3D small target detection method and system based on laser radar

Abstract

The invention provides a 3D small target detection method and system based on a laser radar, and belongs to the field of computer vision. The original point cloud information and the point cloud pillar information are input into a trained point-enhanced attention perception pillar network, point cloud features and pillar features are fused through a point-pillar cross attention module, and fused pillar features are extracted; based on the fused pillar features, dynamically adjusting feature weights from feature dimensions and point number dimensions through a pillar double-attention module, and extracting noise reduction optimization pillar features; based on noise reduction optimization pillar features, fusing different resolution feature information through a multi-resolution set abstraction module, and extracting multi-scale features; and performing target classification and position regression based on the multi-scale features to obtain a 3D small target detection result. The original point cloud features and various attention mechanisms are utilized to enhance the pillar features, and the problem of context and multi-resolution information loss in an existing pillar-based method is relieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, and in particular to a 3D small target detection method and system based on a laser radar. BACKGROUND

[0002] Point cloud three-dimensional object detection based on a laser radar is a core technology for realizing target category recognition and position positioning by collecting point cloud information such as three-dimensional coordinates and reflection intensity of targets in space through a laser radar, and has an irreplaceable role in the fields of automatic driving, intelligent security, etc. In particular, in the automatic driving scenario, small dynamic targets such as pedestrians and cyclists need to be accurately identified relying on this technology, which is directly related to driving safety and system reliability.

[0003] Compared with image two-dimensional detection which only needs to process planar pixel information, the inherent sparsity (low proportion of effective points in space), disorder (no fixed arrangement order of point clouds) and uneven distribution (large difference in point density between target edges and background) characteristics of point clouds make three-dimensional detection need to overcome more complex feature extraction problems. With the increasing demand for small target detection accuracy in automatic driving, how to solve the problem of missed detection and false detection of small targets such as pedestrians and cyclists due to less point cloud information and easy background interference has become a key direction for current technical breakthrough.

[0004] In the prior art, point-based methods directly perform neighborhood search and feature aggregation on points, which can retain more complete geometric information, but a large number of additional parameters result in low computational efficiency and large memory consumption, making it difficult to meet the real-time application requirements. Although the voxel pillar-based method can greatly reduce parameters and improve efficiency by gridizing point clouds into columnar charts and mapping them into two-dimensional pseudo images, the local context information of small targets is easily lost in the discretization process of original point clouds, and multi-resolution features cannot be considered at the same time, ultimately resulting in difficulty in meeting the requirements of small target detection accuracy and high safety performance of automatic driving. SUMMARY

[0005] To solve the above problems, the present application provides a 3D small target detection method and system based on a laser radar, which combines the context information of original point clouds with the inherent features of pillars to carry out three-dimensional object detection, thereby retaining both the local information extracted by point-based methods and the global information captured by pillar-based methods.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a 3D small target detection method based on a laser radar, comprising: obtaining laser radar original point cloud information and point cloud pillar information to be detected; input the original point cloud information and the point cloud pillar information into the trained point-enhanced attention perception pillar network, wherein the point cloud features and the pillar features are fused through a point-pillar cross attention module to extract fusion pillar features; based on the fusion pillar features, the feature weights are dynamically adjusted from the feature dimension and the point number dimension through a pillar double attention module to extract denoising optimization pillar features; based on the denoising optimization pillar features, different resolution feature information is fused through a multi-resolution set abstraction module to extract multi-scale features; target classification and position regression are performed based on the multi-scale features to obtain a 3D small target detection result.

[0007] In a second aspect, the present application provides a 3D small target detection system based on a laser radar, comprising: a data acquisition unit configured to acquire original point cloud information and point cloud pillar information of a laser radar to be detected; a feature extraction unit configured to input the original point cloud information and the point cloud pillar information into a trained point-enhanced attention perception pillar network, wherein the point cloud features and the pillar features are fused through a point-pillar cross attention module to extract fusion pillar features; based on the fusion pillar features, the feature weights are dynamically adjusted from the feature dimension and the point number dimension through a pillar double attention module to extract denoising optimization pillar features; based on the denoising optimization pillar features, different resolution feature information is fused through a multi-resolution set abstraction module to extract multi-scale features; a target detection unit configured to perform target classification and position regression based on the multi-scale features to obtain a 3D small target detection result.

[0008] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to implement the steps of the 3D small target detection method based on a laser radar according to the first aspect.

[0009] In a fourth aspect, the present application provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the 3D small target detection method based on a laser radar according to the first aspect.

[0010] Compared with the prior art, the present application has the following advantages: This invention utilizes a point-pillar cross-attention module to effectively fuse point cloud features and pillar features. The extracted association-enhanced pillar features fully integrate the local details of the point cloud with the global representation of the pillars, laying a solid foundation for subsequent detection. The pillar dual-attention module dynamically adjusts weights from both feature and point count dimensions. Based on the association-enhanced pillar feature extraction, noise-reduced and optimized pillar features suppress noise interference, strengthen effective features, and improve feature quality. The multi-resolution set abstraction module fuses feature information from different resolutions. Based on the noise-reduced and optimized pillar feature extraction, multi-scale fusion target features are extracted, taking into account the feature representation of small targets at different scales, enabling accurate capture of small targets at all scales. Finally, classification and regression are performed based on the multi-scale fusion target features, significantly improving the accuracy and positioning precision of 3D small target detection. This allows for more efficient and accurate identification of various 3D small targets, providing strong technical support for fields such as autonomous driving and intelligent security that rely on LiDAR detection.

[0011] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0012] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0013] Figure 1 The main flowchart of a 3D small target detection method based on lidar provided in an embodiment of the present invention; Figure 2 A detailed flowchart of a 3D small target detection method based on lidar provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the point-column cross-attention module provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0015] Example 1 like Figure 1 As shown, this embodiment discloses a 3D small target detection method based on lidar, including the following steps: S1: Obtain the raw point cloud information and point cloud support information of the lidar to be detected; S2: Input the original point cloud information and point cloud pillar information into the trained point-enhanced attention perception pillar network. Specifically, the point cloud features and pillar features are fused using a point-pillar cross-attention module to extract fused pillar features. Based on the fused pillar features, the feature weights are dynamically adjusted from the feature dimension and point count dimension using a pillar dual-attention module to extract denoised and optimized pillar features. Based on the denoised and optimized pillar features, multi-resolution set abstraction module fuses feature information from different resolutions to extract multi-scale features. S3: Based on the multi-scale features, perform target classification and position regression to obtain 3D small target detection results.

[0016] Next, combined Figure 2 This embodiment provides a detailed description of a 3D small target detection method based on lidar.

[0017] This embodiment proposes a Point Augmented Attention Perceptual Pillar Network (PAPNet). It is an end-to-end trainable 3D object detector, mainly composed of three components: a Point-Pillar Cross Attention Module (PCM), a Pillar Dual Attention Module (PDAM), and a Multi-Resolution Set Abstraction Module (MSAM). Based on this network, LiDAR point clouds are taken as input, and 3D bounding boxes of objects on the road are output.

[0018] I. Point-Column Cross-Attention Module (PCM) The point-pillar cross-attention module integrates the features of the original point cloud with pillar features of different dimensions through the multi-head cross-attention (MHCA) mechanism.

[0019] like Figure 3 As shown, the point-pillar cross-attention module, as the first stage of the point-enhanced attention perception pillar network, aims to integrate the original point cloud features with pillar features of different dimensions, thereby making more comprehensive use of the original point cloud information.

[0020] The point-pillar cross-attention module consists of two branches, processing the raw point cloud information and the point cloud pillar information respectively. The first branch includes multiple cascaded FCS sub-modules, and the second branch includes multiple cascaded attention mechanism sub-modules. The number of FCS sub-modules and attention mechanism sub-modules is the same.

[0021] Specifically, the raw point cloud information is first preprocessed; for example, downsampling is applied to remove redundant points. Simultaneously, the raw point cloud information is divided into pillars to obtain point cloud pillar information. The raw point cloud information can be point cloud information of a road scene, including the geometric shapes of targets such as vehicles, pedestrians, and road signs. The preprocessed raw point cloud information is then input into the FCS submodule. The FCS submodule consists of a combination of cascaded fully connected layers and shared convolutional layers. It should be understood that the process of obtaining the point cloud pillar information is feasible for those skilled in the art.

[0022] Among them, the point features of the point cloud are enhanced by a fully connected layer (FC) to capture its potential semantic information, thereby obtaining enhanced point cloud features.

[0023] Then, a shared convolutional layer is used to compress the enhanced point cloud features into a more concentrated representation, generating a one-dimensional point cloud feature vector. (i=1,2,3). This contains key semantic information for subsequent processing.

[0024] The shared convolutional layer achieves feature vector dimensionality reduction through three 1D convolutional layers, with both input and output channels fixed at 1.

[0025] In point cloud feature extraction, the sliding method of the convolutional kernel directly affects the dimensionality and detail preservation ability of the feature map. The stride of a 1D convolutional layer defines the distance the convolutional kernel slides with each iteration: a larger stride reduces the dimensionality of the output sequence more, but may result in the loss of local detail information; padding adds zero values ​​to the beginning and end of the input sequence to control the length of the output sequence and prevent excessively rapid dimensionality reduction due to the kernel size and stride. By properly setting the stride and padding parameters, it is possible to preserve the original feature information to the maximum extent while ensuring computational efficiency.

[0026] Specifically, regarding the adjustment of the step size s:

[0027] Let the convolution kernel dimension be . The input vector enhances the feature dimension of the point cloud. The output vector dimension is To ensure that the convolution kernels are uniformly distributed across the feature vectors, the stride is set to... The larger of the rounded-up value and 1 can be used to flexibly adapt to different feature dimensions. When the input feature dimension is large and fast dimensionality reduction is required, a larger step size can improve computational efficiency. When the input feature dimension is small, the step size is at least 1, which can prevent feature information from being lost quickly due to excessive dimensionality reduction. This ensures both the dimensionality reduction rate and reasonable computational efficiency at different feature scales, while retaining key features.

[0028] Then, adjust the padding p to ensure that the dimensionality-reduced output vector maintains the desired dimension, preserving the original feature information to the greatest extent possible.

[0029] By calculating an appropriate padding value p, the feature map size can be kept consistent, which is beneficial for feature fusion operations between subsequent layers. At the same time, it avoids the premature loss of boundary information in the point cloud data, enhances the model's ability to capture detailed features of targets such as vehicles and pedestrians, and lays a solid foundation for subsequent point cloud analysis tasks.

[0030] By calculating a reasonable step size (s) and padding (p), not only is the dimension of the feature vector reduced, but the number of model parameters is also significantly reduced, thereby improving computational efficiency.

[0031] Define the point cloud feature vector obtained after passing through the FCS block. The dimension satisfies relation 4 =2 = This achieves progressively increasing representation capacity within each block. Subsequently, Used as keys in the Multi-Head Cross-Attention (MHCA) mechanism Sum .

[0032] Furthermore, the point cloud pillar information is input into the attention mechanism submodule. The attention mechanism submodule includes a multi-head self-attention (MHSA) submodule and a multi-head cross-attention (MHCA) submodule.

[0033] Let the point cloud pillar information be The dimensions are (P, N, C), where P is the number of pillars, N is the maximum number of points per pillar, and C is the feature dimension. Point cloud pillar information. The attention mechanism submodule of the input point-pillar cross attention module first performs MHSA encoding on the input point, and then uses the encoding result as the query for MHCA. ,and The MHCA operation is performed on the keys and values, and finally the features output by the MHCA are input into a feedforward network (FFN) consisting of two fully connected layers to obtain the processed fused pillar features.

[0034] Specifically, let X be the input point cloud support information. Represents the softmax function. , , These are mapping functions for query, key, and value, respectively.

[0035] In this process, Multi-Head Self-Attention (MHSA) first generates a query, key, and value from the input feature X using a mapping function. The query is then multiplied by the transpose of the key and processed by a softmax function to obtain attention weights. These weights are then multiplied by the values ​​to finally output the encoded first pillar feature, which is used to encode the pillar feature itself.

[0036] The multi-head cross-attention (MHCA) input includes the first pillar features obtained based on the multi-head self-attention submodule and the point cloud feature vector obtained based on the FCS submodule. .

[0037] Generate queries from the first pillar features, from Generate keys and values ​​in the middle, and the subsequent calculations are consistent with MHSA. Used to integrate the pillar features with the original point cloud features to obtain the second pillar features.

[0038]

[0039] The two branches have multiple levels of interconnection and one-to-one module correspondence. After multiple rounds of feature extraction, the second pillar features obtained from the last-level multi-head cross-attention submodule are input into a feedforward network (FFN) composed of two fully connected layers to obtain the processed fused pillar features. .

[0040] In this embodiment, compared to previous methods that only extract pillar features, the point-pillar cross-attention module enhances feature extraction by mining additional information from the original point cloud. Multi-head self-attention (MHSA) encodes the pillar features themselves, focusing on internal correlations and mining their details and patterns; multi-head cross-attention (MHCA) integrates pillar features with original point cloud features, establishing cross-modal connections and supplementing point cloud information beyond the pillar features. The combination of these two approaches not only makes the pillar features themselves more accurate but also incorporates additional information from the original point cloud. Through multi-level concatenation, module correspondence, and multi-round collaborative feature extraction, the final fused features are more comprehensive, enhancing the detection capability for small targets, etc.

[0041] By employing a two-branch, multi-level concatenated feature extraction structure, progressive feature refinement and cross-scale fusion are achieved, effectively avoiding the loss of local contextual information for small targets. The modules in the two branches correspond one-to-one, ensuring not only the symmetry and consistency of feature processing but also promoting information complementarity between different feature spaces through an interaction mechanism. This effectively captures local details and global structural features in point clouds, enhancing the model's ability to understand the complex geometry of target objects or environments. Simultaneously, through adaptive step size and padding strategies, original feature information is preserved to the maximum extent while maintaining computational efficiency, providing richer and more robust feature representations for subsequent tasks, thereby improving overall model performance.

[0042] II. Pillar-type Dual Attention Module (PDAM) The pillar-based dual attention module is used to dynamically adjust the weights of pillar features and suppress noisy features.

[0043] PDAM consists of two branches, both based on fusion pillar features generated by PCM. (P,N,C) are used as input.

[0044] Specifically, the feature dimension attention branch performs max pooling and average pooling along the axis corresponding to feature dimension C, respectively generating feature fusion descriptors. and , representing the average pooling feature and the max pooling feature, respectively. Each feature is then passed to a multilayer perceptron (MLP), which consists of two fully connected layers with weights of and . , The outputs of the two fully connected layers are summed element-wise, and then passed through a sigmoid function. Processing yields the first attention score vector. :

[0045] The other point-count-dimensional attention branch also employs the same operation, focusing on the point-count N-related part, thus obtaining two additional spatial context descriptors. and Second attention score vector Represented as:

[0046] Two attention score vectors and The complete attention matrix is ​​obtained by element-wise multiplication and merging. Then, this matrix is ​​combined with the fusion pillar features. Multiplying (P, N, C) yields the attention-weighted denoising optimization pillar features. (P,N,C).

[0047] Finally, these features After encoding, the images are scattered back to the original pillar positions, thus creating a pseudo-image of size (C,H,W).

[0048] In this embodiment, the pillar-based dual attention module can mine feature associations from different dimensions through a dual-branch structure, achieving accurate feature extraction and fusion. This allows the model to better capture local details and global structure of the point cloud, providing rich and robust features for subsequent tasks. Simultaneously, during attention calculation, effective features can be assigned higher weights, weakening the impact of noisy features, suppressing noise interference, and improving feature quality and model performance.

[0049] III. Multi-resolution collection abstraction module MSAM To generate feature representations incorporating multi-resolution information, a Multi-Resolution Ensemble Abstraction Module (MSAM) is proposed. This lightweight architecture reduces the depth of the backbone network by decreasing the number of blocks, resulting in a concise and efficient design with only three convolutional layers. This significantly accelerates inference speed, reduces computational and memory overhead, and makes the pipeline suitable for real-time applications. Furthermore, compared to traditional convolutional networks that progressively downsample and then upsample features, MSAM can preserve high-resolution and low-resolution features in parallel.

[0050] MSAM can be represented as a series of blocks (S, L, F), where each block runs with a specific stride S, contains L 3x3 two-dimensional convolutional layers, each with F output channels, followed by batch normalization and ReLU activation. The first convolutional layer in each block uses... The step size is determined to ensure that the block receives a step size of [number missing]. After inputting the value, the operation is performed in steps S.

[0051] like Figure 2 As shown, the noise reduction and optimization pillar features are input into the multi-resolution ensemble abstraction module, and the downsampling operation generates high, medium, and low-resolution features in a top-down path. Then, based on the features of the first layer, convolution operations are applied to blocks of each resolution, thereby aggregating information from different scales. Next, MSAM introduces three convolutions and combines downsampling and upsampling operations at the three resolution levels from top to bottom. This process fuses high, medium, and low-resolution information, enabling the model to effectively capture multi-scale features.

[0052]

[0053] in, This represents the feature map of the i-th layer. Indicates the previous layer The feature maps are convolved, and the convolution kernel is... .

[0054] All multi-resolution features are concatenated to form a combined representation, and then deconvolution is performed to adjust the output resolution.

[0055] This embodiment utilizes the Multi-Resolution Aggregate Abstraction Module (MSAM) to simplify the network by reducing the number of blocks with a lightweight architecture. It employs convolutions and upsampling operations with specific strides to preserve high- and low-resolution features in parallel. Furthermore, it fuses information from different scales through cross-sampling and upsampling, thereby accelerating inference speed, reducing computational and memory overhead, and adapting to real-time applications. Simultaneously, it promotes information exchange between multiple resolutions, generating features rich in multi-scale information. This enables the model to have good feature perception capabilities for targets of different sizes, improving performance in tasks such as detection.

[0056] Finally, these features are processed by a simple single-shot detector (SSD), which includes classification and regression branches. The classification branch is responsible for determining the object's category, such as distinguishing between vehicles, pedestrians, and road signs in a driving scenario. The regression branch is used to accurately predict the object's position, such as determining the specific coordinates of a vehicle on the road and the relative position of a pedestrian to a vehicle. Taking a driving scenario as an example, after the LiDAR collects point clouds containing vehicles and pedestrians, and after processing by the preceding modules to obtain multi-scale features, the SSD can quickly classify these targets and regress their accurate positions in the road scene, providing crucial information for autonomous driving decisions.

[0057] To address the shortcomings of existing LiDAR 3D small target detection technologies, such as the tendency of single-point-based methods to lose global information and the difficulty of preserving local details in pillar-based methods, as well as the susceptibility of small target features to noise interference and insufficient multi-scale feature fusion leading to low detection accuracy, this embodiment designs a point-pillar cross-attention module to fuse local features of point clouds with global features of pillars, overcoming the deficiencies of traditional methods in single-dimensional feature extraction. A pillar dual-attention module dynamically adjusts weights from the dimensions of features and point count to effectively suppress noise and solve the problem of small target features being interfered with. An innovative multi-resolution set abstraction module efficiently fuses features of different resolutions, overcoming the limitation of weak multi-scale information integration capabilities in traditional architectures. The target detection network constructed by the overall solution achieves synergistic improvement in local and global information, noise reduction optimization, and multi-scale features, significantly improving the accuracy and reliability of 3D small target detection and providing a better solution for related technology applications.

[0058] Example 2 This embodiment provides a 3D small target detection system based on lidar, including: The data acquisition unit is configured to acquire the raw point cloud information and point cloud support information of the lidar to be detected; The feature extraction unit is configured to input the original point cloud information and point cloud pillar information into a trained point-enhanced attention-based perceptual pillar network. Specifically, it extracts fused pillar features by fusing point cloud features and pillar features through a point-pillar cross-attention module; based on the fused pillar features, it extracts denoised and optimized pillar features by dynamically adjusting feature weights from the feature dimension and point count dimension through a pillar dual-attention module; and based on the denoised and optimized pillar features, it extracts multi-scale features by fusing feature information from different resolutions through a multi-resolution set abstraction module. The target detection unit is configured to perform target classification and position regression based on the multi-scale features to obtain 3D small target detection results.

[0059] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a 3D small target detection method based on lidar as described in Embodiment 1 above.

[0060] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the 3D small target detection method based on lidar as described in Embodiment 1 above.

[0061] The steps or modules involved in Embodiments 2 to 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A laser radar-based 3D small target detection method, characterized by, The method comprises the following steps: acquiring laser radar raw point cloud information and point cloud pillar information to be detected; inputting the raw point cloud information and the point cloud pillar information into a trained point-enhanced attention perception pillar network, wherein point cloud features and pillar features are fused through a point-pillar cross-attention module to extract fused pillar features; based on the fused pillar features, feature weights are dynamically adjusted from feature dimensions and point number dimensions through a pillar double-attention module to extract denoising optimization pillar features; based on the denoising optimization pillar features, different resolution feature information is fused through a multi-resolution set abstraction module to extract multi-scale features; based on the multi-scale features, target classification and position regression are performed to obtain 3D small target detection results.

2. The method of claim 1, wherein the method is based on a laser radar. The raw point cloud information is divided into pillars to construct point cloud pillar information.

3. The method of claim 1, wherein the method is based on a laser radar. The point-pillar cross-attention module fuses point cloud features and pillar features to extract fused pillar features, and specifically comprises: The point-pillar cross-attention module comprises a first branch and a second branch; the first branch comprises a plurality of FCS sub-modules connected in series, and each FCS sub-module comprises a fully connected layer and a shared convolutional layer connected in series; the second branch comprises a combination of a plurality of multi-head self-attention sub-modules and multi-head cross-attention sub-modules connected in series; The raw point cloud information is input into the first branch, and the point cloud feature vector is obtained through the plurality of FCS sub-modules connected in series; The pillar features are input into the second branch, and the first pillar features are obtained by encoding through the multi-head self-attention sub-modules and serving as queries; the point cloud feature vector is taken as keys and values, and the multi-head cross-attention calculation is performed through the multi-head cross-attention sub-modules to obtain the second pillar features; After a plurality of rounds of feature extraction, the second pillar features obtained by the last multi-head cross-attention sub-module are input into a feedforward network to obtain the fused pillar features.

4. The method of claim 3, wherein the method is based on a laser radar. The FCS sub-module, the multi-head self-attention sub-module, and the combination of the multi-head cross-attention sub-module correspond to each other.

5. The method of claim 3, wherein the method is based on a laser radar. The shared convolutional layer comprises three 1D convolutional layers, the dimensions of the point cloud feature vector are adjusted by adjusting the step and padding in the 1D convolutional layers, and specifically: ； ； wherein s and p are 1D convolution layer step and padding respectively, is the dimension of the convolution kernel, is the dimension of the enhanced point cloud feature obtained through the fully connected layer, is the dimension of the output point cloud feature vector.

6. The method of claim 1, wherein the method is a laser radar based 3D small target detection method. Based on the fused pillar features, the feature weights are dynamically adjusted from the feature dimensions and the point number dimensions through the pillar double-attention module to extract the denoising optimization pillar features, and specifically comprises: The pillar double-attention module comprises a feature dimension attention branch and a point number dimension attention branch; In the feature dimension attention branch, the fused pillar features are respectively subjected to global average pooling and global maximum pooling, and the two feature vectors are respectively input into a multi-layer perceptron and then combined to generate a first attention score vector; In the point number dimension attention branch, the same pooling and multi-layer perceptron processing are performed on the fused pillar features along different dimensions to generate a second attention score vector; The two attention score vectors are combined to form a complete attention matrix, and the attention matrix is multiplied with the fused pillar features to obtain the denoising optimization pillar features.

7. The method of claim 1, wherein the method is a laser radar based 3D small target detection method. The multi-resolution set abstraction module comprises three layers of processing; In the first layer of processing, the denoising optimization pillar features are input into a first convolutional layer to generate high, medium, and low resolution features through step-by-step downsampling operations; The second layer processing is a convolution operation on the high, medium and low resolution features respectively; The third layer processing is a combination of down-sampling and up-sampling operations of the high, medium and low resolution features from top to bottom to obtain multi-scale fusion features.

8. A lidar-based 3D small target detection system, characterized by, The method comprises the following steps: The data acquisition unit is configured to acquire laser radar raw point cloud information and point cloud pillar information to be detected; The feature extraction unit is configured to input the raw point cloud information and point cloud pillar information into a trained point-enhanced attention perception pillar network, wherein point cloud features and pillar features are fused through a point-pillar cross attention module to extract fusion pillar features; based on the fusion pillar features, a pillar double attention module is used to dynamically adjust feature weights from feature dimensions and point number dimensions to extract noise reduction optimization pillar features; based on the noise reduction optimization pillar features, a multi-resolution set abstraction module is used to fuse different resolution feature information to extract multi-scale features; The target detection unit is configured to perform target classification and position regression based on the multi-scale features to obtain a 3D small target detection result.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the 3D small target detection method based on a laser radar in any one of claims 1-7.

10. 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 steps of the 3D small target detection method based on a laser radar in any one of claims 1-7.

Images