LiDAR 3D Target Detection Methods and Related Equipment

Abstract

This invention provides a 3D target detection method and related equipment using lidar. The method includes: voxelizing point cloud data acquired by lidar; adding global features to each voxel using a global attention-based voxel feature encoding method to establish relationships between voxels; obtaining feature maps through a feature fusion network employing a dilated pyramid pooling mechanism and a channel attention mechanism; and performing target detection using a CenterHead detection head on the obtained feature maps. The beneficial effects of this invention are: the introduction of global attention for voxel feature encoding results in richer feature representations, while the dilated pyramid pooling mechanism based on channel attention further expands the receptive field, enabling the network to better understand the details and contextual relationships of the target at different scales. The combination of these two methods effectively improves the detection performance for small targets.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a lidar 3D target detection method and related equipment. Background Technology

[0002] LiDAR-based 3D object detection is a key technology in autonomous driving. It utilizes LiDAR sensors to acquire point cloud data of the environment and analyzes and processes this data to detect and identify objects in 3D space. Traditional LiDAR-based object detection methods primarily rely on the geometric features of point cloud data, such as point cloud clustering, object segmentation, and feature extraction. These methods typically use traditional computer vision techniques, such as filtering, clustering, and classification algorithms, to process the point cloud data and achieve object detection through object modeling and feature matching. However, with the development of deep learning technology, LiDAR-based 3D object detection is gradually shifting towards end-to-end object detection using deep neural networks.

[0003] Voxel-based 3D object detection algorithms are currently one of the mainstream methods. They transform unordered point cloud data into ordered voxel representations, then extract features through a 3D convolutional network to obtain the object's feature information. Voxelization not only preserves the feature information of the point cloud but also effectively improves the network's processing speed. Currently, CenterPoint is a popular voxel-based 3D object detection algorithm. It first voxels the point cloud (i.e., divides it into grids of a certain size on the XYZ plane), then encodes the features of each voxel using meanVFE (mean Voxel Feature Encoding). Next, 3D object detection is achieved through a convolutional network and the Center detection head. Because the CenterPoint network uses voxelized representations and a regular feature structure, it can achieve high-speed object detection through parallel computation and efficient inference algorithms. This gives the CenterPoint network advantages in real-time applications and high-efficiency computing environments.

[0004] Currently, 3D object detection algorithms typically employ average voxel encoding and dynamic average voxel encoding to encode voxel features. However, in practical applications, separating each voxel's points leads to a loss of connection between points and between voxels, which is detrimental to the network. Simultaneously, the receptive field of the network is also a significant factor affecting network performance; traditional 3D object detection feature fusion networks struggle to acquire a large receptive field. Summary of the Invention

[0005] The technical problem to be solved by this invention is: a 3D target detection method and related equipment for lidar, which aims to solve the technical problem of insufficient receptive field of the network.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a 3D target detection method using lidar, comprising the following steps:

[0007] The point cloud data acquired by the lidar is processed into voxels;

[0008] A global attention-based voxel feature encoding method is used to add global features to each voxel, thereby establishing a connection between voxels;

[0009] Feature maps are obtained by using a feature fusion network that employs a hollow pyramid pooling mechanism and a channel attention mechanism.

[0010] The obtained feature map is then used for target detection using the CenterHead detection head.

[0011] Furthermore, a global attention-based voxel feature encoding method is used to add global features to each voxel, establishing connections between voxels. Specifically, this includes...

[0012] F of all voxels v = (N,5,4) is reconstructed into a two-dimensional tensor F by combining the second and third dimension features. v1 = (N,20), two-dimensional tensor F v1 It contains features of global points, where N represents the number of voxels;

[0013] Tensor F v1 The result obtained after multilayer perceptron is related to F. m Tensors F of the same size v2 = (N,4);

[0014] The tensor F obtained by average voxel encoding m With tensor F v2 The summation yields the voxel feature F, which has global characteristics. o = (N, 4).

[0015] Furthermore, the lidar 3D target detection method also includes the following steps:

[0016] The offset of each point from the centroid is added to the features of the LiDAR points, resulting in the input feature set as follows: Where x i ,y i ,z i v represents the coordinates of each point. x ,v y ,v z For each body mass coordinate, R 7 Represents a seven-dimensional tensor;

[0017] The feature tensor F is obtained through average voxel encoding.m1 = (N, 7);

[0018] The feature tensor F m1 = (N,7) is used to obtain the feature tensor F through average voxel encoding based on a global attention mechanism. v3 = (N,4);

[0019] F m2 With F v3 The summation yields the voxel feature F, which has global characteristics. o1 = (N, 7).

[0020] Furthermore, obtaining feature maps by employing a hollow pyramid pooling mechanism specifically includes,

[0021] For the features F obtained through the backbone network 16x Perform a 3x3 residual convolution block to obtain feature F. in Among them, F in =δB((C(F) 16x )));

[0022] In the above formula, C represents a 3x3 convolution operation, B represents the BatchNorm normalization operation, and δ represents the ReLU activation function.

[0023] Then F in Through five different branches, four of which are dilated convolutions with different dilation scales (3, 6, 12, and 18 respectively), and 3x3 kernels, the features obtained after convolution are: F dilation_3 ,F dilation_6 ,F dilation_12 ,F dilation_18 ;in:

[0024] F dilation_3 =C dilation＝3 (F in )

[0025] F dilation_6 =C dilation＝6 (F in )

[0026] F dilation_12 =C dilation＝12 (F in )

[0027] F dilation_18 =C dilation＝18 (F in )

[0028] In the above formula, C dilation_3This represents a dilated convolution operation with padding of 3, dilation of 3, and a 3x3 kernel; C dilation_6 This represents a dilated convolution operation with padding of 6, dilation of 6, and a 3x3 kernel; C dilation_12 This represents a dilated convolution operation with padding of 12, dilation of 12, and a 3x3 kernel; C dilation_18 This represents a dilated convolution operation with padding of 18, dilation of 18, and a kernel of 3x3.

[0029] Then F in The feature F is obtained through another branch of a convolutional layer with a kernel size of 1x1. 1x1 ;

[0030] Finally, all features are fused together and dimensionality is reduced using a 1x1 convolutional layer to obtain F. out The formula is shown below:

[0031]

[0032] In the above formula, C p1 This represents a convolution operation with a 1x1 kernel. This represents the tensor concatenation operation.

[0033] Furthermore, the feature map acquisition method using a feature fusion network employing a channel attention mechanism specifically includes:

[0034] The feature map F obtained from the backbone network 16x Through two branches, one branch is to take F 16x After max pooling, and then through a multilayer perceptron, F is obtained. max_out ;

[0035] F max_out =L(M(F) 16x ))

[0036] In the above formula, M represents the max pooling layer operation, and L represents the multilayer perceptron (MLP) operation.

[0037] The other branch is to take F 16x F is obtained after passing through an average pooling layer and then a multilayer perceptron. avg_out ;

[0038] F avg_out =L(A(F) 16x ))

[0039] In the above formula, A represents the average pooling layer operation, and L represents the multilayer perceptron (MLP) operation;

[0040] The obtained feature F max_out and feature F avg_out The summation, followed by the sigmoid activation function, yields the feature F. channel_out The formula is shown below:

[0041] F channel_out =S(F max_out +F avg_out )

[0042] In the above formula, S represents the sigmoid activation function operation;

[0043] The obtained feature F channel_out As a weight, it is combined with F obtained in the previous step. out Multiplying them together yields the final feature F. out_16 The formula is shown below:

[0044]

[0045] In the above formula, This indicates that the tensor is multiplied element by element.

[0046] Furthermore, the lidar 3D target detection method also includes the following steps:

[0047] Feature F 8x Become feature F out_8 ;

[0048] Feature F out_16 and feature F 8x The feature F of the input detection head is obtained by fusion; the formula is shown below:

[0049]

[0050] In the above formula, This represents the tensor concatenation operation, C p2 δ represents a convolution operation with a 1x1 kernel, B represents the BatchNorm normalization operation, and δ represents the ReLU activation function operation.

[0051] The present invention also provides a lidar 3D target detection device, comprising:

[0052] The voxelization processing module is used to perform voxelization processing on the point cloud data acquired by the lidar.

[0053] The voxel relationship establishment module is used to add global features to each voxel using a global attention-based voxel feature encoding method, so that voxels can establish relationships between voxels.

[0054] The feature map detection module is used to obtain feature maps by employing a feature fusion network with a hollow pyramid pooling mechanism and a channel attention mechanism.

[0055] The target detection module is used to perform target detection using the CenterHead detection head on the obtained feature map.

[0056] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the laser radar 3D target detection method as described above.

[0057] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, can implement the lidar 3D target detection method as described above.

[0058] The beneficial effects of this invention are as follows: by introducing global attention to the feature encoding of voxels, a richer feature representation is brought about, while the hollow pyramid pooling mechanism based on channel attention further expands the receptive field, enabling the network to better understand the details and contextual relationships of the target at different scales. The combination of the two effectively improves the detection performance of small targets. Attached Figure Description

[0059] The specific structure of the present invention will now be described in detail with reference to the accompanying drawings.

[0060] Figure 1 This is a flowchart of the lidar 3D target detection method according to an embodiment of the present invention;

[0061] Figure 2 This is a schematic diagram of the channel attention void pyramid pooling method according to an embodiment of the present invention;

[0062] Figure 3 This is a structural diagram of the channel attention module according to an embodiment of the present invention;

[0063] Figure 4 The results of the BEV view test in this embodiment of the invention;

[0064] Figure 5 These are test results of adding the invention to different voxel networks according to embodiments of the invention;

[0065] Figure 6 This is a schematic block diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0068] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0069] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0070] like Figure 1 As shown, an embodiment of the present invention is: a 3D target detection method using lidar, comprising the following steps:

[0071] S10. Perform voxelization processing on the point cloud data acquired by the lidar;

[0072] S20. Use a global attention-based voxel feature encoding method to add global features to each voxel, so that a connection is established between voxels.

[0073] S30. Feature maps are obtained by using a feature fusion network with a hollow pyramid pooling mechanism and a channel attention mechanism.

[0074] S40. Use the CenterHead detection head to perform target detection on the obtained feature map.

[0075] In one specific embodiment, step S20, adding global features to each voxel using a global attention-based voxel feature encoding method to establish connections between voxels, specifically includes:

[0076] F of all voxels v = (N,5,4) is reconstructed into a two-dimensional tensor F by combining the second and third dimension features. v1 = (N,20), two-dimensional tensor F v1 It contains features of global points, where N represents the number of voxels;

[0077] Tensor F v1 The result obtained after multilayer perceptron is related to F. m Tensors F of the same sizev2 = (N,4);

[0078] The tensor F obtained by average voxel encoding m With tensor F v2 The summation yields the voxel feature F, which has global characteristics. o = (N, 4).

[0079] In this scheme, the global-mean Voxel Encoding (global-meanVFE) based on the global attention mechanism adds global features to the mean Voxel Encoding (meanVFE).

[0080] In voxel representation, V = {p} i =[x i ,y i ,z i ,r i ] T ∈R 4} i＝1...t Let a non-empty voxel contain t≤T lidar points, where p i Including the XYZ coordinates of the i-th point, r i It is the received reflectivity, R 4 This represents a four-dimensional tensor. Typically, in mean Voxel Encoding (meanVFE), the value of t is 5, indicating that at most 5 LiDAR points are retained within a voxel. Therefore, the feature representation of a voxel is a three-dimensional tensor F. v = (N, 5, 4), where N represents the number of voxels. Then, the features of these five points are summed and averaged to obtain the feature F of a single point. m = (N,4), where the characteristics of this point represent the characteristics of this voxel.

[0081] In one specific embodiment, the lidar 3D target detection method further includes the step of...

[0082] The offset of each point from the centroid is added to the features of the LiDAR points, resulting in the input feature set as follows: Where x i ,y i ,z i v represents the coordinates of each point. x ,v y ,v z For each body mass coordinate, R 7 Represents a seven-dimensional tensor;

[0083] The feature tensor F is obtained through average voxel encoding. m1 = (N, 7);

[0084] The feature tensor F m1= (N,7) is used to obtain the feature tensor F through average voxel encoding based on a global attention mechanism. v3 = (N,4);

[0085] F m2 With F v3 The summation yields the voxel feature F, which has global characteristics. o1 = (N, 7).

[0086] In this scheme, the relationship between global points is considered by average voxel encoding based on the global attention mechanism, and then the relationship between voxels is constructed to obtain a more refined voxel feature representation.

[0087] In one specific embodiment, such as Figure 2 As shown, step S30, obtaining the feature map by employing the hollow pyramid pooling mechanism, specifically includes:

[0088] For the features F obtained through the backbone network 16x Perform a 3x3 residual convolution block to obtain feature F. in Among them, F in =δB((C(F) 16x )));

[0089] In the above formula, C represents a 3x3 convolution operation, B represents the BatchNorm normalization operation, and δ represents the ReLU activation function.

[0090] Then F in Through five different branches, four of which are dilated convolutions with different dilation scales (3, 6, 12, and 18 respectively), and 3x3 kernels, the features obtained after convolution are: F dilation_3 ,F dilation_6 ,F dilation_12 ,F dilation_18 ;in:

[0091] F dilation_3 =C dilation＝3 (F in )

[0092] F dilation_6 =C dilation＝6 (F in )

[0093] F dilation_12 =C dilation＝12 (F in )

[0094] F dilation_18 =C dilation＝18 (F in )

[0095] In the above formula, C dilation_3 This represents a dilated convolution operation with padding of 3, dilation of 3, and a 3x3 kernel; C dilation_6 This represents a dilated convolution operation with padding of 6, dilation of 6, and a 3x3 kernel; C dilation_12 This represents a dilated convolution operation with padding of 12, dilation of 12, and a 3x3 kernel; C dilation_18 This represents a dilated convolution operation with padding of 18, dilation of 18, and a kernel of 3x3.

[0096] Then F in The feature F is obtained through another branch of a convolutional layer with a kernel size of 1x1. 1x1 ;

[0097] Finally, all features are fused together and dimensionality is reduced using a 1x1 convolutional layer to obtain F. out The formula is shown below:

[0098]

[0099] In the above formula, C p1 This represents a convolution operation with a 1x1 kernel. This represents the tensor concatenation operation.

[0100] In one specific embodiment, such as Figure 3 As shown, step S30, obtaining feature maps through a feature fusion network employing a channel attention mechanism, specifically includes:

[0101] The feature map F obtained from the backbone network 16x Through two branches, one branch is to take F 16x After max pooling, and then through a multilayer perceptron, F is obtained. max_out ;

[0102] F max_out =L(M(F) 16x ))

[0103] In the above formula, M represents the max pooling layer operation, and L represents the multilayer perceptron (MLP) operation.

[0104] The other branch is to take F 16x F is obtained after passing through an average pooling layer and then a multilayer perceptron. avg_out ;

[0105] F avg_out =L(A(F) 16x ))

[0106] In the above formula, A represents the average pooling layer operation, and L represents the multilayer perceptron (MLP) operation;

[0107] The obtained feature F max_out and feature F avg_out The summation, followed by the sigmoid activation function, yields the feature F. channel_out The formula is shown below:

[0108] F channel_out =S(F max_out +F avg_out )

[0109] In the above formula, S represents the sigmoid activation function operation;

[0110] The obtained feature F channel_out As a weight, it is combined with F obtained in the previous step. out Multiplying them together yields the final feature F. out_16 The formula is shown below:

[0111]

[0112] In the above formula, This indicates that the tensor is multiplied element by element.

[0113] In this technical solution, the following was also added to the hollow space convolutional pooling pyramid: Figure 3 The channel attention mechanism shown.

[0114] In one specific embodiment, the lidar 3D target detection method further includes the step of...

[0115] Following the two steps above, feature F can be obtained. 16x Become feature F out_16 Similarly, feature F can be... 8x Become feature F out_8 ;

[0116] Feature F out_16 and feature F 8x The feature F of the input detection head is obtained by fusion; the formula is shown below:

[0117]

[0118] In the above formula, This represents the tensor concatenation operation, C p2 δ represents a convolution operation with a 1x1 kernel, B represents the BatchNorm normalization operation, and δ represents the ReLU activation function operation.

[0119] The lidar 3D target detection method in the above embodiments significantly improves the accuracy of small target detection without sacrificing detection speed by employing global average voxel encoding to add global features to each voxel and increasing the network receptive field through the hollow pyramid pooling module with channel attention.

[0120] The method proposed in this embodiment of the invention is tested below. All test data are based on the KITTI dataset, including 7481 training samples and 7518 test samples. To evaluate model performance, the training samples are divided into 3712 training sets and 3769 validation sets. Throughout the training process, only the LiDAR point cloud data from the training sets is used, and training and validation are performed on three categories: cars, pedestrians, and bicycles. The evaluation for each category is divided into three difficulty levels: easy, medium, and hard, based on target size, truncation, and occlusion status. The average accuracy (AP_R40) is used to evaluate the target detection results for each of the three categories, and the average AP (mAP) for the medium difficulty level is calculated as the overall performance metric of the model.

[0121] Testing was conducted using the PyTorch framework, with training and evaluation performed on a hardware environment equipped with a 3090Ti GPU and an Intel i7-12700K CPU. The network was trained for 80 epochs using the Adam optimizer, with a mini-batch size of 16. The initial learning rate was 0.006, dynamically adjusted using a OneCycle strategy. The maximum number of points per voxel was limited to 5.

[0122] like Figure 4 As shown, the evaluation results of this invention in bird's-eye view (BEV) are presented and compared with other existing object detection algorithms. The results show that this invention demonstrates a significant advantage over the baseline model Centerpoint, with the mAP value increasing from 68.57% to 73.41%. Particularly noteworthy is the over 10% improvement in AP value for small object detection.

[0123] In addition, such as Figure 5 The diagram shows the detection results when step one of the present invention is added to other voxel networks. It can be seen that in the pedestrian category of medium difficulty, the detection accuracy is improved by 1.28%, 1.10%, and 2.36%, respectively, further validating the effectiveness of the proposed method in small target detection.

[0124] Another embodiment of the present invention is a lidar 3D target detection device, comprising:

[0125] The voxelization processing module is used to perform voxelization processing on the point cloud data acquired by the lidar.

[0126] The voxel relationship establishment module is used to add global features to each voxel using a global attention-based voxel feature encoding method, so that voxels can establish relationships between voxels.

[0127] The feature map detection module is used to obtain feature maps by employing a feature fusion network with a hollow pyramid pooling mechanism and a channel attention mechanism.

[0128] The target detection module is used to perform target detection using the CenterHead detection head on the obtained feature map.

[0129] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned LiDAR 3D target detection device can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0130] The aforementioned lidar 3D target detection device can be implemented as a computer program, which can, for example... Figure 6 It runs on the computer device shown.

[0131] Please see Figure 6 , Figure 6 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0132] See Figure 6 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0133] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a lidar 3D target detection method.

[0134] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0135] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a laser radar 3D target detection method.

[0136] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0137] The processor 502 is used to run the computer program 5032 stored in the memory to implement the lidar 3D target detection method as described above.

[0138] It should be understood that, in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0139] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0140] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the lidar 3D target detection method as described above.

[0141] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0142] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0143] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0144] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for 3D target detection using lidar, characterized in that, Including the following steps: The point cloud data acquired by the lidar is processed into voxels; A global attention-based voxel feature encoding method is used to add global features to each voxel, thereby establishing a connection between voxels; Feature maps are obtained by using a feature fusion network that employs a hollow pyramid pooling mechanism and a channel attention mechanism. The obtained feature map is used for target detection using the CenterHead detection head; Specifically, the method of adding global features to each voxel using a global attention-based voxel feature encoding method to establish relationships between voxels includes... Features of all voxels By combining the second and third-dimensional features, a two-dimensional tensor is reconstructed. Two-dimensional tensor It contains features of global points, where N represents the number of voxels; tensor The result obtained after multilayer perceptron Tensors of the same size ; Tensor obtained by average voxel encoding With tensor Adding them together yields voxel features with global characteristics. ; Specifically, obtaining feature maps using the hollow pyramid pooling mechanism includes... Features obtained through the backbone network Perform a The residual convolutional blocks yield features ;in, ; In the above formula, C represents Convolution operation, This indicates the BatchNorm normalization operation. Represents the ReLU activation function; Then The convolution uses five different branches, four of which are dilated convolutions with different levels of holes. Their padding and dilation scales are 3, 6, 12, and 18, respectively, and the convolution kernels are all... The features obtained after convolution are as follows: ;in: In the above formula, This indicates that the padding is 3, the dilation is 3, and the convolution kernel is... The dilated convolution operation; This indicates that the padding is 6, the dilation is 6, and the convolution kernel is... The dilated convolution operation; This indicates that the padding is 12, the dilation is 12, and the convolution kernel is... The dilated convolution operation; This indicates that the padding is 18, the dilation is 18, and the convolution kernel is... The dilated convolution operation; Then Through another convolution kernel of size The branches of the convolutional layer are used to obtain features. ; Finally, all features are fused together using a convolutional kernel. Dimensionality reduction of the convolutional layer is obtained The formula is shown below: In the above formula, This indicates that the convolution kernel is... Convolution operation, This represents the tensor concatenation operation.

2. The lidar 3D target detection method as described in claim 1, characterized in that: It also includes steps, The offset of each point from the centroid is added to the features of the LiDAR points, resulting in the input feature set as follows: ,in Coordinates of each physical fitness level , Represents a seven-dimensional tensor; The feature tensor is obtained through average voxel encoding. ; feature tensor The feature tensor is obtained through average voxel encoding based on a global attention mechanism. ; Will and Adding them together yields voxel features with global characteristics. .

3. The lidar 3D target detection method as described in claim 2, characterized in that: The feature map acquisition method using a feature fusion network employing a channel attention mechanism specifically includes... Feature maps obtained from the backbone network Through two branch roads, one of which is... After max pooling, and then through a multilayer perceptron, the result is obtained. ; In the above formula, This indicates the max pooling layer operation. This represents the operation of a multilayer perceptron (MLP). Another branch is to After passing through an average pooling layer, and then through a multilayer perceptron, the result is obtained. ; In the above formula, This indicates the average pooling layer operation. This represents the operation of a multilayer perceptron (MLP). The obtained features and characteristics The features are obtained by summing the ingredients and then applying the sigmoid activation function. The formula is shown below: In the above formula, This represents the sigmoid activation function operation; The obtained features As a weight, and obtained from the previous step Multiply to obtain the final feature. The formula is shown below: In the above formula, This indicates that the tensor is multiplied element by element.

4. The lidar 3D target detection method as described in claim 3, characterized in that: It also includes steps, Features Become a feature ; Features and characteristics The features of the input detection head are obtained by fusion. The formula is shown below: In the above formula, This represents the tensor concatenation operation. Indicates that the convolution kernel is Convolution operation, This indicates the BatchNorm normalization operation. This indicates the ReLU activation function operation.

5. A lidar 3D target detection device, characterized in that, include: The voxelization processing module is used to perform voxelization processing on the point cloud data acquired by the lidar. The voxel relationship establishment module is used to add global features to each voxel using a global attention-based voxel feature encoding method, so as to establish a relationship between voxels. The feature map detection module is used to obtain feature maps by employing a feature fusion network with a hollow pyramid pooling mechanism and a channel attention mechanism. The target detection module is used to perform target detection on the obtained feature map using the CenterHead detection head; The voxel relationship establishment module is specifically used for: Features of all voxels By combining the second and third-dimensional features, a two-dimensional tensor is reconstructed. Two-dimensional tensor It contains features of global points, where N represents the number of voxels; tensor The result obtained after multilayer perceptron Tensors of the same size ; Tensor obtained by average voxel encoding With tensor Adding them together yields voxel features with global characteristics. ; The feature map detection module is specifically used for: Features obtained through the backbone network Perform a The residual convolutional blocks yield features ;in, ; In the above formula, C represents Convolution operation, This indicates the BatchNorm normalization operation. Represents the ReLU activation function; Then The convolution uses five different branches, four of which are dilated convolutions with different levels of holes. Their padding and dilation scales are 3, 6, 12, and 18, respectively, and the convolution kernels are all... The features obtained after convolution are as follows: ;in: In the above formula, This indicates that the padding is 3, the dilation is 3, and the convolution kernel is... The dilated convolution operation; This indicates that the padding is 6, the dilation is 6, and the convolution kernel is... The dilated convolution operation; This indicates that the padding is 12, the dilation is 12, and the convolution kernel is... The dilated convolution operation; This indicates that the padding is 18, the dilation is 18, and the convolution kernel is... The dilated convolution operation; Then Through another convolution kernel of size The branches of the convolutional layer are used to obtain features. ; Finally, all features are fused together using a convolutional kernel. Dimensionality reduction of the convolutional layer is obtained The formula is shown below: In the above formula, This indicates that the convolution kernel is... Convolution operation, This represents the tensor concatenation operation.

6. A computer device, characterized in that: The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the lidar 3D target detection method as described in any one of claims 1 to 4.

7. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, can implement the lidar 3D target detection method as described in any one of claims 1 to 4.

Images