Target Detection Method Based on Frequency Domain Enhancement and Attention Recalibration
By employing a target detection method that combines frequency domain enhancement and attention recalibration, the problems of information loss and insufficient feature fusion in sparse point cloud data are solved, enabling efficient and stable target detection in complex traffic scenarios.
Patent Information
- Application Number
- CN202611225667.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-08-13
AI Technical Summary
Existing 3D point cloud target detection methods tend to lose height information and local geometric details when processing sparse point cloud data, and the fusion of multi-scale features is insufficient, resulting in inadequate detection performance in complex traffic scenarios.
A target detection method using frequency domain enhancement and attention recalibration is adopted. The spatial self-calibration module of frequency domain enhancement and the spatial self-calibration module of residual column are used to improve the perception ability of local geometric features, and the high-quality fusion of multi-scale features is achieved through a dynamic attention pixel reorganization feature pyramid network.
It significantly improves the feature representation capability of sparse point clouds, enhances the detection stability of distant and occluded targets, reduces background noise interference, and maintains high computational performance.
Smart Images

Figure CN122780926B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of 3D point cloud target detection and autonomous driving technology, and particularly to a target detection method based on frequency domain enhancement and attention recalibration. Background Technology
[0002] With the rapid development of autonomous driving technology, 3D object detection, as a core task of perception systems, plays a crucial role in enabling vehicles to understand complex real-world scenes and achieve environmental semantic modeling. Compared to traditional 2D images, 3D point cloud data contains rich spatial geometric information. LiDAR, as the primary sensor for acquiring point clouds, possesses advantages in detection accuracy and reliability that traditional vision solutions lack, thanks to its precise ranging capabilities and resistance to interference from changes in lighting.
[0003] However, 3D point cloud data is characterized by its discrete, sparse, and non-uniform distribution, which poses a significant challenge to achieving high-precision object detection. To handle disordered point clouds, researchers have proposed voxelization and cylinderization methods. While voxelization enhances the structure of point clouds, it suffers from high computational overhead. Cylinderization, represented by PointPillars (LANG AH, VORA S, CAESAR H, et al.PointPillars: fast encoders for object detection from point clouds[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach: IEEE, 2019: 12697-12705.), a 3D object detection method based on fast point cloud encoding, has a clear advantage in detection speed. However, in the process of compressing 3D spatial features into a 2D pseudo-image, it is easy to weaken the information in the height direction and local geometric details. This lack of fine-grained feature representation capability directly leads to poor model performance in scenarios involving distant targets, small targets, and extremely sparse point clouds.Although in recent years, several 3D object detection methods have emerged, such as CenterPoint (YIN T, ZHOU X, KRAHENBUHL P. Center-based 3D object detection and tracking[C] / / Proceedings of the IEEE / CVFConference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2021: 11784-11793.), which uses the target center point as the key representation for detection, and PillarNet (SHI GS, LI RF, MAC. PillarNet: real-time and high-performance pillar-based 3D object detection[C] / / Proceedings of the 17th European Conference on Computer Vision (ECCV). Tel Aviv: Springer, A series of improved methods, such as 2022, have improved performance from the perspectives of anchorless detection and deep feature encoding. However, in complex traffic scenarios, bottleneck problems such as easy loss of local geometric details, insufficient fusion of multi-scale features, and easy interference of background on target feature response still exist.
[0004] To address the aforementioned issues, there is an urgent need for an efficient algorithm that can effectively enhance local structure perception, suppress background noise, and achieve refined feature fusion, thereby improving the practicality and reliability of LiDAR perception. Summary of the Invention
[0005] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a target detection method based on frequency domain enhancement and attention recalibration.
[0006] Technical solution: The target detection method based on frequency domain enhancement and attention recalibration of the present invention includes the following steps:
[0007] S1, acquire the original LiDAR point cloud data, perform columnar division and feature encoding on the point cloud data, and convert it into two-dimensional pseudo-image features;
[0008] S2, input the two-dimensional pseudo-image features into the backbone network, extract multi-scale features through the frequency domain enhanced spatial self-correction module, in the spatial self-correction module, the features are dimensionally calibrated through the residual column spatial self-correction module, and the corrected multi-scale features are output.
[0009] S3. The corrected multi-scale features are fed into the dynamic attention pixel recombination feature pyramid network. The spatial saliency score is generated by the dynamic recalibration mechanism and converted into recalibration weights. The features are spatially calibrated to obtain the calibrated multi-level features. The pixel recombination operator is used to map the feature maps of each level to a unified spatial resolution. Then, they are spliced and fused along the channel dimension to output the final fused features of the neck network.
[0010] S4 inputs the final fused features into the detection head, predicts the target category, 3D center coordinates, 3D size and orientation angle in parallel, and outputs the 3D target detection results.
[0011] Preferably, step S2 includes:
[0012] Using the two-dimensional pseudo-image features X as the input features of the backbone network, with the current feature transformation step size... Branch selection is based on the criteria for judgment, when When it is determined that the process is in a cross-stage feature transformation process, the first branch is used for frequency domain detail enhancement processing; when When the process is in the feature extraction stage, the second branch is used to perform spatial self-correction enhancement to obtain the corrected multi-scale features.
[0013] Preferably, the step of using the first branch for frequency domain detail enhancement includes:
[0014] The two-dimensional pseudo-image feature X is input into the first branch. In the first branch, two parallel branches are entered at the same time: the stride convolution branch and the HDWT wavelet decomposition branch.
[0015] In the stride convolution branch, the two-dimensional pseudo-image features X undergo a convolution operation with a stride of 2 to obtain data-driven downsampling features. , represented as: ,in, This indicates a batch normalization operation. This indicates a step size of 2. convolution;
[0016] In the HDWT wavelet decomposition branch, the two-dimensional pseudo-image feature X first enters the decomposition process... An extended layer consisting of convolution, batch normalization, and ReLU activation function yields extended features. Subsequently, extended features The input features are fed into the core HDWT submodule, where four sets of Haar wavelet filters with fixed parameters are used to perform two-dimensional spatial frequency decomposition on the input features, resulting in low-frequency subbands. With high-frequency detail subband ;
[0017] Given the Haar wavelet basis functions, the four corresponding Haar wavelet filters are defined as follows: Based on the above filters, the HDWT submodule performs feature analysis. The frequency domain decomposition operation process is represented as follows: ,in, This represents a convolution operation with a stride of 2;
[0018] Subsequently, the frequency components are concatenated along the channel dimension to obtain preliminary downsampling features, represented as: ,in, This indicates a step size of 2. convolution, This represents the LeakyReLU activation function. Indicates feature splicing;
[0019] Initial downsampling features The frequency domain enhancement features are formed by processing through the projection layer and importing them into the residual column space self-correction module for global refinement. , represented as: , ,in, express convolution, Represents the linear rectification activation function. This indicates that the residual column space self-correction module is used for processing;
[0020] Finally, the downsampling features Frequency domain enhancement features Element-by-element addition and fusion are performed to obtain the final feature representation of the first branch. , represented as: .
[0021] Preferably, the step of using the second branch for spatial self-correction enhancement processing includes:
[0022] In the second branch, the two-dimensional pseudo-image feature X first passes through... A channel adjustment layer consisting of convolution, batch normalization, and ReLU activation functions performs adaptive recalibration of feature mapping and channel dimensions, as follows: ,in, express Convolution operation;
[0023] Introduction Convolutional layers model local spatial context information to extract geometric structural features and local semantic information within the same-scale neighborhood. These features are then subjected to batch normalization and ReLU nonlinear transformation to obtain enhanced local features. , represented as: ,in, This represents a local spatial modeling convolution operation;
[0024] Local enhancement features The input is fed into the residual column space self-correction module for global context modeling and feature refinement to obtain the features. , represented as: ;
[0025] A residual connection mechanism is introduced to connect the input feature X with the features after convolution. Element-by-element addition and fusion are performed, and finally, the output features of the second branch are... Represented as: .
[0026] Preferably, the specific implementation process of the residual column space self-correction module includes:
[0027] Receive feature maps extracted from the previous branch It is processed in parallel by spatial perception branch, column perception branch and self-correction branch;
[0028] In the spatial perception branch, feature maps First, the input features are processed through a global average pooling layer and a global max pooling layer to obtain average pooling features and max pooling features that reflect the spatial response intensity. Then, the two pooling features are concatenated along the channel dimension, and spatial dependencies are modeled using a convolutional layer to generate a spatial attention weight map. Finally, the weights are normalized using a sigmoid activation function, and the input features are recalibrated element-wise to obtain the spatially perceptual features. The expression is: In the formula, This represents the Sigmoid activation function. For Hadama accumulation, Cascading operations representing the channel dimension. This represents the global average pooling operator. This represents the global max-pooling operator;
[0029] In the column-oriented sensing branch, an adaptive pooling operator is first used along the feature map. Column-oriented average pooling and column-oriented max pooling operations are performed on the height dimension to extract global column-oriented context information. The results of the two pooling operations are then fused to generate a column-oriented descriptor. Subsequently, the column-oriented descriptor is input into a non-linear dimension compression and dilation encoder, which encodes the column-oriented features through two-level channel mapping, batch normalization, and a non-linear activation function to generate column-oriented attention weights with channel discrimination capabilities. Finally, the generated column-oriented attention weights are used to perform weighted calibration on the original features to obtain the enhanced column-oriented perceptual features. The expression is: In the formula, and These represent the output dimensions as follows: Column-oriented adaptive average pooling and max pooling operators; and This involves two channel mapping operations, one for feature compression and the other for feature activation. and These represent the LeakyReLU activation function and the batch normalized layer, respectively.
[0030] In the self-calibration branch, the feature map First, downsampling is performed using average pooling to expand the feature receptive field and aggregate global contextual information; then, the downsampled features are sequentially input into... Feature encoding is performed using convolutional layers, batch normalization layers, and the LeakyReLU activation function. Next, an upsampling operator is used to restore the encoded features to their original resolution, and then element-wise addition is performed with the input features to fuse local and global contextual information. Finally, a self-calibrating weight matrix is generated using the Sigmoid activation function. The expression is: Where AvgPool represents the average pooling operation. Represents convolution operation; Upsample represents upsampling operation.
[0031] Finally, the residual column space self-correction module concatenates the spatial sensing features and column-direction sensing features through channels, and performs adaptive gating adjustment using self-correction weights. The residual features are then added to the original input through an identity mapping to obtain the final output features. , represented as: .
[0032] Preferably, step S3 includes:
[0033] Input features through After channel expansion via convolutional mapping, the system enters the DAR branch of the dynamic recalibration mechanism, which generates spatial saliency scores through a nonlinear bottleneck layer consisting of two convolutional layers, a ReLU activation function, and a Sigmoid activation function. The expression is: Spatial saliency score This is transformed into dynamic recalibration weights through a dynamic recalibration mechanism. It is used to perform nonlinear modulation on the original branch characteristics, and its expression is: In the formula, For the intensity parameter of weighted regulation;
[0034] The features after dynamic recalibration and weight modulation are multiplied element-wise with the original branch features to obtain a preliminary enhanced feature map.
[0035] The enhanced feature map is spatially reorganized using the pixel reshuffle operator PixelShuffle, mapping the feature map to a preset uniform spatial resolution to obtain enhanced features at each level. ;
[0036] Enhance features at each level By splicing along the channel dimension, the final fused features of the neck network are obtained. , represented as: ,in, This indicates the number of input feature levels.
[0037] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:
[0038] 1. This invention significantly enhances the backbone network's ability to perceive local geometric features of discrete point clouds by designing a frequency-domain enhanced spatial self-correction module (FESSCM). Addressing the issue that traditional cylindrical methods easily lose height information and edge details during 3D spatial compression, this invention introduces Haar discrete wavelet transform for frequency domain decomposition, enabling adaptive capture of fine geometric features such as target edge contours. Simultaneously, through a stage attribute judgment mechanism, the optimal processing path can be automatically matched based on the feature stride, emphasizing detail enhancement during cross-stage transformations and spatial self-correction during intra-stage refinement. This effectively reduces unnecessary computational overhead while improving feature discriminative power.
[0039] 2. This invention effectively alleviates the feature weakening and distortion problems caused by sparse point clouds through the integrated Residual Column Space Self-Correction (RCSSC) module. The Column-Aware Branch (CAB) in the RCSSC module models global column-direction dependencies through adaptive pooling and generates channel-discriminative attention weights to calibrate the original features. This column-direction context aggregation mechanism enables the model to adaptively adjust the feature responses of different spatial regions, significantly mitigating the problem of geometric information loss due to sparse point clouds for distant targets and enhancing the model's fine-grained expressive ability in complex traffic scenarios.
[0040] 3. The Dynamic Attention Pixel Recombination Feature Pyramid Network (DAPSFPN) constructed in this invention achieves high-quality fusion and target enhancement of multi-scale features. This network utilizes a dynamic recalibration mechanism to nonlinearly modulate features based on spatial saliency scores, adaptively enhancing feature representations of high-response target regions and accurately suppressing background noise interference, significantly improving the detection stability for occluded targets and distant small targets. Furthermore, by replacing traditional upsampling with the pixel recombination operator PixelShuffle, this invention achieves multi-scale spatial scale alignment while preserving spatial detail information to the greatest extent, providing more refined feature data for final category prediction and 3D attribute regression. Attached Figure Description
[0041] Figure 1 This is a flowchart of the present invention;
[0042] Figure 2 This is a block diagram for object detection;
[0043] Figure 3 This is a schematic diagram of the Frequency Domain Enhanced Spatial Self-Correcting Module (FESSCM).
[0044] Figure 4 This is a schematic diagram of the Residual Column Space Self-Correcting Module (RCSSC).
[0045] Figure 5 This is a schematic diagram of the DAPSM module;
[0046] Figure 6 It is the overall structure of DAPSFPN;
[0047] Figure 7 This is a visualization comparison of the object detection performance of PointPillars and FARNet in complex traffic scenarios;
[0048] Figure 8 This is a visualization comparison of the object detection performance of PointPillars and FARNet in a car-occluded scene. Detailed Implementation
[0049] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.
[0050] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.
[0051] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.
[0052] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0053] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0054] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.
[0055] Combination Figure 1 and Figure 2 As shown in the figure, the target detection method based on frequency domain enhancement and attention recalibration described in this embodiment includes the following steps:
[0056] S1: Acquire raw LiDAR point cloud data, perform columnar division and feature encoding on the point cloud data, and convert it into two-dimensional pseudo-image features.
[0057] S2 inputs the two-dimensional pseudo-image features into the backbone network, extracts multi-scale features through the frequency domain enhanced spatial self-calibration module FESSCM, and performs dimensional calibration on the features through the residual column spatial self-calibration module in the spatial self-calibration module, outputting the calibrated multi-scale features.
[0058] Preferably, step S2 includes:
[0059] Using the two-dimensional pseudo-image features X as the input features of the backbone network, with the current feature transformation step size... Branch selection is based on the criteria for judgment, when When it is determined that the process is in a cross-stage feature transformation process, the first branch is used for frequency domain detail enhancement processing; when When the process is in the feature extraction stage, the second branch is used to perform spatial self-correction enhancement to obtain the corrected multi-scale features.
[0060] Combination Figure 3 As shown, in one example, the steps for frequency domain detail enhancement using the first branch include:
[0061] The two-dimensional pseudo-image feature X is input into the first branch. In the first branch, two parallel branches are entered at the same time: the stride convolution branch and the HDWT wavelet decomposition branch.
[0062] In the stride convolution branch, the two-dimensional pseudo-image features X undergo a convolution operation with a stride of 2 to obtain data-driven downsampling features. , represented as: ,in, This indicates a batch normalization operation. This indicates a step size of 2. convolution;
[0063] In the HDWT wavelet decomposition branch, the two-dimensional pseudo-image feature X first enters the decomposition process... An extended layer consisting of convolution, batch normalization, and ReLU activation function yields extended features. Subsequently, extended features The input features are fed into the core HDWT submodule, where four sets of Haar wavelet filters with fixed parameters are used to perform two-dimensional spatial frequency decomposition on the input features, resulting in low-frequency subbands. With high-frequency detail subband ;
[0064] Given the Haar wavelet basis functions, the four corresponding Haar wavelet filters are defined as follows: Based on the above filters, the HDWT submodule performs feature analysis. The frequency domain decomposition operation process is represented as follows: ,in, This represents a convolution operation with a stride of 2;
[0065] Subsequently, the frequency components are concatenated along the channel dimension to obtain preliminary downsampling features, represented as: ,in, This indicates a step size of 2. convolution, This represents the LeakyReLU activation function. Indicates feature splicing;
[0066] Initial downsampling features The frequency domain enhancement features are formed by processing through the projection layer and importing them into the residual column space self-correction module for global refinement. , represented as: , ,in, Represents a 1×1 convolution. Represents the linear rectification activation function. This indicates that the residual column space self-correction module is used for processing;
[0067] Finally, the downsampling features Frequency domain enhancement features Element-by-element addition and fusion are performed to obtain the final feature representation of the first branch. , represented as: .
[0068] In one example, the steps for spatial self-correction enhancement using the second branch include:
[0069] In the second branch, the two-dimensional pseudo-image feature X first passes through... A channel adjustment layer consisting of convolution, batch normalization, and ReLU activation functions performs adaptive recalibration of feature mapping and channel dimensions, as follows: ,in, express Convolution operation;
[0070] Introduction Convolutional layers model local spatial context information to extract geometric structural features and local semantic information within the same-scale neighborhood. These features are then subjected to batch normalization and ReLU nonlinear transformation to obtain enhanced local features. , represented as: ,in, This represents a local spatial modeling convolution operation;
[0071] Local enhancement features The input is fed into the residual column space self-correction module for global context modeling and feature refinement to obtain the features. , represented as: ;
[0072] A residual connection mechanism is introduced to connect the input feature X with the features after convolution. Element-by-element addition and fusion are performed, and finally, the output features of the second branch are... Represented as: .
[0073] Combination Figure 4 As shown in the example, the specific implementation process of the Residual Column Space Self-Correcting Module (RCSSC) includes:
[0074] Receive feature maps extracted from the previous branch The SCB is processed in parallel by the spatial sensing branch SAB, the column sensing branch CAB, and the self-correction branch.
[0075] In the spatial perception branch, feature maps First, the input features are processed through a global average pooling layer and a global max pooling layer to obtain average pooling features and max pooling features that reflect the spatial response intensity. Then, the two pooling features are concatenated along the channel dimension, and spatial dependencies are modeled using a convolutional layer to generate a spatial attention weight map. Finally, the weights are normalized using a sigmoid activation function, and the input features are recalibrated element-wise to obtain the spatially perceptual features. The expression is: In the formula, This represents the Sigmoid activation function. For Hadama accumulation, Cascading operations representing the channel dimension. This represents the global average pooling operator. This represents the global max-pooling operator;
[0076] In the column-oriented sensing branch, an adaptive pooling operator is first used along the feature map. Column-oriented average pooling and column-oriented max pooling operations are performed on the height dimension to extract global column-oriented context information. The results of the two pooling operations are then fused to generate a column-oriented descriptor. Subsequently, the column-oriented descriptor is input into a non-linear dimension compression and dilation encoder, which encodes the column-oriented features through two-level channel mapping, batch normalization, and a non-linear activation function to generate column-oriented attention weights with channel discrimination capabilities. Finally, the generated column-oriented attention weights are used to perform weighted calibration on the original features to obtain the enhanced column-oriented perceptual features. The expression is: In the formula, and These represent the output dimensions as follows: Column-oriented adaptive average pooling and max pooling operators; and This involves two channel mapping operations, one for feature compression and the other for feature activation. and These represent the LeakyReLU activation function and the batch normalized layer, respectively.
[0077] In the self-calibration branch, the feature map First, downsampling is performed using average pooling to expand the feature receptive field and aggregate global contextual information; then, the downsampled features are sequentially input into... Feature encoding is performed using convolutional layers, batch normalization layers, and the LeakyReLU activation function. Next, an upsampling operator is used to restore the encoded features to their original resolution, and then element-wise addition is performed with the input features to fuse local and global contextual information. Finally, a self-calibrating weight matrix is generated using the Sigmoid activation function. The expression is: ,in, This indicates the average pooling operation. represents convolution operation, BN represents batch normalization operation, and Upsample represents upsampling operation;
[0078] Finally, the residual column space self-correction module concatenates the spatial sensing features and column-direction sensing features through channels, and performs adaptive gating adjustment using self-correction weights. The residual features are then added to the original input through an identity mapping to obtain the final output features. , represented as: .
[0079] S3. The corrected multi-scale features are fed into the Dynamic Attention Pixel Recombination Feature Pyramid Network (DAPSFPN). The spatial saliency score is generated by the dynamic recalibration mechanism and converted into recalibration weights. The features are spatially calibrated to obtain the calibrated multi-level features. The pixel recombination operator is used to map the feature maps of each level to a unified spatial resolution. Then, they are spliced and fused along the channel dimension to output the final fused features of the neck network.
[0080] Combination Figure 5 and Figure 6 As shown, the Dynamic Attention Pixel Reassembly Feature Pyramid Network (DAPSFPN) receives outputs from different stages of the backbone network. Multi-scale feature mappings at each level are processed and fed into corresponding core upsampling units (DAPSMs) for feature enhancement and spatial decoding. Each level of input feature is processed by the corresponding DAPSM upsampling unit to generate enhanced features that have been recalibrated in both the channel and spatial dimensions. .
[0081] In one example, step S3 includes:
[0082] Input features through After channel expansion via convolutional mapping, the system enters the DAR branch of the dynamic recalibration mechanism, which generates spatial saliency scores through a nonlinear bottleneck layer consisting of two convolutional layers, a ReLU activation function, and a Sigmoid activation function. The expression is: ;
[0083] Spatial saliency score This is transformed into dynamic recalibration weights through a dynamic recalibration mechanism. It is used to perform nonlinear modulation on the original branch characteristics, and its expression is: In the formula, For the intensity parameter of weighted regulation;
[0084] The features after dynamic recalibration and weight modulation are multiplied element-wise with the original branch features to obtain a preliminary enhanced feature map.
[0085] The enhanced feature map is spatially reorganized using the pixel reshuffle operator PixelShuffle, mapping the feature map to a preset uniform spatial resolution to obtain enhanced features at each level. ;
[0086] Enhance features at each level By splicing along the channel dimension, the final fused features of the neck network are obtained. , represented as: ,in, This indicates the number of input feature levels.
[0087] S4 inputs the final fused features into the detection head, predicts the target category, 3D center coordinates, 3D size and orientation angle in parallel, and outputs the 3D target detection results.
[0088] The method described in this invention is denoted as FARNet. By employing an improved FARNet based on frequency domain enhancement and attention recalibration, and inputting the original point cloud from the KITTI dataset, the original point cloud acquired by the LiDAR is visualized in 3D space using the open-source 3D data processing library Open3D. The 3D bounding boxes of the target object's true and predicted positions are then labeled. The detection results are as follows: Figure 7 and Figure 8 As shown, Figure 7 A visualization comparison of the 3D and 2D object detection performance of PointPillars and FARNet in complex traffic scenes, such as... Figure 7 As shown in the orange circle on the left of (a), PointPillar exhibits significant missed detections when dealing with small targets at a distance or in sparsely distributed point clouds. This phenomenon reflects the limitation of traditional PointPillars methods, which easily lead to the loss of subtle local spatial features during spatial feature dimensionality reduction, making it difficult for the model to extract effective semantic information from a very small number of point clouds. In contrast, as... Figure 7As shown in the yellow elliptical area in (b), FARNet has a more complete detection response for labeled targets at a distance and in sparse areas of the point cloud, indicating that the frequency domain enhancement and attention recalibration mechanism based on the present invention helps to improve the model's ability to discriminate the local geometric structure of sparse targets.
[0089] Figure 8 This presentation compares the detection results of FARNet and PointPillars in vehicle occlusion scenarios, including... Figure 8 As shown in the orange circle in (a) of the diagram, PointPillars missed detecting partially visible occluded vehicles in the distance, indicating that relying solely on local columnar features may be insufficient to fully reconstruct the spatial structure of the target when the target point cloud information is incomplete. In contrast, Figure 8 In (b) of the results, FARNet provides relatively complete detection results for occluded vehicles. This is because the frequency domain enhancement module in FARNet can preserve target edges and local geometric details to a certain extent, and the residual column space self-correction module helps to enhance the feature response modeling between different spatial regions; at the same time, DAPSPPN improves the multi-scale feature fusion effect through dynamic attention recalibration and pixel recombination mechanisms. The above results show that FARNet has better target modeling capabilities in occluded and sparse point cloud scenarios, and can reduce the impact of complex background interference on detection results.
[0090] Table 1 shows the performance of different methods on the KITTI dataset validation set. For performance comparison, this example selected several mainstream comparison methods: VoxelNet (ZHOU Y, TUZEL O. VoxelNet: End-to-end learning for pointcloud based 3D object detection[C] / / Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City: IEEE, 2018:4490-4499.), SECOND (YAN Y, MAO Y, LI B. SECOND: Sparsely embedded convolutional detection[C] / / Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Salt Lake City: IEEE, 2018: 872-880.), and Voxel R-CNN (DENG J, SHI S, LI P, et al. Voxel R-CNN: Towards highperformance voxel-based 3D object detection[C] / / Proceedings of the AAAIConference on Artificial Intelligence. Virtual: AAAI Press, 2021, 35(2):1201-1209.) is a voxel-based 3D object detection method. This type of method discretizes sparse and irregular point clouds into regular voxel structures and combines them with 3D convolutional networks for feature extraction, achieving a good balance between detection accuracy and computational efficiency. PointPillars is a typical method based on columnar representation. It uses columnar voxels to encode point clouds and transforms the 3D detection problem into a 2D pseudo-image processing process, thus possessing high inference speed and real-time detection capability.3DSSD (YANG ZT, SUN YN, LIU S, et al. 3DSSD: Point-based 3D singlestage object detector[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2020: 11040-11048.) and PointRCNN (SHI SS, WANG XG, LI H S. PointRCNN: 3D object proposal generation and detection from point cloud[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach: IEEE, 2019: 770-779.) are point-level feature learning methods. Both extract target features directly from the original point cloud, which can fully preserve the fine-grained geometric information and spatial structure features of the point cloud. VoxT-GNN (ZHENG QW, WU S, WEI J H. VoxT-GNN: A 3D object detection approach from point cloud based on voxel-level transformer and graph neural network[J]. Information Processing & Management, 2025, 62(4): 104155.) integrates the modeling ideas of graph neural networks, and achieves effective modeling of long-distance dependent information in complex scenes by constructing graph structure relationships between point cloud features. The above comparison methods cover a variety of mainstream technical routes such as voxel-based, column-based, point-level, and graph neural network-based methods, which can verify the effectiveness and advancement of the proposed method from different perspectives. The 3D target detection accuracy under different difficulty levels was statistically analyzed in three categories: vehicles, pedestrians, and cyclists. As shown in the table, the FARNet method proposed in this invention achieves high detection accuracy in multiple categories and difficulty indicators, especially when dealing with occluded and distant targets, it shows stronger robustness. Compared with the benchmark algorithm PointPillars, its average accuracy in the vehicle category is improved by 2.5%.
[0091] Table 1. Comparison of AP40 scores for different 3D object detection methods on the KITTI dataset.
[0092] ;
[0093] Table 2 presents the ablation experiments for different improved modules. As shown in Table 2, the frequency domain enhanced spatial self-calibration module FESSCM and its built-in residual column spatial self-calibration module RCSSC improved the mAP of the improved network by 1.8% compared to the original network, demonstrating the effectiveness of frequency domain detail enhancement and column-oriented context modeling in representing sparse point cloud features. The dynamic attention pixel reconstruction feature pyramid network DAPSPPN further improved the mAP of the improved network by 1.2%, reflecting the advantages of the dynamic recalibration mechanism in suppressing background noise and enhancing target features. The overall network structure achieved significant improvements in all metrics compared to the original network. This ablation experiment demonstrates the effectiveness of each module in the improved FARNet target detection method of this invention, achieving accurate detection of point cloud targets in complex traffic scenarios.
[0094] Table 2. Different module combinations on the KITTI validation set Ablation test results
[0095] ;
[0096] Table 3 analyzes the computational overhead and inference speed. This example tested the model's inference speed and latency under the same hardware environment. In Table 3, Latency represents latency, and Memory represents memory usage. Table 3 shows that the stage attribute judgment mechanism in the FESSCM module effectively avoids unnecessary computational overhead through adaptive path selection. Although the introduction of frequency domain enhancement and attention mechanisms increases the computational load, thanks to the efficient spatial transformation capability of the pixel recombination operator, the improved overall network only shows a minimal increase in inference latency compared to the original network, maintaining a high real-time detection speed. These experimental results demonstrate that this invention significantly improves detection accuracy while maintaining algorithm execution efficiency, meeting the real-time requirements of autonomous driving environmental perception.
[0097] Table 3. Comparison of computational efficiency of different methods on the KITTI validation set.
[0098] ;
[0099] In summary, the target detection method based on frequency domain enhancement and attention recalibration proposed in this invention achieves accurate and stable detection of point cloud targets in complex scenes while maintaining high operating efficiency compared to the original cylindrical detection method.
Claims
1. A target detection method based on frequency domain enhancement and attention recalibration, characterized in that, Includes the following steps: S1, acquire the original LiDAR point cloud data, perform columnar division and feature encoding on the point cloud data, and convert it into two-dimensional pseudo-image features; S2, input the two-dimensional pseudo-image features into the backbone network, extract multi-scale features through the frequency domain enhanced spatial self-correction module, in the spatial self-correction module, the features are dimensionally calibrated through the residual column spatial self-correction module, and the corrected multi-scale features are output. S3, the corrected multi-scale features are fed into the dynamic attention pixel reorganization feature pyramid network, and the spatial saliency score is generated by the dynamic recalibration mechanism and converted into recalibration weights. The features are spatially calibrated to obtain the calibrated multi-level features. The pixel recombination operator is used to map the feature maps of each level to a uniform spatial resolution, and then they are spliced and fused along the channel dimension to output the final fused features of the neck network. S4 inputs the final fused features into the detection head, predicts the target category, 3D center coordinates, 3D size and orientation angle in parallel, and outputs the 3D target detection results; Step S2 includes: Using the two-dimensional pseudo-image features X as the input features of the backbone network, and using the stride s of the current feature transformation as the criterion for branch selection, when... When it is determined that the process is in a cross-stage feature transformation process, the first branch is used for frequency domain detail enhancement processing; when When the process is in the feature extraction stage, the second branch is used to perform spatial self-correction enhancement to obtain corrected multi-scale features. The specific implementation process of the residual column space self-correction module includes: Receive feature maps extracted from the previous branch It is processed in parallel by spatial perception branch, column perception branch and self-correction branch; In the spatial perception branch, feature maps First, the input features are processed through a global average pooling layer and a global max pooling layer to obtain average pooling features and max pooling features that reflect the spatial response intensity. Then, the two pooling features are concatenated along the channel dimension, and spatial dependencies are modeled using a convolutional layer to generate a spatial attention weight map. Finally, the weights are normalized using a sigmoid activation function, and the input features are recalibrated element-wise to obtain the spatially perceptual features. The expression is: , In the formula, This represents the Sigmoid activation function. For Hadama accumulation, express convolution, Cascading operations representing channel dimensions; This represents the global average pooling operator. This represents the global max-pooling operator; In the column-oriented sensing branch, an adaptive pooling operator is first used along the feature map. Column-oriented average pooling and column-oriented max pooling operations are performed on the height dimension to extract global column-oriented context information. The results of the two pooling operations are then fused to generate a column-oriented descriptor. Subsequently, the column-oriented descriptor is input into a non-linear dimension compression and dilation encoder, which encodes the column-oriented features through two-level channel mapping, batch normalization, and a non-linear activation function to generate column-oriented attention weights with channel discrimination capabilities. Finally, the generated column-oriented attention weights are used to perform weighted calibration on the original features to obtain the enhanced column-oriented perceptual features. The expression is: , In the formula, and These represent column-oriented adaptive average pooling and max pooling operators with an output size of (1, W), respectively. and This involves two channel mapping operations, one for feature compression and the other for feature activation. In the self-calibration branch, the feature map First, downsampling is performed using average pooling to expand the receptive field and aggregate global contextual information. Then, the downsampled features are sequentially input into a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function for feature encoding. Next, an upsampling operator is used to restore the encoded features to their original resolution, and they are then element-wise added to the input features to fuse local and global contextual information. Finally, a self-calibrating weight matrix SC is generated using the Sigmoid activation function, with the expression: , Where AvgPool represents the average pooling operation, Conv 3×3 Represents convolution operation; Upsample represents upsampling operation. Finally, the residual column space self-correction module concatenates the spatial sensing features and column-direction sensing features through channels, and performs adaptive gating adjustment using self-correction weights. The residual features are then added to the original input through an identity mapping to obtain the final output features. , is represented as: 。 2. The target detection method based on frequency domain enhancement and attention recalibration according to claim 1, characterized in that, The steps for frequency domain detail enhancement using the first branch include: The two-dimensional pseudo-image feature X is input into the first branch. In the first branch, two parallel branches are entered at the same time: the stride convolution branch and the HDWT wavelet decomposition branch. In the stride convolution branch, the two-dimensional pseudo-image features X undergo a convolution operation with a stride of 2 to obtain data-driven downsampling features. , is represented as: , in, This indicates a batch normalization operation. This indicates a step size of 2. convolution; In the HDWT wavelet decomposition branch, the two-dimensional pseudo-image feature X first enters the decomposition process... An extended layer consisting of convolution, batch normalization, and ReLU activation function yields extended features. Subsequently, extended features The input features are fed into the core HDWT submodule, where four sets of Haar wavelet filters with fixed parameters are used to perform two-dimensional spatial frequency decomposition on the input features, resulting in low-frequency subbands. With 3 high-frequency detail sub-bands ; Given the Haar wavelet basis functions, the four corresponding Haar wavelet filters are defined as follows: , Based on the above filters, the HDWT submodule performs feature analysis. The frequency domain decomposition operation process is represented as follows: , in, This represents a convolution operation with a stride of 2; Subsequently, the frequency components are concatenated along the channel dimension to obtain preliminary downsampling features, represented as: , in, This indicates a step size of 2. convolution, This represents the LeakyReLU activation function; Indicates feature splicing; This represents a 3×3 convolution with a stride of 1; Initial downsampling features The frequency domain enhancement features are formed by processing through the projection layer and importing them into the residual column space self-correction module for global refinement. , is represented as: , , in, express Convolution operation; Represents the linear rectification activation function; This indicates that the residual column space self-correction module is used for processing; Finally, the downsampling features Frequency domain enhancement features Element-by-element addition and fusion are performed to obtain the final feature representation of the first branch. , is represented as: 。 3. The target detection method based on frequency domain enhancement and attention recalibration according to claim 2, characterized in that, The steps for spatial self-correction enhancement using the second branch include: In the second branch, the two-dimensional pseudo-image feature X first passes through... A channel adjustment layer consisting of convolution, batch normalization, and ReLU activation functions performs adaptive recalibration of feature mapping and channel dimensions, as follows: , in, This represents a 1×1 convolution operation; A 3×3 convolutional layer is introduced to model local spatial context information, used to extract geometric structural features and local semantic information within the same-scale neighborhood. Subsequently, batch normalization and ReLU nonlinear transformation are applied to obtain enhanced local features. , is represented as: , in, This represents a local spatial modeling convolution operation; express convolution; Local enhancement features The input is fed into the residual column space self-correction module for global context modeling and feature refinement to obtain the features. , is represented as: ; A residual connection mechanism is introduced to connect the input feature X with the features after convolution. Element-by-element addition and fusion are performed, and finally, the output features of the second branch are... Represented as: 。 4. The target detection method based on frequency domain enhancement and attention recalibration according to claim 3, characterized in that, Step S3 includes: Input features through After channel expansion via convolutional mapping, the system enters the DAR (Dynamic Recalibration) branch, which generates a spatial saliency score S through a nonlinear bottleneck layer consisting of two convolutional layers, a ReLU activation function, and a Sigmoid activation function. The expression for this score is: , The spatial saliency score S is transformed into a dynamic recalibration weight through a dynamic recalibration mechanism. It is used to perform nonlinear modulation on the original branch characteristics, and its expression is: , In the formula, γ is the weighted control intensity parameter; The features after dynamic recalibration and weight modulation are multiplied element-wise with the original branch features to obtain a preliminary enhanced feature map. The enhanced feature map is spatially reorganized using the pixel reshuffle operator PixelShuffle, mapping the feature map to a preset uniform spatial resolution to obtain the enhanced features F at each level. i ; Enhance the features F at each level i By splicing along the channel dimension, the final fusion feature F of the neck network is obtained. out , is represented as: ,in, This indicates the number of input feature levels.
