of this disclosure. Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0085] like Figure 6As shown, the electronic device 500 may include a processing device 501, a ROM 502, a RAM 503, a bus 504, an input / output (I / O) interface 505, an input device 506, an output device 507, a storage device 508, and a communication device 509. The processing device (e.g., a central processing unit, a graphics processor, etc.) 501 can perform various appropriate actions and processes to implement the methods of the embodiments described herein, based on a program stored in the read-only memory (ROM) 502 or a program loaded from the storage device 508 into the random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device 500. The processing device 501, ROM 502, and RAM 503 are interconnected via the bus 504. The input / output (I / O) interface 505 is also connected to the bus 504.

[0086] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts, thereby implementing the non-maximum suppression method as described above. In such embodiments, the computer program can be downloaded and installed from a network via communication device 509, or installed from storage device 508, or installed from ROM 502. When the computer program is executed by processing device 501, it performs the functions defined in the methods of embodiments of this disclosure.

[0087] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0088] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: determine a feasible region corresponding to a second feature map based on a first target detection box of a preset 3D target in a first feature map, wherein the first feature map is a historical feature map preceding the second feature map; determine a plurality of reference detection boxes from candidate detection boxes of the preset 3D target in the second feature map according to the feasible region; determine a rotation reference vector based on the plurality of reference detection boxes and the first target detection box; for each non-maximum reference detection box, determine the intersection-union ratio (IUGR) between the non-maximum reference detection box and a target reference detection box; wherein the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection boxes are the reference detection boxes among the plurality of reference detection boxes excluding the target reference detection box; determine a similarity compensation value for the non-maximum reference detection box based on the rotation reference vector and the direction vector of the non-maximum reference detection box; and determine the similarity between the non-maximum reference detection box and the target reference detection box based on the IUGR and the similarity compensation value.

[0089] Optionally, when one or more of the above-described procedures are executed by the electronic device, the electronic device may also perform other steps described in the above embodiments.

[0090] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0091] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

Claims

1. A nonmaximum suppression method, characterized in that, The method includes: Based on the first target detection box of the preset 3D target in the first feature map, a feasible region corresponding to the second feature map is determined. The first feature map is a historical feature map before the second feature map. Based on the feasible region, multiple reference detection boxes are determined from the candidate detection boxes of the preset 3D target in the second feature map; Determining a rotation reference vector based on the plurality of reference detection boxes and the first target detection box includes: determining the average value of the center point coordinates of the plurality of reference detection boxes; and constructing the rotation reference vector with the center point of the first target detection box as the starting point and the point corresponding to the average value as the ending point. For each non-maximum reference detection box, the intersection-union ratio (IUGR) between the non-maximum reference detection box and the target reference detection box is determined; the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection box is the reference detection box among the plurality of reference detection boxes excluding the target reference detection box. The similarity compensation value of the non-maximum reference detection box is determined based on the rotation reference vector and the orientation vector of the non-maximum reference detection box; The similarity between the non-maximum reference detection box and the target reference detection box is determined based on the intersection-union ratio and the similarity compensation value.

2. The method according to claim 1, characterized in that, The determination of the feasible region corresponding to the second feature map by the first target detection box based on the preset 3D target in the first feature map includes: The feasible region is determined based on the prediction speed of the preset 3D target indicated by the first target detection box.

3. The method according to claim 2, characterized in that, Determining the feasible region based on the prediction speed of the preset 3D target indicated by the first target detection box includes: In a preset coordinate system, with the center point of the first target detection box as the origin, the image region defined by the following function is determined as the feasible region: in, This indicates the predicted velocity of the preset 3D target indicated by the first target detection box along the x-axis. This indicates the predicted velocity of the preset 3D target indicated by the first target detection box along the y-axis. This represents the acceleration of the preset 3D target along the x-axis. dt represents the acceleration of the preset 3D target along the y-axis, and dt represents the acquisition time difference between the original data frames corresponding to the first feature map and the second feature map, respectively.

4. The method according to claim 1, characterized in that, The step of determining the similarity compensation value of the non-maximum reference detection box based on the rotation reference vector and the orientation vector of the non-maximum reference detection box includes: Determine the distance between the rotation reference vector and the direction vector of the non-maximum reference detection box; The rotation similarity penalty term for the non-maximum reference detection box is determined based on the distance using the following formula: Where cost represents the rotation similarity penalty term, and d represents the distance; The similarity compensation value is determined based on the rotation similarity penalty term according to the following formula: T = k × cost, where T represents the similarity compensation value and k represents the preset weight.

5. The method according to claim 1, characterized in that, Determining the similarity between the non-maximum reference detection box and the target reference detection box based on the intersection-union ratio and the similarity compensation value includes: The similarity is determined based on the following formula: Wherein, Metric represents the similarity, IOU represents the intersection-union ratio, and T represents the similarity compensation value.

6. The method according to claim 1, characterized in that, The step of determining a reference detection box from the candidate detection boxes of the preset 3D target in the second feature map based on the feasible region includes: Candidate detection boxes whose area falls within the feasible region and reaches a preset ratio are determined as reference detection boxes.

7. A nonmaximum suppression device, characterized in that, include: The first determining module is used to determine the feasible region corresponding to the second feature map based on the first target detection box of the preset 3D target in the first feature map, wherein the first feature map is a historical feature map preceding the second feature map; The second determining module is used to determine multiple reference detection boxes from the candidate detection boxes of the preset 3D target in the second feature map based on the feasible region; The third determining module is used to determine a rotation reference vector based on the plurality of reference detection boxes and the first target detection box, including: determining the average value of the center point coordinates of the plurality of reference detection boxes; and constructing the rotation reference vector with the center point of the first target detection box as the starting point and the point corresponding to the average value as the ending point. The fourth determining module is used to determine the intersection-union ratio (IUR) between the non-maximum reference detection box and the target reference detection box for each non-maximum reference detection box; the target reference detection box is the reference detection box with the highest confidence among the plurality of reference detection boxes, and the non-maximum reference detection box is the reference detection box among the plurality of reference detection boxes excluding the target reference detection box; The fifth determining module is used to determine the similarity compensation value of the non-maximum reference detection box based on the rotation reference vector and the direction vector of the non-maximum reference detection box; The sixth determining module is used to determine the similarity between the non-maximum reference detection box and the target reference detection box based on the intersection-union ratio and the similarity compensation value.

8. An electronic device, characterized in that, The electronic device includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Images