A supply-demand driven information selection method for V2X perception

Abstract

The application discloses a supply-demand driven information selection method for V2X perception, comprising the following steps: step 1: feature extraction and spatial confidence evaluation; step 2: perception demand mask and perception supply mask generation; step 3: supply-demand driven information selection; step 4: multi-source feature fusion; and step 5: perception result output. The method realizes high-value BEV feature selection and sparse transmission under the communication budget constraint by explicitly modeling the perception demand of the host vehicle and the perception supply capability of the cooperative node. The method evaluates the perception reliability of the host vehicle and the cooperative node at different BEV spatial positions by constructing a spatial confidence map, calculates the collaborative value tensor by using the continuous perception demand mask and the continuous perception supply mask, and generates a transmission mask according to the communication budget to reduce redundant feature transmission. Meanwhile, the collaborative perception efficiency and the three-dimensional target detection robustness under the bandwidth limited condition are improved through sparse feature reconstruction, coordinate alignment and multi-head attention fusion.

Description

Technical Field

[0001] This invention belongs to the field of intelligent connected vehicles and vehicle-road cooperative perception technology, specifically involving a supply-demand driven information selection method for V2X perception. Background Technology

[0002] In recent years, with the development of autonomous driving and intelligent connected vehicle technologies, vehicles and roadside units can share perception information through V2X communication, thereby alleviating problems such as limited sensor field of view, occlusion, sparse long-distance targets, and difficulty in covering blind spots in complex traffic scenarios that exist in single-vehicle perception. Vehicle-road cooperative perception, by integrating multi-source perception information from the master vehicle, cooperative vehicles, or roadside nodes, helps to improve the completeness and robustness of 3D target detection.

[0003] However, in real-world V2X collaborative perception scenarios, communication bandwidth and latency resources are limited. Directly transmitting all point cloud data or the complete BEV feature map would result in significant communication overhead and redundant transmission, impacting the real-time performance of the collaborative perception system. While existing methods can reduce communication volume through feature compression, spatial filtering, or confidence filtering, they still suffer from issues such as insufficient modeling of the master vehicle's perception requirements, inadequate matching of the supply capabilities of collaborating nodes, redundant transmission across multiple nodes, and unclear communication budget constraints.

[0004] In autonomous driving cooperative perception scenarios, existing technologies have the following shortcomings:

[0005] (1) Insufficient adaptability of communication strategies: Existing methods mostly rely on a single feature compression or spatial filtering strategy, which makes it difficult to balance perception accuracy and communication efficiency in bandwidth-limited environments. This can easily lead to redundant feature transmission or missing transmission in key areas, reducing the perception robustness in complex scenarios.

[0006] (2) Insufficient modeling of the main vehicle's perception requirements: Existing solutions usually do not fully quantify the main vehicle's perception reliability and perception supplementation requirements at different BEV spatial locations. It is difficult to accurately distinguish between areas where the main vehicle already has high confidence and areas with low confidence due to occlusion, long distance or sparse point cloud. It is easy to allocate communication resources to areas where the main vehicle does not need to supplement.

[0007] (3) Insufficient assessment of the supply capacity of collaborative nodes: Existing solutions often select information based solely on whether the collaborative nodes have observation information or local confidence, without fully judging whether the collaborative nodes can provide effective supplementation at the corresponding BEV spatial location. This may lead to the wrong selection of areas where the main vehicle demand is strong but the supply capacity of the collaborative nodes is insufficient.

[0008] (4) Imperfect supply and demand matching mechanism: Existing information selection methods usually do not jointly model the master vehicle's perceived demand and the collaborative node's perceived supply capacity. They lack a collaborative value assessment mechanism based on "strong demand and strong supply", which leads to supply and demand mismatch, limited collaborative information contribution, and waste of computing and communication resources.

[0009] (5) The problem of repeated transmission by multiple nodes is prominent: When multiple cooperating nodes participate in cooperation at the same time, different cooperating nodes may have overlapping observations of the same blind zone or the same target area. If there is no demand update mechanism based on the received area, it is easy for multiple cooperating nodes to repeatedly transmit similar BEV features, which limits bandwidth utilization and the integrity of cooperative perception. Summary of the Invention

[0010] The purpose of this invention is to address the shortcomings of existing technologies by proposing a supply-demand driven information selection method for V2X perception. This method achieves high-value BEV feature selection and sparse transmission under communication budget constraints by explicitly modeling the master vehicle's perception demand and the cooperative nodes' perception supply capabilities. The method assesses the perception reliability of the master vehicle and cooperative nodes at different BEV spatial locations by constructing a spatial confidence map, calculates the collaborative value tensor using continuous perception demand and continuous perception supply masks, and generates a transmission mask based on the communication budget to reduce redundant feature transmission. Simultaneously, it improves the collaborative perception efficiency and 3D target detection robustness under bandwidth-constrained conditions through sparse feature reconstruction, coordinate alignment, and multi-head attention fusion.

[0011] The technical solution to achieve the objective of this invention is:

[0012] A supply-demand driven information selection method for V2X awareness includes the following steps:

[0013] 1) The master vehicle and at least one cooperating node acquire local point cloud data respectively, and generate corresponding BEV feature maps through a feature extraction network. and ;

[0014] 2) The main vehicle's BEV feature map Generate the spatial confidence map of the master vehicle Collaborating nodes are based on their own BEV feature maps Generate a spatial confidence graph of collaborative nodes. The spatial confidence map is used to characterize the perceived reliability of the master vehicle or cooperative node at different BEV spatial locations.

[0015] 3) Based on the spatial confidence map of the main vehicle Generate a perceptual demand mask The perception requirement mask is used to characterize the perception supplementation requirements of the main vehicle at different BEV spatial locations.

[0016] 4) Based on the spatial confidence map of the collaborative nodes Generate a perceptual supply mask The perception supply mask is used to characterize the ability of cooperative nodes to provide effective perception supplementation at different BEV spatial locations;

[0017] 5) Based on the aforementioned perception requirement mask and the perception supply mask Calculate the collaborative value tensor The collaborative value tensor is used to characterize the collaborative contribution value of the collaborative node to the master vehicle at each BEV spatial location.

[0018] 6) Under a given communication budget constraint, expand and sort the cooperative value tensor according to its spatial location, and select the top K BEV spatial locations with the highest cooperative value to generate a transmission mask, where K is determined by the number of BEV spatial grids and the communication budget ratio.

[0019] 7) The cooperating nodes, based on the transport mask, extract the BEV feature map. Select corresponding spatial location features to form sparse BEV features, and send the sparse BEV features and corresponding spatial index or coordinate transformation information to the main vehicle.

[0020] 8) The master vehicle reconstructs the received sparse BEV features based on the transmission mask and spatial index, and aligns the reconstructed cooperative node BEV features to the master vehicle coordinate system.

[0021] 9) The main vehicle will display the BEV feature map. BEV feature map of aligned cooperative nodes The fusion process is performed to obtain fused BEV features, and the three-dimensional target detection results are output based on the fused BEV features.

[0022] A vehicle-road cooperative perception system is provided to implement the aforementioned supply-demand driven information selection method for V2X perception. The system includes:

[0023] (1) Feature extraction module, used to generate BEV feature map of the main vehicle. Generate BEV feature map of collaborating node j. ;

[0024] (2) Spatial confidence generation module, used to generate spatial confidence map of the main vehicle. Generate a spatial confidence graph of collaborative nodes j. ;

[0025] (3) Perception demand mask generation module, set on the main vehicle side, is used to generate a spatial confidence map of the main vehicle. Generate a continuous sensing demand mask ;

[0026] (4) A sensing supply mask generation module is set on the collaborative node side and is used to generate a spatial confidence map of the collaborative node. Generate a continuous sensing supply mask ;

[0027] (5) Collaborative value calculation module, set on the collaborative node side, used to calculate the continuous sensing demand mask. and the continuous sensing supply mask Calculate the collaborative value tensor ;

[0028] (6) Budget selection module, used to perform Top-K selection on the cooperative value tensor under a given communication budget constraint and generate a transmission mask;

[0029] (7) A communication module, used to transmit the corresponding sparse BEV features, spatial index, timestamp or coordinate transformation information according to the transmission mask;

[0030] (8) A reconstruction and alignment module, located on the main vehicle side, is used to reconstruct the sparse BEV features based on the transmission mask and spatial index, and align them to obtain aligned cooperative BEV features. ;

[0031] (9) Fusion module, located on the main vehicle side, used to employ a multi-head attention mechanism to analyze the main vehicle's BEV characteristics. and aligned collaborative BEV features The features of the fused BEV are obtained by fusion.

[0032] (10) Detection module, used to output the three-dimensional bounding box, target category and corresponding confidence level of the target based on the fused BEV features.

[0033] The beneficial effects of this technical solution are:

[0034] (1) Reduce invalid feature transmission: This technical solution uses continuous sensing demand masking. and continuous sensing supply mask Joint computation of collaborative value tensor This allows collaborating nodes to prioritize transmitting BEV region characteristics where the main vehicle has strong demand and its own supply capacity, reducing invalid transmissions to high-confidence areas of the main vehicle or low-reliability areas of the collaborating nodes.

[0035] (2) Reduce redundant communication overhead: Under the given communication budget constraint B, this technical solution performs Top-K selection based on the cooperative value tensor and generates a transmission mask. Only high-value sparse BEV features are transmitted, thereby reducing the communication overhead caused by transmitting the complete BEV feature map.

[0036] (3) Suppressing repeated transmissions by multiple nodes: When multiple cooperating nodes participate in the collaboration, this technical solution uses the received transmission mask. Update continuous sensing requirement mask This reduces the need for subsequent cooperating nodes to transmit data repeatedly to the already covered BEV space area, thereby improving the utilization rate of communication resources.

[0037] (4) Enhance collaborative perception capability in bandwidth-constrained scenarios: This technical solution effectively supplements the sparse BEV features transmitted by the collaborative nodes to the main vehicle BEV space through sparse feature reconstruction, coordinate alignment and multi-head attention fusion, which helps to improve the robustness of 3D target detection in occlusion, long-distance targets and complex traffic scenarios.

[0038] (5) Balancing detection accuracy and communication efficiency: This technical solution can reduce communication overhead while maintaining high detection accuracy, and is suitable for V2X collaborative sensing scenarios with limited communication bandwidth. Attached Figure Description

[0039] Figure 1 Flowchart for an embodiment;

[0040] Figure 2 This is a schematic diagram of the structure for selecting supply and demand driven information in the embodiment;

[0041] Figure 3 The curve showing the relationship between AP@0.5 and communication overhead on the DAIR-V2X dataset in the example;

[0042] Figure 4 This is a visualization result of a multi-vehicle collaborative perception scenario in the embodiment. Detailed Implementation

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this is not intended to limit the scope of the invention.

[0044] Example:

[0045] Reference Figure 1 A supply-demand driven information selection method for V2X awareness includes the following steps:

[0046] Step 1: Feature Extraction and Spatial Confidence Assessment: The main vehicle i uses its local point cloud data Processing is performed on the local point cloud data of the collaborating node j. The data is processed and divided into columnar sections or voxels within a preset sensing range. The preset sensing range is [-70.4, 70.4] meters on the X-axis and [-40, 40] meters on the Y-axis, with a pillar resolution of 0.16m × 0.16m. Each point cloud is then encoded using a PointPillars feature extraction network to generate a BEV feature map. and The BEV feature map has dimensions H=504 and W=200, where H and W represent the number of grids in the two spatial dimensions of the BEV feature map, respectively. Subsequently, the master vehicle and collaborating nodes input their respective generated BEV feature maps into the local 3D object detection head of their respective nodes. The local 3D object detection head of the corresponding node outputs the center position, category, 3D bounding box, and detection confidence of the candidate object, and combines this with the point cloud density prior to generate a spatial confidence map. The master vehicle generates... Collaborative node j is generated Taking the main vehicle as an example, the center position of the m-th candidate target output by the detection head is ( , The detection confidence level is The initial spatial confidence level is:

[0047]

[0048] Further combining the point cloud density or effective observation density at the BEV spatial location (x,y) The spatial confidence map is obtained as follows:

[0049]

[0050] in, Let be the detection confidence score for the m-th candidate target, and σ be the spatial diffusion scale parameter of the confidence score. This represents the number of point clouds or the effective observation density within the grid location (x,y), θ is the density normalization threshold, α is the fusion weight between the detection confidence and the prior point cloud density, and clip(·) indicates that the value is restricted to the range [0,1].

[0051] Collaborating node j is also based on its local BEV feature map. Together with the local 3D object detection head of the corresponding node, generate a collaborative node spatial confidence map in the manner described above. .

[0052] Step 2: Generation of perceived demand and perceived supply masks: based on the master vehicle spatial confidence map Generate a continuous sensing demand mask This is used to represent the intensity of the sensing supplementation demand of the main vehicle at different BEV spatial locations. The continuous sensing demand mask is represented as:

[0053]

[0054] Where k is the scaling factor, This is the threshold for determining demand. When S i (x,y) is higher than hour, A value close to 0 indicates that the main vehicle's need for supplemental perception in that area is relatively weak; when Below hour, An increase indicates a strong need for additional perception from the main vehicle in that area.

[0055] Based on the spatial confidence graph of collaborative nodes Generate a continuous sensing supply mask This represents the ability of a collaborative node to provide effective sensing supplementation at different BEV spatial locations. The continuous sensing supply mask is represented as:

[0056]

[0057] in, This is the supply determination threshold. When... Higher than hour, An increase indicates that the collaborating nodes have a strong perception supplementation capability in this area.

[0058] Step 3: Supply and demand driven information selection, such as... Figure 2 As shown: The main vehicle will continuously sense the demand mask. Send to collaborating nodes, which then continuously sense and supply the mask. Continuous sensing demand mask with the main vehicle Calculate the collaborative value tensor This is used to evaluate the complementary value of the cooperating node j to the master vehicle at each BEV spatial location:

[0059]

[0060] in, The higher the value of , the stronger the BEV spatial location simultaneously meets the main vehicle's perception requirements and the stronger the supply capacity of the collaborating nodes, and the more priority it should be as a transmission area. Given a communication budget constraint B, assuming the number of BEV spatial grids is N = H × W, the number of allowed transmission spatial locations is:

[0061]

[0062] Where B represents the communication budget proportion, and K represents the number of BEV spatial locations that can be selected under the current communication budget. The collaborative value tensor... Expanding by spatial location and sorting by cooperative value from high to low, the BEV spatial locations corresponding to the top K cooperative values ​​in the sorting results are selected to form a candidate transport location set. And according to the set of candidate transmission locations Generate transport mask :

[0063]

[0064] in, This represents the set of the top K BEV spatial locations in the collaborative value ranking; =1 indicates that the feature of the cooperating node j at the BEV space location (x,y) has been selected and needs to be transmitted. =0 indicates that the features at this location are not transmitted. Cooperative nodes use the transmission mask. BEV feature map generated from the collaborating node j itself Select the sparse BEV feature corresponding to the spatial location, and package the sparse BEV feature and its spatial index, timestamp, or coordinate transformation information into a communication message. The data is sent to the master vehicle. To reduce redundant transmissions of the same BEV spatial region by multiple collaborating nodes, after the master vehicle receives sparse BEV features from a collaborating node, it updates the continuous sensing requirement mask based on the transmission mask corresponding to the received features. Let... If the BEV spatial location has been supplemented by collaborating nodes, then the updated continuous sensing demand mask is represented as:

[0065]

[0066] in, This represents the continuous sensing requirement mask before the update. This represents the updated continuous sensing demand mask. Through this update method, the demand values ​​corresponding to the BEV space regions already supplemented by collaborating nodes are reduced, thereby reducing the need for subsequent collaborating nodes to repeatedly transmit data to already covered areas.

[0067] Step 4: Multi-source feature fusion: The master vehicle receives communication messages from cooperating node j. The communication message includes sparse BEV features, spatial index, timestamp, and coordinate transformation information. The host vehicle uses the spatial index and transmission mask... The received sparse BEV features are reconstructed into a collaborative BEV feature map with the same size as the main vehicle's BEV feature map; for For BEV spatial locations where =1, fill the corresponding sparse BEV features into that location; for The BEV spatial location with a value of 0 is zero-padded to maintain the integrity of the spatial dimensions of the BEV feature tensor. Subsequently, the master vehicle performs coordinate transformations based on the coordinate transformation relationship from the collaborating node to the master vehicle. The reconstructed collaborative BEV feature map is transformed to the master vehicle coordinate system to obtain the aligned collaborative BEV feature map. To fuse the main vehicle BEV feature map with the aligned collaborative BEV feature map, a multi-head attention mechanism is introduced. This mechanism uses the main vehicle BEV feature map... As query features, the aligned collaborative BEV feature map As key features and value features, respectively, linear mapping is performed:

[0068]

[0069]

[0070]

[0071] For the m-th attention head, the output is represented as:

[0072]

[0073] in, , , These represent the query, key, and value corresponding to the m-th attention head, respectively. γ is the scaling factor for the feature channels. After obtaining the outputs of multiple attention heads, the outputs of each attention head are concatenated at the channel level, and then weighted and fused with the original BEV features of the main vehicle through a dynamic fusion factor γ to obtain the final fused BEV features:

[0074]

[0075] in, This represents the fusion feature at the BEV spatial location (x, y). Indicates the original BEV characteristics of the main vehicle. to This represents the output of M attention heads, where γ is the dynamic fusion factor.

[0076] Step 5: Perception Result Output and Example Evaluation: After space filling and feature fusion in Step 4, fused BEV features are generated. These fused BEV features are input into the 3D target detection head, which outputs the target's 3D bounding box, target category, and corresponding confidence score. In example evaluation, the average accuracy (AP) at different IoU thresholds can be used as a detection performance evaluation index, such as AP@0.3, AP@0.5, and AP@0.7; simultaneously, the number of transmitted features or communication overhead can be used to evaluate communication efficiency. Figure 4 As shown, Figure 4 This document demonstrates the visualization results of this embodiment in a multi-vehicle cooperative perception scenario, including non-cooperative perception results, spatial confidence maps, perception requirement masks, cooperative node confidence maps, selection masks, redundant filtering regions, and fused 3D detection results and BEV detection results. Figure 4 As can be seen, this embodiment can determine the perception requirements based on the low-confidence area of ​​the main vehicle, and select sparse BEV features with supplementary value for transmission in combination with the perception supply capability of the cooperative nodes, thereby reducing redundant transmission and improving the 3D target detection effect in occluded areas, distant targets or low-confidence areas of the main vehicle.

[0077] Experimental setup:

[0078] The proposed method was validated using the DAIR-V2X vehicle-to-everything (V2X) cooperative dataset. The experiments were implemented using the OpenCOOD cooperative perception framework, with PyTorch as the deep learning framework. The training and testing hardware platform was an NVIDIA Tesla P100 16GB. The basic detection network used PointPillars, and the dataset was divided into training, validation, and test sets in a 7:1:2 ratio.

[0079] The point cloud perception range was set to [-70.4, 70.4] meters on the X-axis and [-40, 40] meters on the Y-axis. The pillar resolution was 0.16m × 0.16m, and the generated BEV feature map size was H=504 and W=200. During training, the AdamW optimizer was used with a weight decay of 1e-2, an initial learning rate of 1e-3, and a cosine annealing learning rate strategy. Training lasted for 30 epochs with a batch size of 4.

[0080] Evaluation indicators:

[0081] In the experimental evaluation, the average accuracy (AP) under different IoU thresholds was used as the detection performance evaluation index, and communication overhead was used to evaluate communication efficiency. Communication overhead measures the amount of data required for the cooperating node to transmit BEV features to the host vehicle; the single-frame communication overhead is defined as:

[0082]

[0083] in, Indicates the transmission mask. The number of non-zero positions in the transmission mask represents the number of BEV space grids that are actually selected and allowed to be transmitted; C represents the channel dimension of each valid feature grid; and 4 represents the number of bytes occupied by a single feature value when using the float32 format.

[0084] Ablation experiment results:

[0085] To verify the effectiveness of the supply-demand driven dynamic selection mechanism, an ablation study was conducted on the DAIR-V2X dataset to compare the proposed method with the static screening strategy. The results are shown in the table below:

[0086]

[0087] As shown in the table, the method in this example maintains relatively stable detection accuracy at AP@0.3, AP@0.5, and AP@0.7, while having lower communication overhead than the static screening strategy. Figure 3 As shown in the figure, the results demonstrate that the supply-demand driven dynamic selection mechanism can perform feature filtering based on the master vehicle's perceived demand and the cooperative nodes' perceived supply capacity, and reduce redundant feature transmissions through demand mask updates, thereby reducing communication overhead while maintaining detection performance.

Claims

1. A supply-demand driven information selection method for V2X sensing, characterized in that, Includes the following steps: 1) The master vehicle and at least one cooperating node acquire local point cloud data respectively, and generate corresponding BEV feature maps through a feature extraction network. and ; 2) The main vehicle's BEV feature map Generate the spatial confidence map of the master vehicle Collaborating nodes are based on their own BEV feature maps Generate a spatial confidence graph of collaborative nodes. The spatial confidence map is used to characterize the perceived reliability of the master vehicle or cooperative node at different BEV spatial locations. 3) Based on the spatial confidence map of the main vehicle Generate a perceptual demand mask The perception requirement mask is used to characterize the perception supplementation requirements of the main vehicle at different BEV spatial locations. 4) Based on the spatial confidence map of the collaborative nodes Generate a perceptual supply mask The perception supply mask is used to characterize the ability of cooperative nodes to provide effective perception supplementation at different BEV spatial locations; 5) Based on the aforementioned perception requirement mask and the perception supply mask Calculate the collaborative value tensor The collaborative value tensor is used to characterize the collaborative contribution value of the collaborative node to the master vehicle at each BEV spatial location. 6) Under a given communication budget constraint, expand and sort the cooperative value tensor according to its spatial location, and select the top K BEV spatial locations with the highest cooperative value to generate a transmission mask, where K is determined by the number of BEV spatial grids and the communication budget ratio. 7) Collaborating nodes based on the transmission mask From BEV feature map Selected from Based on the characteristics of the corresponding spatial location, sparse BEV features are formed, and the sparse BEV features and the corresponding spatial index or coordinate transformation information are sent to the host vehicle. 8) The master vehicle reconstructs the received sparse BEV features based on the transmission mask and spatial index, and aligns the reconstructed cooperative node BEV features to the master vehicle coordinate system. 9) The main vehicle will display the BEV feature map. , and the aligned collaborative node BEV feature map The fusion process is performed to obtain fused BEV features, and the three-dimensional target detection results are output based on the fused BEV features.

2. The supply-demand driven information selection method for V2X sensing according to claim 1, characterized in that: Step 1) involves generating the corresponding BEV feature map, which includes the following sub-steps: (1.1) The main vehicle and the cooperating nodes respectively acquire local point cloud data, and perform columnar division or voxelization on the point cloud data within the preset perception range; (1.2) The PointPillars feature extraction network is used to encode the segmented point cloud data to obtain columnar features, and the columnar features are mapped to the BEV space to generate a BEV feature map. (1.3) The preset sensing range is the X-axis [ , meters, Y-axis , [meters, Pillar resolution is] × The generated BEV feature map has a size of H×W, where , , , , , H and W are parameters determined based on the sensing task, sensor configuration, or dataset settings.

3. The supply-demand driven information selection method for V2X sensing according to claim 2, characterized in that, The spatial confidence map of the main vehicle in step 2) The result is obtained by fusing candidate target detection confidence in the BEV space with prior point cloud density. The center position of the m-th candidate target output by the 3D target detection head based on the BEV feature map is ( , The detection confidence level is The initial spatial confidence level is: By further combining the point cloud density or effective observation density Density(x,y) at the BEV spatial location (x,y), a spatial confidence map is obtained: Where σ is the confidence spatial diffusion scale parameter, α is the fusion weight between detection confidence and point cloud density prior, θ is the density normalization threshold, and clip(·) indicates that the value is restricted to the range [0,1]; the cooperative node j is also based on the BEV feature map. Together with the local 3D object detection head of the corresponding node, generate a collaborative node spatial confidence map in the manner described above. .

4. The supply-demand driven information selection method for V2X sensing according to claim 3, characterized in that, In steps 3) and 4), both the perceived demand mask and the perceived supply mask are continuous masks; wherein, the perceived demand mask is based on the spatial confidence map of the master vehicle. The generated information represents the intensity of collaborative perception demand of the master vehicle at different BEV spatial locations; the perception supply mask is based on the spatial confidence map of the collaborative nodes. Generate, used to represent the ability of collaborative nodes to provide perception supplementation at different BEV spatial locations.

5. A supply-demand driven information selection method for V2X sensing according to claim 1, characterized in that, The generation of the perceived demand mask and perceived supply mask in steps 3) and 4) includes the following steps: (1) Based on the spatial confidence map of the main vehicle Generate a continuous sensing demand mask The continuous sensing requirement mask is represented as: Where k is the scaling factor, The threshold for determining demand; when Higher than hour, (x,y) approaching 0 indicates that the main vehicle's perception supplementation needs are weak in this area; when Below hour, An increase in (x,y) indicates a strong need for additional perception from the main vehicle in that area. (2) Based on the spatial confidence map of the collaborative nodes Generate a continuous sensing supply mask The continuous sensing supply mask is represented as: in, The supply determination threshold; when Higher than hour, An increase indicates that the collaborating nodes have a strong perception supplementation capability in this area.

6. A supply-demand driven information selection method for V2X sensing according to claim 5, characterized in that, Steps 5) to 7) include the following sub-steps: (1) Mask based on the continuous sensing requirements of the master vehicle Continuous sensing supply mask for collaborative nodes Calculate the collaborative value tensor: in, Used to represent the collaborative supplementary value of the cooperating node j to the master vehicle at the BEV spatial location (x,y); (2) Under the given communication budget constraint B, let the number of BEV spatial grids N = H × W, and the number of spatial locations allowed for transmission be: The collaborative value tensor Expand the BEV spatial locations and sort them in descending order, then select the top K BEV spatial locations with the highest cooperation value to generate a transmission mask. ,in: in, This represents the set of the top K BEV spatial locations in the collaborative value ranking; (3) The cooperating nodes determine the transmission mask. From the BEV feature map of collaborative node j Select the sparse BEV feature corresponding to the spatial location, and send the sparse BEV feature and its spatial index, timestamp or coordinate transformation information to the host vehicle.

7. A supply-demand driven information selection method for V2X sensing according to claim 6, characterized in that, In step (3), the master vehicle updates the sensing requirement mask based on the received transmission mask when multiple cooperating nodes participate in the collaboration: in, This indicates the BEV spatial location that has been supplemented by cooperating nodes, thereby reducing the need for subsequent cooperating nodes to repeatedly transmit data to the already covered area.

8. A supply-demand driven information selection method for V2X sensing according to claim 6, characterized in that, In step 9), the fusion process includes the following sub-steps: (9.1) The master vehicle receives the sparse BEV features, spatial index, timestamp, and coordinate transformation information sent by the cooperating node, and processes them according to the transmission mask. Using spatial indexing, the sparse BEV features are reconstructed into a collaborative BEV feature map with the same size as the master vehicle BEV feature map, where untransmitted locations are zero-padded. (9.2) Based on the coordinate transformation relationship from the cooperating node to the master vehicle The reconstructed collaborative BEV feature map is transformed from the collaborative node coordinate system to the master vehicle coordinate system to obtain the aligned collaborative BEV feature map. ; (9.3) The main vehicle BEV feature map Aligned collaborative BEV feature map Input the attention fusion module to obtain fused BEV features and the fused BEV features Input a 3D target detection head and output a 3D bounding box, target category, and corresponding confidence score.

9. A supply-demand driven information selection method for V2X sensing according to claim 8, characterized in that, In step (9.3), the attention fusion module uses a multi-head attention mechanism to fuse the main vehicle BEV feature map and the aligned collaborative BEV feature map, using the main vehicle BEV feature map as the basis for the fusion. As query features, the aligned collaborative BEV feature map As key features and value features, respectively, linear mapping is performed: For the m-th attention head, the output is represented as: in, , , These represent the query, key, and value corresponding to the m-th attention head, respectively. The scaling factor represents the feature channel. The outputs of multiple attention heads are concatenated to obtain collaborative attention features, which are then weighted and fused with the main vehicle BEV features to obtain fused BEV features. Where γ is the fusion weight and M is the number of attention heads.

10. A vehicle-road cooperative perception system, characterized in that, This system is used to implement a supply-demand driven information selection method for V2X sensing as described in any one of claims 1-9, the system comprising: (1) Feature extraction module, used to generate BEV feature map of the main vehicle. Generate BEV feature map of collaborating node j. ; (2) Spatial confidence generation module, used to generate spatial confidence map of the main vehicle. Generate a spatial confidence graph of collaborative nodes j. ; (3) Perception demand mask generation module, set on the main vehicle side, is used to generate a spatial confidence map of the main vehicle. Generate a continuous sensing demand mask ; (4) A sensing supply mask generation module is set on the collaborative node side and is used to generate a spatial confidence map of the collaborative node. Generate a continuous sensing supply mask ; (5) Collaborative value calculation module, set on the collaborative node side, used to calculate the continuous sensing demand mask. and the continuous sensing supply mask Calculate the collaborative value tensor ; (6) Budget selection module, used to perform Top-K selection on the cooperative value tensor under a given communication budget constraint and generate a transmission mask; (7) A communication module, used to transmit the corresponding sparse BEV features, spatial index, timestamp or coordinate transformation information according to the transmission mask; (8) A reconstruction and alignment module, located on the main vehicle side, is used to reconstruct the sparse BEV features based on the transmission mask and spatial index, and align them to obtain aligned cooperative BEV features. ; (9) Fusion module, located on the main vehicle side, used to employ a multi-head attention mechanism to analyze the main vehicle's BEV characteristics. and aligned collaborative BEV features The features of the fused BEV are obtained by fusion. (10) Detection module, used to output the three-dimensional bounding box, target category and corresponding confidence level of the target based on the fused BEV features.

Images