A three-dimensional target detection method and device fusing coordinate attention and multi-scale features

Abstract

This invention discloses a method and apparatus for 3D target detection that integrates coordinate attention and multi-scale features, relating to the field of 3D target detection technology. The method includes: acquiring raw 3D point cloud data, extracting features, and generating 2D bird's-eye view features; performing orthogonal one-dimensional feature aggregation on the 2D bird's-eye view features to extract direction-aware features and determine spatial attention weights; reweighting the 2D bird's-eye view features to generate enhanced multi-scale features; weighting and aggregating the enhanced multi-scale features to generate a shared feature map; performing attribute prediction processing based on the shared feature map to obtain a center point heatmap and 3D bounding box attributes, and performing geometric co-optimization on the 3D bounding box attributes to output the detection results. This invention decouples global pooling into one-dimensional pooling in the horizontal and vertical directions, retaining the coordinate priors of each spatial location. The bidirectional attention weights generated on this basis can adaptively enhance the feature response at the target's location.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional target detection technology, and in particular to a three-dimensional target detection method and apparatus that integrates coordinate attention and multi-scale features. Background Technology

[0002] In autonomous driving environmental perception, 3D point cloud target detection based on LiDAR is a core component for achieving high-precision spatial positioning. To improve end-to-end inference efficiency, mainstream methods (such as anchor-free detection paradigms like CenterPoint) achieve a balance between computational cost and detection accuracy by combining 3D sparse convolution with center point heatmap prediction, and have now become the mainstream baseline. However, when faced with extremely sparse point cloud distributions and targets of vastly different scales, existing architectures still suffer from the following three major technical shortcomings in practical applications: First, feature dimensionality reduction leads to the loss of precise spatial coordinate priors, resulting in insufficient sparse semantic modeling capabilities. During feature extraction, the global two-dimensional pooling operation completely erases precise spatial coordinate priors. Traditional feature aggregation and attention mechanisms (such as CBAM, graph convolution, etc.) struggle to achieve a balance between long-range contextual dependencies and precise location. When the aggregation receptive field expands, the already sparse geometric structure of small targets (such as distant pedestrians) is easily assimilated by the background, leading to severe feature diffusion.

[0003] Second, traditional multi-scale feature fusion suffers from the problem of high-level semantics overshadowing low-level spatial details. Traditional multi-scale feature fusion (such as FPN) often relies on unidirectional or equally weighted topological structures. This crude addition mechanism easily leads to high-level strong semantic features overshadowing low-level weak spatial details, hindering the accurate perception of small-scale targets and thus causing a high false negative rate.

[0004] Third, independent regression of multidimensional attributes leads to spatial misalignment and gradient conflicts. Existing 3D bounding box prediction widely uses the smoothed L1 loss function, treating the target's length, width, height, and orientation angle as independent scalars for isolated regression, ignoring the strong coupling of physical volume. In 3D scenes, especially for objects with extreme aspect ratios (such as cyclists), even small angular deviations can easily cause significant spatial misalignment, leading to severe gradient inefficiency and conflicts in the early stages of model training, thus limiting the upper limit of detection accuracy. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a three-dimensional target detection method and device that integrates coordinate attention and multi-scale features.

[0006] To solve the above technical problems, the technical solution of the present invention is as follows: A 3D target detection method integrating coordinate attention and multi-scale features includes: Acquire raw 3D point cloud data and extract features from the raw 3D point cloud data to generate 2D bird's-eye view features; One-dimensional feature aggregation in orthogonal directions is performed on the two-dimensional bird's-eye view features to extract direction-aware features. Spatial attention weights are determined based on the direction-aware features. The spatial attention weights are then used to reweight the two-dimensional bird's-eye view features to generate enhanced multi-scale features. By using a bidirectional information flow path across scales, the enhanced multi-scale features at different resolutions are weighted and aggregated using normalized weights obtained through adaptive learning to generate a unified shared feature map. The shared feature map is subjected to decoupled attribute prediction processing to obtain a target center point heatmap and three-dimensional bounding box attributes. The three-dimensional bounding box attributes are then output as the three-dimensional target detection result. The computational parameters required for generating the enhanced multi-scale features, generating the shared feature map, and performing the attribute prediction processing are all determined through collaborative optimization during the training phase using a joint loss function. This joint loss function integrates the three-dimensional volume intersection-union constraint, the aspect ratio difference penalty term, and the orientation angle difference penalty term.

[0007] As a preferred embodiment of the 3D target detection method integrating coordinate attention and multi-scale features described in this invention, wherein: one-dimensional feature aggregation in orthogonal directions is performed on the two-dimensional bird's-eye view features to extract direction-aware features, including: One-dimensional pooling kernels with spatial ranges of (H, 1) and (1, W) are used to aggregate each channel of the input features along the horizontal and vertical directions, respectively, to obtain the horizontal and vertical perceptual feature maps.

[0008] As a preferred embodiment of the 3D target detection method integrating coordinate attention and multi-scale features described in this invention, determining the spatial attention weights based on the orientation-aware features includes: The horizontal and vertical perceptual feature maps are spliced ​​together in the spatial dimension, and the features are fused through a shared convolutional layer and a nonlinear activation function to generate intermediate fused features. The intermediate fusion features are separated along the spatial dimension and then passed through independent convolutional layers and sigmoid activation functions to generate horizontal and vertical attention weights.

[0009] As a preferred embodiment of the 3D target detection method integrating coordinate attention and multi-scale features described in this invention, the step of weighted aggregation of the enhanced multi-scale features at different resolutions through a cross-scale bidirectional information flow path includes: Add jump residual connection edges between input and output nodes at the same level, and remove intermediate nodes with only one input edge; Introducing a minimum value The fast normalization fusion strategy is based on the weight parameters learned by network adaptive learning to perform the weighted aggregation.

[0010] As a preferred embodiment of the 3D target detection method fusing coordinate attention and multi-scale features described in this invention, wherein: in the fast normalization fusion strategy, the calculation method of the output feature O of each node is as follows: ,in, Let i be the i-th input feature of this node. and These are the weighted parameters obtained through adaptive learning. This is the sum of the weights of all input features for the current node.

[0011] As a preferred embodiment of the 3D target detection method fusing coordinate attention and multi-scale features described in this invention, the joint loss function is defined as follows: ,in, The formula for predicting the volume intersection-union ratio (IU / U) between a 3D bounding box and the actual 3D bounding box is as follows: In the formula, To predict the intersection volume between the 3D bounding box and the true 3D bounding box, To predict the volume of a 3D bounding box, The volume of the true 3D bounding box. It is a smoothing constant; Penalty item for aspect ratio difference; For dynamic adaptive weights; This is a penalty term for differences in orientation angle; To balance the hyperparameters.

[0012] As a preferred embodiment of the 3D target detection method that integrates coordinate attention and multi-scale features as described in this invention, wherein: the aspect ratio difference penalty term The calculation formula is: ,in, and These are the width and length of the actual bounding box, respectively. and These represent the width and length of the predicted bounding box, respectively.

[0013] As a preferred embodiment of the 3D target detection method that integrates coordinate attention and multi-scale features as described in this invention, wherein: the orientation angle difference penalty term The calculation formula is: ,in, To predict the orientation angle, This is the actual orientation angle.

[0014] As a preferred embodiment of the 3D target detection method that integrates coordinate attention and multi-scale features as described in this invention, wherein: the dynamic adaptive weights The calculation formula is: .

[0015] The present invention also provides a three-dimensional target detection device that integrates coordinate attention and multi-scale features, comprising: The point cloud acquisition module is used to acquire raw 3D point cloud data and extract features from the raw 3D point cloud data to generate 2D bird's-eye view features. The coordinate attention module is used to perform one-dimensional feature aggregation in orthogonal directions on the two-dimensional bird's-eye view features, extract direction-aware features, determine spatial attention weights based on the direction-aware features, and reweight the two-dimensional bird's-eye view features using the spatial attention weights to generate enhanced multi-scale features. The weighted bidirectional feature pyramid fusion module is used to weight and aggregate the enhanced multi-scale features at different resolutions through a cross-scale bidirectional information flow path and by using normalized weights learned through adaptive learning to generate a unified shared feature map. The detection output module is used to perform decoupled attribute prediction processing on the shared feature map to obtain the target center point heatmap and three-dimensional bounding box attributes, and output the three-dimensional bounding box attributes as the three-dimensional target detection result. Among them, the computational parameters required for generating enhanced multi-scale features, generating shared feature maps, and performing attribute prediction processing are all determined through collaborative optimization during the training phase using a joint loss function. The joint loss function integrates the 3D volume intersection-union constraint, the aspect ratio difference penalty term, and the orientation angle difference penalty term.

[0016] The beneficial effects of this invention are: (1) This invention decouples global pooling into one-dimensional pooling in the horizontal and vertical directions, explicitly preserving the coordinate priors of each spatial location. The bidirectional attention weights generated on this basis can adaptively enhance the feature response at the target's location and suppress irrelevant background. Ablation experiments show that, with only this mechanism introduced, the detection accuracy of the Pedestrian category's Moderate is improved from 67.93% to 68.21%, and the Cyclist category from 86.65% to 87.47%. When used in conjunction with other modules, the recall rate for small targets is significantly improved, effectively overcoming the problem of easily missed detections of small targets in sparse point clouds.

[0017] (2) This invention employs learnable normalized weights to dynamically weight and aggregate features from different scales, and constructs a bidirectional information flow path from top to bottom and bottom to top, enabling the network to adaptively determine the contribution of features at each scale based on the data. Ablation experiments show that introducing this module alone can achieve a balanced increase in detection accuracy across all categories, with Pedestrian Moderate accuracy increasing from 67.93% to 68.85%. When used in conjunction with coordinate attention, the two produce a strong synergistic amplification effect, with Pedestrian Moderate accuracy further jumping to 69.45% and Cyclist Moderate accuracy reaching 88.25%, indicating that the above multi-scale fusion strategy can effectively preserve fine-grained geometric structure without sacrificing semantic richness.

[0018] (3) The 3DD-CIoU loss function designed in this invention unifies volume overlap rate, shape proportion consistency, and orientation angle consistency into a single optimization objective, and adaptively adjusts the penalty intensity of each item through dynamic weights. Ablation experiments show that by further introducing 3DD-CIoU loss on the basis of CA and BiFPN, the accuracy of Cyclist Moderate is improved from 88.25% to 88.79%, and the accuracy of Pedestrian Moderate is improved from 69.45% to 71.02%. Visualization results also show that the bicycle and rider bounding boxes predicted by CBDNet are more accurate in orientation and have a higher fit, verifying the effectiveness of this loss function for geometric constraints; (4) The CBDNet proposed in this invention maintains a level comparable to top algorithms in the Car category (Moderate 87.75%), and achieves the best performance among the comparison algorithms in the highly challenging Pedestrian category (Moderate 71.02%) and Cyclist category (Moderate 88.79%). Especially in extremely complex urban road scenarios, the visualization results show that this invention can successfully recall vehicles parked at a distance, dense pedestrians and non-motorized vehicles that were completely missed by the baseline model, and effectively distinguish different individual instances in highly overlapping point cloud clusters. This fully demonstrates the great potential and industrial value of the method described in this invention in the practical application of autonomous driving. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the three-dimensional target detection method that integrates coordinate attention and multi-scale features provided by this invention; Figure 2 This is a schematic diagram of the CBDNet network structure in this invention; Figure 3 This is a schematic diagram of the coordinate attention (CA) module in this invention; Figure 4 This is a schematic diagram of the weighted bidirectional feature pyramid fusion (BiFPN) module in this invention; Figure 5 This is a visual comparison chart of the detection of the embodiments of the present invention and the benchmark algorithm in a conventional scenario; Figure 6 This is a visual comparison of the detection of the present invention embodiment and the benchmark algorithm in a small target scene; Figure 7 This is a visual comparison chart of the detection of the embodiments of the present invention and the benchmark algorithm in complex scenes; Figure 8 This is a schematic diagram of the three-dimensional target detection device that integrates coordinate attention and multi-scale features provided by the present invention. Detailed Implementation

[0021] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating the 3D target detection method that fuses coordinate attention and multi-scale features, as provided in an embodiment of this application. This method is based on LiDAR point cloud data and is applied to an autonomous driving environmental perception system.

[0023] The network structure (CBDNet) constructed by the above method is as follows: Figure 2 As shown, this method aims to simultaneously improve the global semantic expressiveness, cross-scale interaction capability, and geometric consistency of bounding box regression of features. A brief overview of the method is as follows: First, the original point cloud is processed by voxelization and a 3D sparse convolutional backbone network, and compressed along the height dimension, mapping it into structured 2D bird's-eye view features. In the subsequent 2D bird's-eye view (BEV) feature extraction stage, the network uses stride control to separate high-resolution and low-resolution feature branches, and embeds a coordinate attention module on this basis. This module decouples global pooling into orthogonal one-dimensional feature encoding along the X and Y directions, effectively capturing long-distance local geometric structures while explicitly preserving accurate spatial coordinates. Subsequently, the enhanced multi-scale features enter the lightweight weighted bidirectional feature pyramid fusion module (BiFPN), which uses the network's adaptively learned weight parameters to perform bidirectional feature weighted fusion, and then concatenates them in the spatial dimension through deconvolution upsampling to generate a shared 2D feature map containing rich semantics and high-fidelity positional cues. Finally, the feature map is fed into a decoupled category detection head, which independently predicts the center point heatmap and various 3D bounding box attributes for targets with significant morphological differences. A 3DD-CIoU loss function, which fuses volume intersection-over-union ratio, aspect ratio, and orientation angle penalty, is introduced to perform geometric co-optimization of the bounding boxes. The overall design of this method follows the principles of accurate localization and efficient perception. By introducing a position-sensitive attention mechanism and a weighted feature fusion network, it effectively controls computational overhead while overcoming the challenges of point cloud sparsity and scale disparities, balancing detection accuracy and inference efficiency.

[0024] The overall detection process for the CBDNet network structure described above includes the following steps: Step S1: Obtain the original 3D point cloud data and extract features from the original 3D point cloud data to generate 2D bird's-eye view features.

[0025] Specifically, the process begins by acquiring raw 3D point cloud data obtained from LiDAR scanning. The raw point cloud is then voxelized, and voxel features are extracted using Mean Voxel Feature Encoding (MeanVFE) and a 3D sparse convolutional backbone network (VoxelBackBone8x). These voxel features are then compressed along the height dimension, mapping them to 2D bird's-eye view (BEV) features. Subsequently, a 2D base BEV backbone network (BaseBEVBackbone) is used to separate high-resolution and low-resolution feature branches through stride control.

[0026] Step S2: Perform one-dimensional feature aggregation in orthogonal directions on the two-dimensional bird's-eye view features to extract direction-aware features, determine spatial attention weights based on the direction-aware features, and reweight the two-dimensional bird's-eye view features using the spatial attention weights to generate enhanced multi-scale features.

[0027] Specifically, to address the issue of location information loss caused by two-dimensional pooling, this application designs a CA module. For example... Figure 3As shown, the separated high-resolution and low-resolution BEV features are input into the coordinate attention (CA) module. The specific execution flow of the CA module is as follows: The global compression operation is explicitly decoupled into a pair of one-dimensional pooling kernels with spatial ranges of (H, 1) and (1, W), which are applied to the input features of the c-th channel along the horizontal (X-axis) and vertical (Y-axis) directions, respectively. Aggregation, mathematically expressed as: .

[0028] Subsequently, the orientation-aware feature maps are concatenated in the spatial dimension and then processed through a shared 1x1 convolution and a non-linear activation function. Perform feature fusion to generate intermediate feature f: .

[0029] After separating the fused features, they are transformed using independent 1x1 convolutional transformations and the Sigmoid function. Generate horizontal weights Weights in the vertical direction : .

[0030] Finally, element-wise multiplication is used to broadcast the attention weights in both directions and reweight them back into the original features to obtain enhanced output features. : .

[0031] In this embodiment, the CA module is embedded after each feature level of the two-dimensional basic backbone network, so that BEV features at all scales receive position-sensitive attention enhancement.

[0032] This mechanism preserves precise location priors at extremely low computational cost through a "cross-shaped" approach, overcoming semantic diffusion.

[0033] Step S3: Through a cross-scale bidirectional information flow path, the enhanced multi-scale features at different resolutions are weighted and aggregated using normalized weights obtained through adaptive learning to generate a unified shared feature map.

[0034] Specifically, to address the limitations of traditional pyramid structures that indiscriminately add equal weights, a BiFPN structure is introduced. (See [link to relevant documentation]). Figure 4 The network removes intermediate nodes with only one input edge and adds skip connections between input and output nodes at the same level, simplifying the bidirectional network topology. The fusion process abandons the high-latency Softmax function and instead uses an introduced minimum function. A fast normalization fusion strategy is proposed. In the above fusion strategy, the output feature O of each node is calculated as follows: ,in, Let i be the i-th input feature of this node. and These are the weighted parameters obtained through adaptive learning. This is the sum of the weights of all input features for the current node.

[0035] Taking the 6th layer of the network as an example, its intermediate features With output features Depth-separable convolution ( The calculation process for ) is as follows: Where Resize represents upsampling or downsampling operation (bilinear interpolation is used in this embodiment), and Conv is depthwise separable convolution (3×3 kernel).

[0036] After bidirectional weighted fusion, a unified shared feature map is generated, which simultaneously contains rich semantic information and high-fidelity spatial location clues. This weighting mechanism enables flexible guidance of multi-scale feature responses, effectively avoiding the obscuring of low-level details.

[0037] Step S4: Perform decoupled attribute prediction processing on the shared feature map to obtain the target center point heatmap and 3D bounding box attributes, and output the 3D bounding box attributes as the 3D target detection result.

[0038] Specifically, the shared feature map generated in step S3 is input into the decoupled class detection head (CenterHead). This detection head independently predicts two branches: Classification branch: Output center point heatmap, where the probability of each position represents the confidence level that a certain type of target center exists at that position.

[0039] Regression branch: Outputs the 3D bounding box attributes corresponding to each center point, including center coordinate offset (Δx, Δy), height z, 3D dimensions (l, w, h), and orientation angle r.

[0040] During the training phase, to optimize the learnable parameters in the network (including the convolutional layer parameters that generate attention weights in the coordinate attention module, the normalized weights and their convolutional parameters in the weighted bidirectional feature pyramid fusion module, and the parameters used for classification and regression in the detection head), a 3DD-CIoU loss function was designed to perform volumetric co-constraint. This loss function simultaneously constrains the 3D intersection-over-union ratio, aspect ratio consistency, and orientation angle consistency between the predicted and ground truth bounding boxes. Through gradient backpropagation, the network can automatically adjust the parameters of each module, making the final output 3D bounding boxes geometrically more accurate. The 3DD-CIoU joint loss function is defined as follows: ,in, To predict the volume intersection-over-union ratio of the 3D bounding box and the true 3D bounding box, This is a penalty item for differences in aspect ratio. For dynamic adaptive weights, For the orientation angle difference penalty term, To balance the hyperparameters.

[0041] The specific calculation methods for each item are as follows: 3D volume crossover ratio First, the predicted bounding box parameters (center coordinates, size, orientation) are reconstructed into a 3D cube. Then, the intersection volume of the predicted box and the ground truth box is calculated. And the union volume, thus obtaining: ,in, To predict the intersection volume between the 3D bounding box and the true 3D bounding box, To predict the volume of a 3D bounding box, The volume of the true 3D bounding box. This is the smoothing constant.

[0042] Length-to-width ratio difference penalty item The arctangent function is used to characterize the angle between the diagonals, and the aspect ratio difference between the predicted bounding box and the ground truth bounding box is calculated. ,in, and These are the width and length of the actual bounding box, respectively. and These represent the width and length of the predicted bounding box, respectively.

[0043] Dynamic adaptive weights This weight is based on and Automatic adjustment reduces the contribution of aspect ratio penalty when the overlap between the predicted and ground truth bounding boxes is low, thus avoiding excessive interference. The calculation formula is: .

[0044] Orientation Angle Difference Penalty Calculate the cosine distance between the predicted orientation angle and the actual orientation angle, and apply a smoothing penalty to the angle deviation. The calculation formula is as follows: ,in, To predict the orientation angle, This is the actual orientation angle.

[0045] During the inference phase, the trained detection head is used directly for forward propagation, and the 3D bounding box attributes are output as the detection result without calling the loss function.

[0046] To verify the effectiveness of the above method, the KITTI benchmark dataset was used for evaluation. KITTI is one of the most influential public datasets in the field of autonomous driving and is widely used in 3D perception research for autonomous driving. KITTI provides high-precision data acquired by LiDAR sensors, including 7481 frames of point cloud data with 3D annotations. It covers typical traffic targets such as vehicles, pedestrians, and cyclists.

[0047] To train and evaluate the model, this embodiment adopts the mainstream partitioning scheme, dividing the labeled data into a 3712-frame training set and a 3769-frame validation set. Based on the evaluation criteria defined by KITTI, AP@40 recall positions are used, and the detection accuracy and average precision based on IoU (mAP) are reported for Easy, Moderate, and Hard difficulty levels under the Car (IoU threshold of 0.7) and Pedestrian and Cyclist (IoU threshold of 0.5) categories, respectively, to quantitatively analyze the model performance.

[0048] All experiments were conducted on an Ubuntu 22.04 system using the publicly available OpenPCDet framework. The hardware environment included Python 3.8, PyTorch 1.10.0, CUDA 11.3, and an NVIDIA RTX 4090 GPU. For the KITTI dataset, the effective range of the point cloud was set to X: [0, 69.12] meters, Y: [-39.68, 39.68] meters, and Z: [-3.0, 1.0] meters. The input point cloud was voxelized, with voxels of size 0.05 × 0.05 × 0.1 cubic meters. Each voxel contained a maximum of 32 points, and a maximum of 16,000 non-empty voxels were retained for each scene. During training, the data augmentation strategies employed included: random sampling of 15 cars, 10 pedestrians, and 10 cyclists from the ground truth (GT) database for each scene for target-level data augmentation; random flipping along the X-axis; overall random rotation within the range of [-0.785, 0.785] radians; and global random scaling within the range of [0.95, 1.05].

[0049] The detection model adopts a single-stage CenterPoint no-anchor-box framework: the network first extracts voxel features through Mean Voxel Feature Encoding (MeanVFE) and a 3D sparse convolutional backbone network (VoxelBackBone8x), and compresses and maps them into 2D bird's-eye view (BEV) features along the height dimension; then, it uses a 2D base backbone network (BaseBEVBackbone) to further extract multi-scale features; finally, it directly outputs the target center point heatmap through the CenterHead, and decouples and regresses the target's center coordinates, height, size, and orientation angle, etc., as 3D bounding box attributes.

[0050] The network was trained using the AdamOneCycle optimizer with an initial learning rate of 0.003, weight decay of 0.01, momentum parameter of 0.9, and batch size per GPU of 4. The model was trained for a total of 80 epochs, with the learning rate decayed by 0.1 at the 35th and 45th epochs respectively. To maintain training stability, the maximum norm threshold for gradient clipping was set to 10.

[0051] The training results of the improved CBDNet were compared with those of the original network without modification. The results show that the improved CBDNet model exhibits better accuracy and effectiveness in 3D object detection. The results are shown in Tables 1 and 2.

[0052] Table 1. Comparison of 3D detection accuracy of different algorithms on the KITTI dataset

[0053] In the Car category, CBDNet (Ours) achieved detection accuracies of 87.75% and 83.20% at Moderate and Hard difficulties, respectively. This represents a stable performance improvement over the baseline model CenterPoint. While slightly lower than SECOND and PillarNet, which employ heavily voxelized feature extraction, in this category, its performance still comprehensively surpasses the classic PointPillars algorithm, demonstrating that CBDNet maintains a highly competitive detection level when handling standard-sized targets. For the challenging Pedestrian target category, CBDNet performs exceptionally well, achieving high accuracies of 75.34%, 71.02%, and 67.86% at Easy, Moderate, and Hard difficulties, respectively. The Moderate and Hard levels both achieved the best performance among the compared algorithms. This not only represents a further breakthrough compared to the baseline CenterPoint but also significantly outperforms PointPillars and SECOND by a substantial margin. This strongly demonstrates that the CA coordinate attention mechanism and BiFPN multi-scale fusion module introduced in this network can extremely effectively capture and amplify the fine-grained geometric structure and contextual semantics of small targets in sparse point clouds. In the Cyclist category, CBDNet also exhibits an overwhelming advantage, achieving the highest accuracy of 88.79% and 83.64% at Moderate and Hard difficulties, respectively. Compared to strong competitors such as SECOND and PillarNet, CBDNet shows a significant performance gap. Since the cyclist target has a very large aspect ratio and is extremely sensitive to orientation angle, this excellent performance profoundly verifies that the 3DD-CIoU loss function proposed in this study can provide more accurate geometric co-constraints when regressing size and angle, effectively overcoming the spatial misalignment problem caused by traditional L1 loss.

[0054] Table 2 Ablation experiments of different modules on the KITTI validation set

[0055] Ablation experiments validated the results: Experiments show that introducing only the CA module on a strong baseline still effectively extracts fine-grained features from sparse small targets, improving the Moderate accuracy of Cyclist and Pedestrian by 0.82% and 0.28%, respectively. Introducing only the BiFPN module breaks the bottleneck of unidirectional feature flow through cross-scale fusion, resulting in a balanced increase in detection accuracy across all categories. When CA and BiFPN are used together, they produce a strong synergistic amplification effect, effectively overcoming the challenge of easily missed detections of small targets in sparse point clouds, achieving Moderate accuracy of 69.45% and 88.25% for Pedestrian and Cyclist, respectively.

[0056] Finally, the 3DD-CIoU loss function is further introduced to form the complete CBDNet. This network achieves the ultimate fit of the 3D bounding box by imposing stricter constraints on the center point distance and volume intersection rate.

[0057] Overall, on CenterPoint's extremely high performance benchmark, CBDNet successfully achieved 71.02% and 88.79% accuracy for the Pedestrian and Cyclist classes, respectively, under moderate difficulty, which is of great significance for the industrial application of autonomous driving. Simultaneously, it maintained the same detection level as the benchmark model for the most critical Car class in autonomous driving. This fully demonstrates the superior effectiveness and architectural synergy of the proposed network in complex scene perception and accurate detection of small targets.

[0058] Comparison with the original model (visual analysis): To more intuitively demonstrate the effectiveness of the proposed method, a blind visual comparison was conducted. Figures 5-7 The blue box represents CBDNet, and the green box represents the Baseline.

[0059] In urban road scenarios, distant vehicles and vehicles parked consecutively on the roadside often suffer feature loss due to sparse LiDAR scanning beams and severe mutual occlusion. As shown in the attached figure, the Baseline model exhibits severe missed detections of consecutively parked vehicles at medium to long distances. In contrast, thanks to the deep semantic fusion of the feature pyramid and the enhanced attention mechanism, CBDNet is able to extract effective geometric features from extremely sparse point clouds, successfully recalling distant vehicles missed by the Baseline (as shown in the red circle), effectively improving the system's perception distance and safety redundancy on open roads.

[0060] See Figure 6 Small targets such as pedestrians and bicycles occupy very few voxels in 3D point clouds, and their shapes are easily confused with background noise such as utility poles and tree trunks, making them a traditional challenge in 3D object detection. In small target scenes, Baseline almost completely loses its ability to capture dense pedestrians and bicycles on the roadside. However, CBDNet demonstrates extremely high sensitivity to fine-grained features, accurately outlining pedestrians and non-motorized vehicles on the roadside. This proves that the model in this embodiment can still maintain a very high recall rate when processing instances with a very small number of point clouds.

[0061] See Figure 7When faced with extreme scenes involving long distances and complex lighting structures, the detection network of the Baseline model completely failed, resulting in large blind spots. In contrast, CBDNet demonstrated superior robustness. In the extremely dense region marked in red, CBDNet not only successfully identified the target group but also effectively distinguished different individual instances within highly overlapping point cloud clusters, providing clearly defined and accurately localized 3D bounding boxes. This fully validates CBDNet's excellent global receptive field and local detail capture capabilities.

[0062] Based on the visualization results above, CBDNet demonstrates significantly superior detection performance compared to the baseline across different scales, distances, and extremely complex congested scenarios. It not only drastically reduces the false negative rate but also significantly improves the understanding of weakly characterized targets (small targets, occluded targets), proving the immense application potential of the proposed network architecture in real-world autonomous driving tasks.

[0063] In addition, embodiments of this application also provide a three-dimensional target detection device that integrates coordinate attention and multi-scale features. See also Figure 8 The device includes a point cloud acquisition module, a coordinate attention module, a weighted bidirectional feature pyramid fusion module, and a detection output module.

[0064] Specifically, the point cloud acquisition module is used to acquire raw 3D point cloud data and extract features from the raw 3D point cloud data to generate 2D bird's-eye view features.

[0065] The coordinate attention module is used to aggregate one-dimensional features in orthogonal directions for the features of the two-dimensional bird's-eye view, extract direction-aware features, determine spatial attention weights based on the direction-aware features, and use the spatial attention weights to reweight the features of the two-dimensional bird's-eye view to generate enhanced multi-scale features.

[0066] The weighted bidirectional feature pyramid fusion module is used to weight and aggregate enhanced multi-scale features of different resolutions through a bidirectional information flow path across scales, using normalized weights learned through adaptive learning, to generate a unified shared feature map.

[0067] The detection output module is used to perform decoupled attribute prediction processing on the shared feature map, obtain the target center point heatmap and 3D bounding box attributes, and output the 3D bounding box attributes as the 3D target detection result.

[0068] Therefore, the technical solution of this application provides a point cloud target detection method and device that integrates coordinate attention and multi-scale features, which effectively solves the technical problems in point cloud target detection caused by sparse features, feature submersion caused by multi-scale fusion, and limited accuracy caused by isolated regression of multi-dimensional attributes.

[0069] In addition to the above embodiments, the present invention may have other implementation methods; all technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A three-dimensional target detection method integrating coordinate attention and multi-scale features, characterized in that: include: Acquire raw 3D point cloud data and extract features from the raw 3D point cloud data to generate 2D bird's-eye view features; One-dimensional feature aggregation in orthogonal directions is performed on the two-dimensional bird's-eye view features to extract direction-aware features. Spatial attention weights are determined based on the direction-aware features. The spatial attention weights are then used to reweight the two-dimensional bird's-eye view features to generate enhanced multi-scale features. By using a bidirectional information flow path across scales, the enhanced multi-scale features at different resolutions are weighted and aggregated using normalized weights obtained through adaptive learning to generate a unified shared feature map. The shared feature map is subjected to decoupled attribute prediction processing to obtain a target center point heatmap, and the three-dimensional bounding box attributes are determined based on the heatmap. The three-dimensional bounding box attributes are then output as the three-dimensional target detection result. The computational parameters required for generating the enhanced multi-scale features, generating the shared feature map, and performing the attribute prediction processing are all determined through collaborative optimization during the training phase using a joint loss function. This joint loss function integrates the three-dimensional volume intersection-union constraint, the aspect ratio difference penalty term, and the orientation angle difference penalty term.

2. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 1, characterized in that: The two-dimensional bird's-eye view features are subjected to orthogonal one-dimensional feature aggregation to extract direction-aware features, including: One-dimensional pooling kernels with spatial ranges of (H, 1) and (1, W) are used to aggregate each channel of the input features along the horizontal and vertical directions, respectively, to obtain the horizontal and vertical perceptual feature maps.

3. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 2, characterized in that: Determining spatial attention weights based on the aforementioned orientation-aware features includes: The horizontal and vertical perceptual feature maps are spliced ​​together in the spatial dimension, and the features are fused through a shared convolutional layer and a nonlinear activation function to generate intermediate fused features. The intermediate fusion features are separated along the spatial dimension and then passed through independent convolutional layers and sigmoid activation functions to generate horizontal and vertical attention weights.

4. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 1, characterized in that: The weighted aggregation of the enhanced multi-scale features at different resolutions via a cross-scale bidirectional information flow path includes: Add jump residual connection edges between input and output nodes at the same level, and remove intermediate nodes with only one input edge; Introducing a minimum value The fast normalization fusion strategy is based on the weight parameters learned by network adaptive learning to perform the weighted aggregation.

5. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 4, characterized in that: In the fast normalization fusion strategy, the output feature O of each node is calculated as follows: , among which, I i For the i-th input feature of this node, w i and w j These are the weighted parameters obtained through adaptive learning. ∑ j w j This is the sum of the weights of all input features for the current node.

6. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 1, characterized in that: The joint loss function is defined as follows: L 3DD-CIoU =1- IoU 3D + αv + λL angle ,in, IoU 3D The formula for predicting the volume intersection-union ratio (IU / U) between a 3D bounding box and the actual 3D bounding box is as follows: In the formula, V inter To predict the intersection volume between the 3D bounding box and the true 3D bounding box, V pred To predict the volume of a 3D bounding box, V gt The volume of the true 3D bounding box. It is a smoothing constant; v Penalty item for aspect ratio difference; α For dynamic adaptive weights; L angle This is a penalty term for differences in orientation angle; λ To balance the hyperparameters.

7. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 6, characterized in that: The formula for calculating the aspect ratio difference penalty term v is: , where w gt and l gt These are the width and length of the actual bounding box, w. pred and l pred These represent the width and length of the predicted bounding box, respectively.

8. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 6, characterized in that: The orientation angle difference penalty item L angle The calculation formula is: L angle =1 - cos( r pred - r gt ), where r pred To predict the orientation angle, r gt This is the actual orientation angle.

9. The three-dimensional target detection method fusing coordinate attention and multi-scale features according to claim 6, characterized in that: The dynamic adaptive weight α The calculation formula is: 。 10. A three-dimensional target detection device that integrates coordinate attention and multi-scale features, characterized in that: include: The point cloud acquisition module is used to acquire raw 3D point cloud data and extract features from the raw 3D point cloud data to generate 2D bird's-eye view features. The coordinate attention module is used to perform one-dimensional feature aggregation in orthogonal directions on the two-dimensional bird's-eye view features, extract direction-aware features, determine spatial attention weights based on the direction-aware features, and reweight the two-dimensional bird's-eye view features using the spatial attention weights to generate enhanced multi-scale features. The weighted bidirectional feature pyramid fusion module is used to weight and aggregate the enhanced multi-scale features at different resolutions through a cross-scale bidirectional information flow path and by using normalized weights learned through adaptive learning to generate a unified shared feature map. The detection output module is used to perform decoupled attribute prediction processing on the shared feature map to obtain the target center point heatmap and three-dimensional bounding box attributes, and output the three-dimensional bounding box attributes as the three-dimensional target detection result. The computational parameters required for generating the enhanced multi-scale features, generating the shared feature map, and performing the attribute prediction processing are all determined through collaborative optimization during the training phase using a joint loss function. This joint loss function integrates the three-dimensional volume intersection-union constraint, the aspect ratio difference penalty term, and the orientation angle difference penalty term.

Images