A target detection method and device for compressed point cloud and electronic equipment

Abstract

The present disclosure provides a target detection method and device for compressed point cloud and an electronic device. A first target detection model taking an original point cloud as input is used as a teacher model, and a second target detection model taking a compressed point cloud as input is used as a student model. A feature distillation loss between a feature extraction network for extracting a point cloud feature map in the first target detection model and the second target detection model is determined. An attention distillation loss between a spatial attention module for assigning a corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model and the second target detection model is determined. The second target detection model is trained under the supervision of the feature distillation loss and the attention distillation loss. The compressed point cloud to be detected is input into the trained second target detection model, and a target object in the compressed point cloud to be detected is determined. The detection performance of target detection on the compressed point cloud can be improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of target detection, and in particular, to a target detection method and device for compressed point cloud and electronic equipment. BACKGROUND

[0002] With the development of 3D acquisition technology, three-dimensional data obtained by sensors can provide rich geometric, shape and scale information, which is complementary to 2D images, and 3D data provides an opportunity for machines to better understand the surrounding environment. 3D data has been widely used in different fields, including autonomous driving, robots, remote sensing and medical treatment. Point cloud, as a common form of 3D data, retains the original geometric information in three-dimensional space. Therefore, point cloud is the preferred data form for many scene understanding related applications, such as autonomous driving and robots. Target detection is an important branch of computer vision, and 3D target detection generates a directed 3D box around each detected target with scene point cloud as input. The current popular point cloud target detection technology is mostly based on point or voxel to extract features, and then sends the features to the detection head to complete the detection.

[0003] Compared with image target detection, laser radar point cloud provides reliable depth information, which can be used for accurate positioning of targets and description of target shapes. However, due to non-uniform sampling of three-dimensional space, limited perception range of sensors, occlusion and relative pose, etc., laser radar point cloud is sparse and the point density is uneven. In addition, the amount of point cloud data is large, and compression processing is often required in actual processing and transmission. After compression processing, the data loss introduced by compression further reduces the performance of target detection, which also makes it difficult to deploy related algorithms to actual applications. The current point cloud target detection method is mostly performed on the original point cloud, and this kind of method will have obvious performance degradation after being migrated to the point cloud after compression processing. SUMMARY

[0004] The embodiments of the present disclosure at least provide a target detection method, device and electronic equipment for compressed point cloud, which can improve the detection performance of target detection for point cloud after compression processing, improve the accuracy and detection quality of target detection.

[0005] The embodiments of the present disclosure provide a target detection method for compressed point cloud, comprising:

[0006] A first target detection model taking original point cloud as input is provided, and a second target detection model taking compressed point cloud as input is provided;

[0007] determine a feature distillation loss between the feature extraction network for extracting the point cloud feature map in the first target detection model and the second target detection model;

[0008] determine an attention distillation loss between the spatial attention module for assigning a corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model and the second target detection model;

[0009] train the second target detection model under the supervision of the feature distillation loss and the attention distillation loss;

[0010] input a to-be-detected compressed point cloud into the trained second target detection model, and determine a target object in the to-be-detected compressed point cloud.

[0011] In an optional implementation, the first target detection model comprises:

[0012] a first feature extraction network, a first spatial attention module, and a first detection head connected in sequence.

[0013] In an optional implementation, the first feature extraction network is configured to perform feature extraction on the input original point cloud, and determine a corresponding original point cloud feature map.

[0014] The first spatial attention module is configured to assign a corresponding spatial weight to each pixel in the original point cloud feature map.

[0015] The first detection head is configured to perform prediction on the original point cloud feature map to which the spatial weight is assigned through a plurality of convolution layers, and determine a target object included in the original point cloud.

[0016] In an optional implementation, the second target detection model comprises:

[0017] a second feature extraction network, a second spatial attention module, a channel attention module, and a second detection head connected in sequence, wherein the second detection head is a best transmission detection head based on label assignment.

[0018] In an optional implementation, the second feature extraction network is configured to perform feature extraction on the input compressed point cloud, and determine a corresponding compressed point cloud feature map.

[0019] The second spatial attention module is configured to assign a corresponding spatial weight to each pixel in the compressed point cloud feature map.

[0020] The channel attention module is configured to assign a corresponding channel weight to each channel in the compressed point cloud feature map.

[0021] The second detection head is configured to predict the compressed point cloud feature map to which the spatial weight and the channel weight are assigned through a multi-layer convolutional layer, and determine a target object included in the compressed point cloud.

[0022] In an optional implementation, the feature distillation loss is configured to supervise the second feature extraction network to imitate the process of extracting the point cloud feature map in the first target detection model.

[0023] In an optional implementation, the attention distillation loss is configured to supervise the second spatial attention module to imitate the process of assigning the corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model.

[0024] The embodiments of the present disclosure further provide a target detection device for compressed point clouds, comprising:

[0025] A model construction module is configured to provide a first target detection model taking an original point cloud as input, and a second target detection model taking a compressed point cloud as input;

[0026] A feature distillation module is configured to take the first target detection model as a teacher model, and take the second target detection model as a student model, and determine a feature distillation loss between feature extraction networks for extracting a point cloud feature map in the first target detection model and the second target detection model;

[0027] An attention distillation module is configured to determine an attention distillation loss between spatial attention modules for assigning a corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model and the second target detection model;

[0028] A training module is configured to train the second target detection model under the supervision of the feature distillation loss and the attention distillation loss;

[0029] A detection module is configured to input a compressed point cloud to be detected into the trained second target detection model, and determine a target object in the compressed point cloud to be detected.

[0030] The embodiments of the present disclosure further provide an electronic device, comprising a processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, and the machine readable instructions are executed by the processor to perform the above target detection method for compressed point clouds, or the steps in any one of the possible implementation manners of the above target detection method for compressed point clouds.

[0031] The embodiment of the present disclosure further provides a computer readable storage medium, which stores a computer program. When the computer program is run by a processor, the steps of the above method for target detection on compressed point clouds or any possible implementation manner of the above method for target detection on compressed point clouds are performed.

[0032] The embodiment of the present disclosure further provides a computer program product, which comprises computer programs / instructions. When the computer programs / instructions are executed by a processor, the steps of the above method for target detection on compressed point clouds or any possible implementation manner of the above method for target detection on compressed point clouds are implemented.

[0033] The embodiment of the present disclosure provides a method, device and electronic equipment for target detection on compressed point clouds. The method comprises the following steps: providing a first target detection model taking an original point cloud as input, and providing a second target detection model taking a compressed point cloud as input; taking the first target detection model as a teacher model and the second target detection model as a student model, determining a feature distillation loss between feature extraction networks for extracting a point cloud feature map in the first target detection model and the second target detection model; determining an attention distillation loss between spatial attention modules for assigning corresponding spatial weights to each pixel of the point cloud feature map in the first target detection model and the second target detection model; training the second target detection model under the supervision of the feature distillation loss and the attention distillation loss; and inputting a compressed point cloud to be detected into the trained second target detection model to determine a target object in the compressed point cloud to be detected. The detection performance of target detection on the compressed point cloud can be improved, and the accuracy and detection quality of target detection can be improved.

[0034] In order to make the above objectives, features and advantages of the present disclosure more apparent, the following will specifically describe preferred embodiments in conjunction with the attached drawings, and make a detailed description as follows. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following will briefly introduce the drawings needed to be used in the embodiments. The drawings herein are incorporated into the specification and form a part of the specification, which illustrate the embodiments consistent with the present disclosure, and are used to explain the technical solutions of the present disclosure together with the specification. It should be understood that the following drawings only show some embodiments of the present disclosure, and therefore should not be considered as a limitation to the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor.

[0036] Figure 1A flow chart of a target detection method for compressed point cloud provided by an embodiment of the present disclosure is shown.

[0037] Figure 2 A schematic diagram of a target detection network based on feature attention guidance provided by an embodiment of the present disclosure is shown.

[0038] Figure 3 A schematic diagram of a target detection device for compressed point cloud provided by an embodiment of the present disclosure is shown.

[0039] Figure 4 A schematic diagram of an electronic device provided by an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be described below in connection with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all the embodiments. The components of the embodiments of the present disclosure described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present disclosure.

[0041] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0042] The term “and / or” herein only describes an association relationship, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone. In addition, the term “at least one” herein means any one of multiple or any combination of at least two of multiple, for example, including at least one of A, B and C can mean including any one or more elements selected from the set consisting of A, B and C.

[0043] It is found through research that, compared with image target detection, laser radar point cloud provides reliable depth information and can be used for accurately positioning targets and describing target shapes. However, due to factors such as non-uniform sampling of three-dimensional space, limited perception range of a sensor, occlusion, and relative pose, laser radar point cloud is sparse and has uneven point density. In addition, point cloud data is large, and is often compressed in actual processing and transmission. After compression, the data loss introduced by compression further reduces the performance of target detection, which also makes it difficult to deploy related algorithms to actual applications. Current point cloud target detection methods are mostly performed on original point cloud, and such methods will have obvious performance degradation after being migrated to compressed point cloud.

[0044] Based on the above research, the present disclosure provides a target detection method, device and electronic equipment for compressed point cloud. A first target detection model with original point cloud as input and a second target detection model with compressed point cloud as input are provided. The first target detection model is used as a teacher model, and the second target detection model is used as a student model. A feature distillation loss between feature extraction networks for extracting point cloud feature maps in the first target detection model and the second target detection model is determined. An attention distillation loss between spatial attention modules for assigning corresponding spatial weights to each pixel of the point cloud feature map in the first target detection model and the second target detection model is determined. The second target detection model is trained under the supervision of the feature distillation loss and the attention distillation loss. The target object in the compressed point cloud to be detected is determined by inputting the compressed point cloud to be detected into the trained second target detection model. The detection performance of target detection for compressed point cloud can be improved, and the accuracy and detection quality of target detection can be improved.

[0045] To facilitate understanding of the present embodiment, first, a target detection method for compressed point cloud disclosed by the present embodiment is described in detail. The execution subject of the target detection method for compressed point cloud provided by the present embodiment is generally a computer device with certain computing power, which may, for example, include a terminal device or a server or other processing device. The terminal device may, for example, be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementation manners, the target detection method for compressed point cloud may be realized by a processor calling computer readable instructions stored in a memory.

[0046] Referring to Figure 1As shown, a flowchart of a target detection method for compressed point clouds provided by an embodiment of the present disclosure is shown, and the method comprises steps S101-S105, wherein:

[0047] S101, a first target detection model taking an original point cloud as input is provided, and a second target detection model taking a compressed point cloud as input is provided.

[0048] In a specific implementation, a first target detection model taking an uncompressed original point cloud as input is constructed to locate and describe the target included in the original point cloud, and a second target detection model taking a compressed point cloud after compression processing as input is constructed to locate and describe the target included in the original point cloud.

[0049] Specifically, the first target detection model comprises a first feature extraction network, a first spatial attention module and a first detection head connected in sequence.

[0050] Here, the first feature extraction network is used to extract features for the input original point cloud to determine the corresponding original point cloud feature map; the first spatial attention module is used to assign a corresponding spatial weight to each pixel in the original point cloud feature map; and the first detection head is used to predict the original point cloud feature map after assigning the spatial weight through a plurality of convolution layers to determine the target object included in the original point cloud.

[0051] Further, the second target detection model comprises a second feature extraction network, a second spatial attention module, a channel attention module and a second detection head connected in sequence, wherein the second detection head is a best transmission detection head based on label assignment.

[0052] Among them, the self-attention method based on Transformer has shown extremely advanced performance in some natural language processing and 2D object detection tasks, therefore, the channel attention module can be a channel attention module based on Transformer to obtain good channel weights at a global level.

[0053] Here, the second target detection model is an end-to-end two-stage 3D laser radar point cloud detection network. In an actual two-stage target detection algorithm, the obtained proposal frame in the first stage is discarded in part. From the existing target detection research, if the proposal frame is discarded according to a fixed IOU threshold, not every useful prediction frame can be assigned to a label frame (GT). In addition, due to external reasons such as object size and shape occlusion, some proposal frames are classified as ambiguous frames. These special frames are elements that can improve the generalization ability of the model. In the case of data degradation, due to the reduction of information in the data, the effective proposal frame is more likely to be discarded. Therefore, in order to improve the use of prediction frames, the inventors introduce the label assignment optimal transmission (OTA) proposed for 2D images into the detection head, and the second detection head is a detection head based on label assignment optimal transmission.

[0054] The second feature extraction network is configured to perform feature extraction on the input compressed point cloud to determine a corresponding compressed point cloud feature map. The second spatial attention module is configured to assign a corresponding spatial weight to each pixel in the compressed point cloud feature map. The channel attention module is configured to assign a corresponding channel weight to each channel in the compressed point cloud feature map. The second detection head is configured to perform prediction on the compressed point cloud feature map after the spatial weight and the channel weight are assigned by using a plurality of convolution layers, to determine a target object included in the compressed point cloud.

[0055] S102, taking the first target detection model as a teacher model and the second target detection model as a student model, a feature distillation loss between feature extraction networks in the first target detection model and the second target detection model for extracting a point cloud feature map is determined.

[0056] In a specific implementation, a target detection network based on feature attention guidance is constructed by taking the first target detection model as a teacher model and the second target detection model as a student model. As shown in Figure 2 , it is a schematic diagram of a target detection network based on feature attention guidance provided by an embodiment of the present disclosure.

[0057] Here, in the target detection network based on feature attention guidance, a feature distillation loss between a processing behavior of the first feature extraction network in the first target detection model for performing feature extraction on the input original point cloud to determine a corresponding original point cloud feature map and a processing behavior of the second feature extraction network in the second target detection model for performing feature extraction on the input compressed point cloud to determine a corresponding compressed point cloud feature map is determined.

[0058] The feature distillation loss is configured to supervise the second feature extraction network in the second target detection model to imitate the action process of the first feature extraction network in the first target detection model for extracting a point cloud feature map.

[0059] S103, determine the attention distillation loss between the spatial attention modules in the first target detection model and the second target detection model for assigning corresponding spatial weights to each pixel of the point cloud feature map.

[0060] In a specific implementation, the attention distillation loss between the process of assigning corresponding spatial weights to each pixel in the original point cloud feature map by the first spatial attention module in the first target detection model and the process of assigning corresponding spatial weights to each pixel in the compressed point cloud feature map by the second spatial attention module in the second target detection model.

[0061] The attention distillation loss is used to supervise the second spatial attention module to imitate the process of assigning corresponding spatial weights to each pixel of the point cloud feature map in the first target detection model.

[0062] S104, train the second target detection model under the supervision of the feature distillation loss and the attention distillation loss.

[0063] In a specific implementation, the second target detection model is trained under the supervision of the feature distillation loss and the attention distillation loss, so that the second target detection model processing the compressed point cloud imitates the processing process of the first target detection model processing the original point cloud, thereby reducing the performance decline of target detection caused by data loss due to compression.

[0064] Specifically, in the compression case, due to the point consistency and the jump of the point position, the attribute features extracted by the existing method will become unreliable, because the distortion caused by compression will cause the points to have different degrees of deviation, thereby reducing the detection accuracy. Therefore, under the supervision of the feature distillation loss, the second target detection model taking the compressed point cloud as input can imitate the feature extraction behavior of the detector taking the original point cloud as input as much as possible, thereby obtaining better feature extraction capability.

[0065] Further, under the supervision of the attention distillation loss, the network pays more attention to the key stable points and suppresses the unstable points, so that the second target detection model can obtain a similar attention map as the first target detection model processing the original point cloud when processing the compressed point cloud. In the case of compressed data degradation, the effective attention mechanism can emphasize the region where the object may exist and reduce the attention of the network to the distortion region.

[0066] As a possible implementation, the loss function of the target detection network based on feature attention guidance, i.e., the target detection network taking the first target detection model as the teacher model and the second target detection model as the student model, can be calculated by the following formula:

[0067] Cost ij =L reg (P i ,G i )+L cls (P i ,G i )

[0068] wherein, Cost ij represents a loss function; L reg (P i ,G i ) represents a position loss; L cls (P i , G i ) represents a classification loss; P represents a target detection result; and G represents a ground truth label.

[0069] S105, inputting the to-be-detected compressed point cloud into the trained second target detection model to determine a target object in the to-be-detected compressed point cloud.

[0070] In a specific implementation, the to-be-detected compressed point cloud is input into the trained second target detection model, a feature extraction network in the second target detection model is used to extract features of the input to-be-detected compressed point cloud to determine a corresponding compressed point cloud feature map; a second spatial attention module is used to assign a corresponding spatial weight to each pixel in the compressed point cloud feature map; a channel attention module is used to assign a corresponding channel weight to each channel in the compressed point cloud feature map; and a second detection head is used to predict the compressed point cloud feature map after the spatial weight and the channel weight are assigned, to determine a position and a shape of a target object included in the to-be-detected compressed point cloud through a plurality of convolutional layers.

[0071] The target detection method for compressed point clouds provided by the embodiments of the present disclosure includes the following steps: providing a first target detection model taking an original point cloud as input, and a second target detection model taking a compressed point cloud as input; taking the first target detection model as a teacher model and the second target detection model as a student model, determining a feature distillation loss between a feature extraction network in the first target detection model and a feature extraction network in the second target detection model, determining an attention distillation loss between a spatial attention module in the first target detection model and a spatial attention module in the second target detection model, which are used to assign a corresponding spatial weight to each pixel in a point cloud feature map, training the second target detection model under the supervision of the feature distillation loss and the attention distillation loss, and inputting a to-be-detected compressed point cloud into the trained second target detection model to determine a target object in the to-be-detected compressed point cloud. The detection performance of target detection on the compressed point cloud can be improved, and the accuracy and detection quality of target detection can be improved.

[0072] Those skilled in the art can understand that, in the above method of the specific implementation, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process, and the specific execution order of each step should be determined according to its function and possible internal logic.

[0073] Based on the same inventive concept, the disclosure embodiments also provide a target detection device for compressed point clouds corresponding to the target detection method for compressed point clouds. Since the principle of solving problems in the device of the disclosure embodiments is similar to the above-mentioned target detection method for compressed point clouds of the disclosure embodiments, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described here.

[0074] Please refer to Figure 3 , Figure 3 A schematic diagram of a target detection device for compressed point clouds provided by the disclosure embodiments. As shown in Figure 3 The target detection device for compressed point clouds provided by the disclosure embodiments 300 includes:

[0075] The model construction module 310 is configured to provide a first target detection model taking an original point cloud as input, and a second target detection model taking a compressed point cloud as input.

[0076] The feature distillation module 320 is configured to take the first target detection model as a teacher model, and take the second target detection model as a student model, and determine a feature distillation loss between feature extraction networks for extracting point cloud feature maps in the first target detection model and the second target detection model.

[0077] The attention distillation module 330 is configured to determine an attention distillation loss between spatial attention modules for assigning corresponding spatial weights to each pixel of the point cloud feature map in the first target detection model and the second target detection model.

[0078] The training module 340 is configured to train the second target detection model under the supervision of the feature distillation loss and the attention distillation loss.

[0079] The detection module 350 is configured to input a compressed point cloud to be detected into the trained second target detection model, and determine a target object in the compressed point cloud to be detected.

[0080] The description of the processing flow of each module in the device and the interaction flow between the modules can be referred to the related description in the above-mentioned method embodiments, which will not be described in detail here.

[0081] The device for target detection on compressed point cloud provided by the embodiment of the present disclosure comprises: a first target detection model taking the original point cloud as input, and a second target detection model taking the compressed point cloud as input; the first target detection model is taken as a teacher model, and the second target detection model is taken as a student model; a feature distillation loss between a feature extraction network for extracting a point cloud feature map in the first target detection model and the second target detection model is determined; an attention distillation loss between a spatial attention module for assigning a corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model and the second target detection model is determined; the second target detection model is trained under the supervision of the feature distillation loss and the attention distillation loss; and a target object in a to-be-detected compressed point cloud is determined by inputting the to-be-detected compressed point cloud into the trained second target detection model. The detection performance of target detection on the compressed point cloud can be improved, and the accuracy and detection quality of target detection can be improved.

[0082] Corresponding to Figure 1 The embodiment of the present disclosure also provides an electronic device 400, as shown in Figure 4 The structure schematic diagram of the electronic device 400 provided by the embodiment of the present disclosure comprises:

[0083] The processor 41, the memory 42, and the bus 43; the memory 42 is used for storing execution instructions, including the internal memory 421 and the external memory 422; the internal memory 421 is also called the internal memory, and is used for temporarily storing operation data in the processor 41 and data exchanged with the external memory 422 such as a hard disk; the processor 41 exchanges data with the external memory 422 through the internal memory 421; when the electronic device 400 is running, the processor 41 and the memory 42 communicate through the bus 43, so that the processor 41 executes Figure 1 The steps of the target detection method on the compressed point cloud in the embodiment of the present disclosure.

[0084] The embodiment of the present disclosure also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to perform the steps of the target detection method on the compressed point cloud described in the above method embodiment. The storage medium can be a volatile or non-volatile computer readable storage medium.

[0085] The embodiment of the present disclosure also provides a computer program product comprising computer instructions, which can execute the steps of the target detection method on the compressed point cloud described in the above method embodiment when executed by a processor. For details, please refer to the above method embodiment, which will not be described here.

[0086] The computer program product can be implemented by hardware, software or a combination thereof. In an optional embodiment, the computer program product is embodied in a computer storage medium. In another optional embodiment, the computer program product is embodied in a software product, such as a software development kit (SDK) or the like.

[0087] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the apparatus described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here. In several embodiments provided in the present disclosure, it should be understood that the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some communication interfaces, and can be electrical, mechanical or other forms.

[0088] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment scheme.

[0089] In addition, each functional unit in each embodiment of the present disclosure can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0090] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a nonvolatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present disclosure essentially or the part that contributes to the prior art or 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 number 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 methods described in various embodiments of the present disclosure. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0091] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present disclosure, used to illustrate the technical solutions of the present disclosure, and not to limit them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art who is familiar with the technology in the art can still make modifications or easily think of changes to the technical solutions described in the foregoing embodiments, or make equivalent replacements to some of the technical features; and these modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A method for object detection on compressed point clouds, characterized in that, The method comprises the following steps: providing a first target detection model taking an original point cloud as input, and a second target detection model taking a compressed point cloud as input; determining a feature distillation loss between a feature extraction network for extracting a point cloud feature map in the first target detection model and the second target detection model, taking the first target detection model as a teacher model and the second target detection model as a student model; determining an attention distillation loss between a spatial attention module for assigning a corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model and the second target detection model; training the second target detection model under the supervision of the feature distillation loss and the attention distillation loss; inputting a to-be-detected compressed point cloud into the trained second target detection model to determine a target object in the to-be-detected compressed point cloud.

2. The method of claim 1, wherein, The first target detection model comprises: a first feature extraction network, a first spatial attention module and a first detection head connected in sequence.

3. The method of claim 2, wherein: the first feature extraction network is configured to perform feature extraction on the input original point cloud to determine a corresponding original point cloud feature map; the first spatial attention module is configured to assign a corresponding spatial weight to each pixel in the original point cloud feature map; the first detection head is configured to perform prediction on the original point cloud feature map after the spatial weight is assigned through a plurality of convolutional layers to determine a target object included in the original point cloud.

4. The method of claim 1, wherein, The second target detection model comprises: a second feature extraction network, a second spatial attention module, a channel attention module and a second detection head connected in sequence, wherein the second detection head is a best transmission detection head based on label assignment.

5. The method of claim 4, wherein: the second feature extraction network is configured to perform feature extraction on the input compressed point cloud to determine a corresponding compressed point cloud feature map; the second spatial attention module is configured to assign a corresponding spatial weight to each pixel in the compressed point cloud feature map; the channel attention module is configured to assign a corresponding channel weight to each channel in the compressed point cloud feature map; the second detection head is configured to perform prediction on the compressed point cloud feature map after the spatial weight and the channel weight are assigned through a plurality of convolutional layers to determine a target object included in the compressed point cloud.

6. The method of claim 4, wherein: the feature distillation loss is configured to supervise the second feature extraction network to imitate the process of extracting the point cloud feature map in the first target detection model.

7. The method of claim 4, wherein: the attention distillation loss is configured to supervise the second spatial attention module to imitate the process of assigning the corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model.

8. A target detection device for compressed point clouds, characterized in that, The method comprises the following steps: a model construction module configured to provide a first target detection model taking an original point cloud as input, and a second target detection model taking a compressed point cloud as input; a feature distillation module configured to determine a feature distillation loss between a feature extraction network for extracting a point cloud feature map in the first target detection model and the second target detection model, with the first target detection model as a teacher model and the second target detection model as a student model; an attention distillation module configured to determine an attention distillation loss between a spatial attention module for assigning a corresponding spatial weight to each pixel of the point cloud feature map in the first target detection model and the second target detection model; a training module configured to train the second target detection model under the supervision of the feature distillation loss and the attention distillation loss; a detection module configured to input a to-be-detected compressed point cloud into the trained second target detection model to determine a target object in the to-be-detected compressed point cloud.

9. An electronic device, comprising: comprise: a processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine readable instructions executed by the processor perform the steps of the target detection method for compressed point clouds according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, which is executed by the processor to perform the steps of the target detection method for compressed point clouds according to any one of claims 1 to 7.

Images