Weakly supervised object detection model training method and apparatus

Abstract

The application discloses a target detection model training method and device based on weak supervision, wherein the method comprises the following steps: acquiring original three-dimensional point cloud data and partially labeled three-dimensional point cloud data in a target environment; presetting a plurality of anchor point boxes, and screening the plurality of anchor point boxes according to corresponding point cloud densities to obtain corresponding first candidate boxes; dynamically adjusting the corresponding first candidate boxes to screen the corresponding first candidate boxes to obtain corresponding weak pseudo labels; screening the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes, and purifying pseudo labels according to real label boxes to obtain strong pseudo labels; and training a pre-set neural network according to the weak pseudo labels and the strong pseudo labels to obtain a trained target detection model, thereby reducing the labor and time costs.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a weakly supervised target detection model training method, a computer readable storage medium, a computer device and a weakly supervised target detection model training device. BACKGROUND

[0002] In the related art, most three-dimensional target detectors based on point clouds are implemented in a fully supervised manner, which requires large-scale precisely labeled three-dimensional point cloud data. Labeling point cloud data is very tedious and expensive, and consumes a large amount of manpower and time cost. It is necessary and urgent to reduce the heavy labeling burden. Using weakly supervised learning is generally considered an effective way to save costs. Meanwhile, weakly supervised learning has become a trend in semantic segmentation of point clouds, which provides a reference for the application of point cloud target detection methods. SUMMARY

[0003] The present application aims to at least partially solve one of the above technical problems. To this end, one object of the present application is to provide a weakly supervised target detection model training method, which extracts candidate boxes from the density features of point clouds in a certain area and performs target detection through weak supervision only when part of the input point clouds are labeled, thereby reducing the cost of manpower and time.

[0004] A second object of the present application is to provide a computer readable storage medium.

[0005] A third object of the present application is to provide a computer device.

[0006] A fourth object of the present application is to provide a weakly supervised target detection model training device.

[0007] To achieve the above objects, the first aspect of the present application provides a weakly supervised target detection model training method, comprising the following steps: obtaining original three-dimensional point cloud data and partially labeled three-dimensional point cloud data in a target environment; presetting a plurality of anchor boxes and selecting corresponding first candidate boxes according to the point cloud density corresponding to the plurality of anchor boxes; dynamically adjusting the corresponding first candidate boxes to screen the corresponding first candidate boxes to obtain corresponding weak pseudo labels; screening the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes, and purifying pseudo labels according to real labeled boxes to obtain strong pseudo labels; training a pre-set neural network according to the weak pseudo labels and the strong pseudo labels to obtain a trained target detection model.

[0008] According to the weakly supervised target detection model training method provided in the embodiments of the present application, the anchor points are preset, the anchor points and the point cloud of the input frame in the anchor points are projected to two dimensions, the corresponding candidate frame is obtained according to the density of the point cloud, the candidate frame obtained is screened through the attenuation of height and distance and used as a pseudo label, and the pseudo label and the part of the real label in each frame are used for the training of the point cloud target detection network to obtain the final target detection model. Thus, the labor and time costs are reduced.

[0009] In addition, the weakly supervised target detection model training method provided in the above embodiments of the present application can further have the following additional technical features.

[0010] Optionally, the plurality of anchor point frames are preset, and the corresponding first candidate frame is obtained by screening the point cloud density corresponding to the plurality of anchor point frames, including: within a preset distance, two anchor points in the vertical direction are respectively set with the same center, and the anchor points and the corresponding original three-dimensional point cloud data are projected to a square area through an interpolation method to obtain the anchor point frame; and it is judged whether the point cloud density corresponding to the anchor point frame exceeds a first threshold value, if yes, the anchor point frame is retained, and if no, the anchor point frame is discarded.

[0011] Optionally, the corresponding first candidate frame is dynamically adjusted, including: judging whether the point cloud density in the corresponding first candidate frame is within a preset threshold range; if yes, the first candidate frame is translated according to the projection distance of the original three-dimensional point cloud data in the anchor point.

[0012] Optionally, the original three-dimensional point cloud data and the part of the labeled three-dimensional point cloud data are screened to obtain the corresponding second candidate frame, including: a plurality of anchor point frames are preset, and the corresponding first candidate frame is obtained by screening the point cloud density corresponding to the plurality of anchor point frames, and the anchor point frame after the first screening is dynamically adjusted to be screened for the second time to obtain the corresponding second candidate frame.

[0013] Optionally, the pseudo label is purified according to the real label frame to obtain a strong pseudo label, including: the point cloud coordinates in the part of the labeled real label frame are moved, rotated or flipped to obtain an enhanced pseudo label, so as to filter the second candidate frame to obtain the strong pseudo label.

[0014] Optionally, when the pre-set neural network is trained according to the weak pseudo label and the strong pseudo label, the loss calculation is further performed according to the intersection over union of the weak pseudo label and the strong pseudo label.

[0015] To achieve the above object, the second embodiment of the present application provides a computer readable storage medium, which stores a weakly supervised target detection model program, and the weakly supervised target detection model program is executed by a processor to realize the weakly supervised target detection model method as described above.

[0016] To achieve the above object, the third aspect of the present application provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to realize the weakly supervised target detection model method as described above.

[0017] To achieve the above object, the fourth aspect of the present application provides a weakly supervised target detection model device, comprising: an acquisition module configured to acquire original three-dimensional point cloud data and partially labeled three-dimensional point cloud data in a target environment; a first screening module configured to preset a plurality of anchor boxes and screen the corresponding point cloud density according to the plurality of anchor boxes to obtain corresponding first candidate boxes; a second screening module configured to dynamically adjust the corresponding first candidate boxes to screen the corresponding first candidate boxes to obtain corresponding weak pseudo labels; a purification module configured to screen and process the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes, and purify the pseudo labels according to the real labeled boxes to obtain strong pseudo labels; and a training module configured to train a pre-set neural network according to the weak pseudo labels and the strong pseudo labels to obtain a trained target detection model; thereby reducing the labor and time costs.

[0018] In addition, the weakly supervised target detection model training device according to the above embodiments of the present application can also have the following additional technical features:

[0019] Optionally, the first screening module is further configured to set two anchor points in the vertical direction with the same center within a preset distance, and project the anchor points and the corresponding original three-dimensional point cloud data together into a square region by an interpolation method to obtain the anchor boxes; and determine whether the point cloud density corresponding to the anchor boxes exceeds a first threshold value, if yes, keep it, and if not, discard it. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 Flowchart of the weakly supervised target detection model training method according to an embodiment of the present application;

[0021] Figure 2 Flowchart of the weakly supervised target detection model training method according to an embodiment of the present application;

[0022] Figure 3 Schematic diagram of the first candidate box generation according to an embodiment of the present application;

[0023] Figure 4 Schematic diagram of the effect of the obtained weak pseudo labels according to an embodiment of the present application;

[0024] Figure 5 A schematic diagram of a training block for strong pseudo-labels according to an embodiment of the present invention;

[0025] Figure 6 This is a schematic diagram illustrating the effect of a strong pseudo-label obtained according to an embodiment of the present invention;

[0026] Figure 7 This is a block diagram of a target detection model training device based on weak supervision according to an embodiment of the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0028] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art.

[0029] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0030] Figure 1 This is a flowchart illustrating a weakly supervised target detection model training method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the training method for this weakly supervised object detection model includes the following steps:

[0031] S101, acquire the original 3D point cloud data and partially labeled 3D point cloud data in the target environment.

[0032] It should be noted that the original 3D point cloud data is completely unlabeled, while partially labeled 3D point cloud data can be labeled using some targets in the same frame, or all targets in some frames can be labeled, while all point clouds in the remaining frames are unlabeled. Labeling refers to manually selecting target objects. As an example, the target object can be a vehicle. Suppose there are 20 vehicles in the current environment. You can select 5 vehicles in the same frame and leave the remaining 15 unselected, or you can select all 20 vehicles in some frames and leave them unselected in the remaining frames.

[0033] S102, a plurality of anchor boxes are preset, and a first candidate box corresponding to a point cloud density of the plurality of anchor boxes is selected.

[0034] As an embodiment, the plurality of anchor boxes are preset, and the first candidate box corresponding to the point cloud density of the plurality of anchor boxes is selected, including: within a preset distance, two anchors in a vertical direction are respectively set with a same center, and the anchors and corresponding original three-dimensional point cloud data are projected into a square region by an interpolation method to obtain the anchor boxes; and whether the point cloud density corresponding to the anchor boxes exceeds a first threshold value is judged, and if yes, the anchor boxes are retained, and if no, the anchor boxes are discarded.

[0035] That is, a 3D anchor point is preset, and a higher-scoring anchor point is selected as an output candidate box through prior knowledge and density information.

[0036] As a specific embodiment, as shown in Figure 3 Since the positions of the vehicles in the data set are within a certain height range and there is no stacking phenomenon, anchor points are preset at certain intervals within the range, and two anchors in 0-degree and 90-degree directions are respectively set with a same center; the point cloud density in the candidate box can indicate whether an object is contained therein, and a high density represents a high confidence; the three-dimensional point cloud data is projected onto a front view plane, which can avoid the problem of sparse point cloud in the distance due to the sensor; the projection of the preset anchor points is adjusted to a square region of H*H by an interpolation method, and the point cloud density is represented as N / H^2 if there are N points in the region; since the anchor points are preset and one-to-one correspond to the square projected onto the two-dimensional plane, the number of points in the square can reflect the number of points in the anchor point; a threshold value is set according to the point cloud density in the square, and the part exceeding the threshold value is considered to contain an object, and a certain range is retained, and the candidate box with a low density is considered to have no object and is discarded.

[0037] That is, the corresponding point cloud is framed by presetting the anchor points, and the anchor points and the corresponding point cloud are projected into a two-dimensional square region to form a plurality of square anchor boxes.

[0038] S103, the corresponding first candidate box is dynamically adjusted to screen the corresponding first candidate box to obtain a corresponding weak pseudo label.

[0039] As an embodiment, the corresponding first candidate box is dynamically adjusted, including: judging whether the point cloud density in the corresponding first candidate box is within a preset threshold range; if yes, the first candidate box is translated according to the projection distance of the original three-dimensional point cloud data in the anchor point.

[0040] It should be noted that since the anchor points are preset according to the distance, there is a high probability that the same object cannot be included in the same candidate box, but such conditions are generally concentrated near the edges of the square; therefore, according to this point, the anchor points can be translated in a small range so that they are better aligned with the points they contain.

[0041] As a specific embodiment, the square box with a point cloud density within a certain threshold range is appropriately translated in the x and y axes according to the point projection distance within the anchor point to include points that may be the same object; the dynamically adjusted square box can filter out most of the preset anchor points, and the filtering effect is shown in Figure 4

[0042] S104, filtering the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain a corresponding second candidate box, and purifying the pseudo label according to the real label box to obtain a strong pseudo label.

[0043] As an embodiment, filtering the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain a corresponding second candidate box includes: presetting a plurality of anchor boxes, and performing a first filtering according to the point cloud density corresponding to the plurality of anchor boxes, and dynamically adjusting the anchor boxes after the first filtering to perform a second filtering to obtain a corresponding second candidate box.

[0044] It should be noted that the filtering process of this step is the same as the filtering processes of steps S102 and S103 described above, except that the input data includes partially labeled three-dimensional point cloud data, so this will not be described again.

[0045] As an embodiment, purifying the pseudo label according to the real label box to obtain a strong pseudo label includes: moving, rotating or flipping the point cloud coordinates within the partially labeled real label box to obtain an enhanced pseudo label, so as to filter the second candidate box to obtain a strong pseudo label.

[0046] It should be noted that the candidate boxes generated by the above steps S102 and S103 through point cloud density will have a lot of noise, since the high-density positions are not only cars, but also some other objects, and it is difficult to avoid including some non-car objects, and since the anchor points are preset according to the distance and angle, even if the range is expanded and slightly translated, there will be some deviations in position and angle from the real car, affecting the accuracy of target detection.

[0047] Since a weakly supervised method is used, only partially accurate GT boxes are used, therefore, data augmentation needs to be performed on the input label box to generate more labels.

[0048] ​New data is created by moving or rotating or flipping the points within the real bounding box; specifically, the points are translated, i.e. the coordinates of the points are changed; the label of the bounding box is randomly rotated within a certain range, i.e. the direction of the points is changed, and the rotation angle is [-1.57, +1.57].

[0049] Meanwhile, the results of several epochs can also be enhanced during the training process, i.e. the output features of each layer during the training process are also used as enhanced pseudo labels for training.

[0050] In addition, after the coordinates of the points are modified, a collision detection experiment needs to be performed, i.e. because the scenario in which a car coincides with other objects does not exist in a real scene, all cars are on the same plane, so it is necessary to prevent the situation in which the newly generated bounding box coincides with other bounding boxes.

[0051] Thus, more strong pseudo label boxes are generated by performing data enhancement on the input accurate labels.

[0052] That is, as shown in Figure 5 , the candidate boxes obtained after the screening strategy of steps S102 and S103 using the frames with partial manual labels are subjected to ROI Pooling, the features of the candidate boxes are extracted more finely, and the pseudo labels are purified. The detailed process is as follows: first, the candidate boxes generated in steps S102 and S103 are divided into g sub-grids, then the features of each grid are extracted using Pointnet, the features of these sub-grids are aggregated as the features of the entire candidate box, and the output is obtained through MLP as the finally generated strong pseudo label box. The effect diagram is shown in Figure 6 .

[0053] In S105, the pre-set neural network is trained according to the weak pseudo labels and the strong pseudo labels to obtain a trained target detection model.

[0054] As an embodiment, when the pre-set neural network is trained according to the weak pseudo labels and the strong pseudo labels, the loss is also calculated according to the intersection over union (IOU) of the weak pseudo labels and the strong pseudo labels.

[0055] As a specific embodiment, the IOU matching is performed according to the weak pseudo labels and the strong pseudo labels:

[0056] The car matching IOU threshold is greater than or equal to 0.6 as a positive sample, and less than 0.45 as a negative sample, and the middle one is not calculated for loss.

[0057] The total loss function is defined as follows:

[0058]

[0059] wherein positive sample after passing through a softmax function, negative sample after passing through a softmax function; u i and The loss is calculated only for the positive sample regression, and a cross-entropy loss function is used. and The classification loss is used for regularization of each term; and are two balance coefficients, and 1.5 and 1 are used respectively; and the regression loss uses a SmoothL1 function.

[0060] In summary, as Figure 2 shown in the figure, according to the weakly supervised target detection model training method of the embodiment of the application, the preset anchor point is obtained, and the anchor point and the point cloud of the input frame in the anchor point are projected to two dimensions, the corresponding candidate frame is obtained according to the density of the point cloud; the obtained candidate frame is screened by height and distance attenuation, and used as a pseudo label, and the point cloud target detection network is trained with the labeled part of the real label in each frame to obtain the final target detection model; thus, the manpower and time cost are reduced.

[0061] In order to realize the above-mentioned embodiment, the embodiment of the application provides a computer readable storage medium, which stores a weakly supervised target detection model training program, and the weakly supervised target detection model training program is executed by a processor to realize the weakly supervised target detection model training method as described above.

[0062] According to the computer readable storage medium of the embodiment of the application, the weakly supervised target detection model training program is stored, so that when the processor executes the weakly supervised target detection model training program, the weakly supervised target detection model training method as described above is realized, and thus the preset anchor point is obtained, and the anchor point and the point cloud of the input frame in the anchor point are projected to two dimensions, the corresponding candidate frame is obtained according to the density of the point cloud; the obtained candidate frame is screened by height and distance attenuation, and used as a pseudo label, and the point cloud target detection network is trained with the labeled part of the real label in each frame to obtain the final target detection model, thereby reducing the manpower and time cost.

[0063] In order to realize the above-mentioned embodiment, the embodiment of the application provides a computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the weakly supervised target detection model training method as described above is realized.

[0064] The computer device according to the embodiment of the present application stores the weakly supervised target detection model training program in the memory, so that the processor implements the weakly supervised target detection model training method as described above when executing the weakly supervised target detection model training program. Thus, the corresponding candidate frame is obtained by presetting the anchor point, projecting the anchor point and the point cloud of the input frame in the anchor point to two dimensions, and screening according to the density of the point cloud. The obtained candidate frame is screened by the height and distance attenuation and used as a pseudo label. The pseudo label and the part of the real label in each frame are used for training the point cloud target detection network to obtain the final target detection model, thereby reducing the labor and time cost.

[0065] To implement the above embodiment, the embodiment of the present application further provides a weakly supervised target detection model training device, as shown in the figure. Figure 7 The weakly supervised target detection model training device comprises an acquisition module 10, a first screening module 20, a second screening module 30, a purification module 40 and a training module 50.

[0066] The acquisition module 10 is configured to acquire original three-dimensional point cloud data and part of labeled three-dimensional point cloud data in a target environment. The first screening module 20 is configured to preset a plurality of anchor point frames and screen the corresponding point cloud density to obtain corresponding first candidate frames. The second screening module 30 is configured to dynamically adjust the corresponding first candidate frames to screen the corresponding first candidate frames to obtain corresponding weak pseudo labels. The purification module 40 is configured to screen the original three-dimensional point cloud data and part of the labeled three-dimensional point cloud data to obtain corresponding second candidate frames, and purify the pseudo labels according to the real label frame to obtain strong pseudo labels. The training module 50 is configured to train the pre-set neural network according to the weak pseudo labels and the strong pseudo labels to obtain a trained target detection model.

[0067] As an embodiment, the first screening module 20 is further configured to set two anchor points in the vertical direction with the same center within a preset distance, and project the anchor points and the corresponding original three-dimensional point cloud data together to a square area by an interpolation method to obtain an anchor point frame. It is determined whether the point cloud density corresponding to the anchor point frame exceeds a first threshold. If yes, the anchor point frame is retained. If no, the anchor point frame is discarded.

[0068] It should be noted that the above description and example of the weakly supervised target detection model training method are also applicable to the weakly supervised target detection model training device of the present embodiment, which will not be described here.

[0069] In summary, according to the weakly supervised target detection model device based on the embodiment of the present application, the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data in the target environment are acquired by the acquisition module; the first screening module pre-sets a plurality of anchor point boxes, and screens according to the point cloud density corresponding to the plurality of anchor point boxes to obtain corresponding first candidate boxes; the second screening module dynamically adjusts the corresponding first candidate boxes, so as to screen the corresponding first candidate boxes to obtain corresponding weak pseudo labels; the purification module screens and processes the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes, and purifies the pseudo labels according to the real labeled boxes to obtain strong pseudo labels; the training module trains the pre-set neural network according to the weak pseudo labels and the strong pseudo labels to obtain the trained target detection model; thus, the labor and time costs are reduced.

[0070] Those skilled in the art will appreciate that embodiments of the present application can be supplied as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage media, etc.) having computer-usable program code embodied therein.

[0071] The present application is described in reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce an apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in the flowcharts and / or block diagrams.

[0072] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that implements the functions specified in the flowcharts and / or block diagrams.

[0073] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable devices provide the function of implementing the processes specified in the flowcharts Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0074] It should be noted that the use of any of the terms "first", "second" or the like used in the description and / or in the claims is merely intended to distinguish between similar objects going back to a common technical feature. It is within the scope of the application that a single one of the objects specified with "first" or "second" can be replaced by the other of these terms specifying the same object. The singular forms "a", "an" and "the" include plural referents unless the context clearly dictates otherwise.

[0075] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments without departing from the spirit and scope of the application. Therefore, it should be understood that the appended claims are intended to cover all such modifications and variations as falling within the scope of the application.

[0076] It is apparent that a person skilled in the art can make various changes and modifications to the application without departing from the spirit and scope thereof. Therefore, if these modifications and changes of the application fall within the scope of the claims and their equivalents, it is intended to include them in the application.

[0077] In the description of the application, it should be understood that the terms "first", "second" are used only for descriptive purposes and are not to be construed as indicating or implying relative importance or an indicated number of technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited.

[0078] In the present application, unless specifically defined otherwise, the terms "mounting", "connected", "connection", "fixed", and the like should be construed broadly and do not necessarily mean fixedly connected, but can mean removably connected, or integral; can mean mechanical connection, or electrical connection; can mean direct connection, or indirect connection via an intermediate medium; can mean internal communication between two elements, or interaction between two elements. Those skilled in the art can understand the specific meaning of the above terms in the present application according to the specific circumstances.

[0079] In the present application, unless specifically defined otherwise, the first feature is "on" or "under" the second feature can be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, the first feature "above", "over" and "on" the second feature can be that the first feature is directly above or obliquely above the second feature, or only means that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "under" and "under" the second feature can be that the first feature is directly below or obliquely below the second feature, or only means that the horizontal height of the first feature is less than that of the second feature.

[0080] In the description of the present application, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present application, the illustrative description of the above terms should not be understood as necessarily referring to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the present application and the features of the different embodiments or examples without contradiction.

[0081] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for training a weakly supervised object detection model, the method comprising: The method comprises the following steps: Obtaining original three-dimensional point cloud data and partially labeled three-dimensional point cloud data in a target environment; Presetting a plurality of anchor point boxes and screening according to the point cloud density corresponding to the plurality of anchor point boxes to obtain corresponding first candidate boxes; Performing dynamic adjustment on the corresponding first candidate boxes to screen the corresponding first candidate boxes to obtain corresponding weak pseudo labels; Screening the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes, and purifying pseudo labels according to real label boxes to obtain strong pseudo labels; Training a preset neural network according to the weak pseudo labels and the strong pseudo labels to obtain a trained target detection model; The presetting a plurality of anchor point boxes and screening according to the point cloud density corresponding to the plurality of anchor point boxes to obtain corresponding first candidate boxes comprises: Within a preset distance, two anchor points in the vertical direction are set with the same center, and the anchor points and the corresponding original three-dimensional point cloud data are projected into a square region by an interpolation method to obtain the anchor point boxes; Judging whether the point cloud density corresponding to the anchor point boxes exceeds a first threshold value, and if yes, retaining, and if not, discarding; The dynamic adjustment on the corresponding first candidate boxes comprises: Judging whether the point cloud density in the corresponding first candidate box is within a preset threshold range; If yes, performing translation on the first candidate box according to the projection distance of the original three-dimensional point cloud data in the anchor point; The screening of the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes comprises: Presetting a plurality of anchor point boxes, screening according to the point cloud density corresponding to the plurality of anchor point boxes for the first time, and dynamically adjusting the anchor point boxes after the first screening to perform second screening to obtain corresponding second candidate boxes; The purifying of pseudo labels according to real label boxes to obtain strong pseudo labels comprises: Moving, rotating or flipping the point cloud coordinates in the partially labeled real label boxes to obtain enhanced pseudo labels, and filtering the second candidate boxes to obtain strong pseudo labels. 2.The weakly supervised object detection model training method of claim 1, wherein, When the preset neural network is trained according to the weak pseudo labels and the strong pseudo labels, loss calculation is further performed according to the intersection over union of the weak pseudo labels and the strong pseudo labels.

3. A computer-readable storage medium, characterized in that, A processor executes a weakly supervised target detection model training program stored thereon to implement the weakly supervised target detection model training method according to any one of claims 1-2.

4. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the weakly supervised target detection model training method according to any one of claims 1-2.

5. A weakly supervised based target detection model training apparatus, characterized in that, The method comprises: An acquisition module is configured to acquire original three-dimensional point cloud data and partially labeled three-dimensional point cloud data in a target environment; A first screening module is configured to preset a plurality of anchor point boxes and screen according to the point cloud density corresponding to the plurality of anchor point boxes to obtain corresponding first candidate boxes; A second screening module is configured to screen the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain corresponding second candidate boxes, and purify pseudo labels according to real label boxes to obtain strong pseudo labels; The second screening module is configured to dynamically adjust the corresponding first candidate box, so as to screen the corresponding first candidate box to obtain a corresponding weak pseudo label; The purification module is configured to screen the original three-dimensional point cloud data and the partially labeled three-dimensional point cloud data to obtain a corresponding second candidate box, and purify a pseudo label according to a real label box to obtain a strong pseudo label; The training module is configured to train a pre-set neural network according to the weak pseudo label and the strong pseudo label to obtain a trained target detection model; The first screening module is further configured to: project the anchor points and the corresponding original three-dimensional point cloud data together into a square region by an interpolation method to obtain the anchor box; determine whether the point cloud density corresponding to the anchor box exceeds a first threshold value, and if yes, retain the anchor box, and if not, discard the anchor box; The second screening module is further configured to: determine whether the point cloud density in the corresponding first candidate box is within a preset threshold range; if yes, translate the first candidate box according to the projection distance of the original three-dimensional point cloud data in the anchor point; The purification module is further configured to: preset a plurality of anchor boxes, and perform a first screening on the plurality of anchor boxes according to the point cloud density corresponding to the plurality of anchor boxes, and dynamically adjust the anchor boxes after the first screening to perform a second screening to obtain a corresponding second candidate box; move, rotate or flip the point cloud coordinates in the partially labeled real label box to obtain an enhanced pseudo label, so as to filter the second candidate box to obtain a strong pseudo label.

Images