Vehicle cooperative perception-oriented semantic value driven resource allocation method and device

Abstract

The application provides a semantic value driven resource allocation method and device for vehicle cooperative perception, the method comprises the following steps: in each perception time slot, a roadside unit first calculates a sub-area semantic value, and filters to obtain an effective task set; the perception score index of a candidate service vehicle is calculated for each effective sub-area, and the service vehicle with the highest perception score index is selected as an execution vehicle; the roadside unit generates a resource allocation strategy including a quantization bit number, a bandwidth allocation ratio and an edge computing allocation ratio according to the current system state, the execution vehicle extracts semantic features, and performs adaptive quantization according to the resource allocation strategy; the quantized semantic features are uploaded to the roadside unit through a vehicle-to-road communication link, multi-source edge fusion is performed on the roadside unit, and the fused perception result is fed back to the target vehicle.

Description

Technical Field

[0001] This invention relates to the field of data integration and analysis technology, and more specifically, to a semantic value-driven resource allocation method and apparatus for vehicle collaborative perception. Background Technology

[0002] Currently, autonomous vehicles need to continuously acquire information about surrounding vehicles, pedestrians, and obstacles in complex road scenarios such as urban intersections. While individual vehicles rely on onboard sensors such as LiDAR, millimeter-wave radar, and cameras for environmental perception, these sensors are susceptible to blind spots due to limitations in detection distance, field of view, building obstructions, large vehicle occlusion, and dynamic changes in traffic participants. This results in untimely updates to critical risk areas.

[0003] Existing methods for allocating perception tasks or selecting vehicles typically focus more on the performance quality between vehicles and sub-regions, such as distance, line-of-sight integrity, vehicle resource status, or single latency metrics. However, the urgency, security risks, and novelty requirements of the region itself are not adequately distinguished, which can easily lead to the mixing of region priority with vehicle performance quality and make it difficult to accurately determine the priority perception targets.

[0004] Furthermore, existing collaborative perception resource allocation strategies, if they adopt centralized strategy training or centralized aggregation of historical perception data from multiple vehicles, often require the uploading of a large amount of raw perception data or local experience data, which can easily increase the communication burden and bring data privacy risks. There is still room for improvement in how to complete cross-vehicle strategy learning while retaining local vehicle data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a semantic value-driven resource allocation method and apparatus for vehicle cooperative perception. The method of the present invention constructs a sub-region semantic value model that includes dynamic targets, information age, and blind spot risks. It combines PSI to perform vehicle selection, adaptive quantization bits, and joint allocation of bandwidth and edge computing. This enables roadside units to prioritize scheduling sub-regions that are more valuable for blind spot coverage of target vehicles in each perception time slot, and achieves a trade-off between information value, system latency, and energy consumption.

[0006] To solve the above problems, the technical solution of the present invention is as follows:

[0007] A semantic value-driven resource allocation method for vehicle cooperative perception includes the following steps:

[0008] Within each sensing time slot, the roadside unit first calculates the semantic value of the sub-region and filters it to obtain the effective task set;

[0009] For each valid sub-region, calculate the perception score index of candidate service vehicles, and select the service vehicle with the highest perception score index as the execution vehicle.

[0010] The roadside unit generates a resource allocation strategy based on the current system state, including the number of quantization bits, bandwidth allocation ratio, and edge computing allocation ratio, performs semantic feature extraction from vehicles, and performs adaptive quantization according to the resource allocation strategy.

[0011] The quantized semantic features are uploaded to the roadside unit via the vehicle-to-infrastructure communication link. The roadside unit performs multi-source edge fusion and feeds back the fused perception results to the target vehicle.

[0012] Preferably, the step of calculating the semantic value of a sub-region and filtering it to obtain a set of effective tasks within each sensing time slot specifically includes:

[0013] Set time slot The set of candidate sub-regions within is The service vehicle collection is For each sub-region Semantic value is constructed based on dynamic target indicator variables, information age, and blind spot indicator variables:

[0014]

[0015] in, Subregion In terms of the semantic value of time slot t, Subregion Does a dynamic target exist in time slot t? Subregion Last valid update slot, This indicates the information age of the sub-region. Indicates the AoI scaling factor. This indicates whether the sub-area is located in the blind spot of the target vehicle. Indicates the weight of the blind zone;

[0016] Based on each service vehicle and sub-area The distance between them determines the coverage relationship. Sub-regions with semantic value not lower than the threshold and covered by at least one service vehicle are identified as valid tasks, forming a set of valid tasks.

[0017] Preferably, the step of calculating the perception score index of candidate service vehicles for each valid sub-region and selecting the service vehicle with the highest perception score index as the execution vehicle specifically includes:

[0018] For each valid sub-region From the covered set Select one vehicle to perform the service; candidate vehicles Pair of sub-regions The perceptual rating index is defined as:

[0019]

[0020] in, For distance factor, As an integrity factor, To calculate the potential factor;

[0021] Will be with sub-regions The distance does not exceed the effective sensing radius. The service vehicles constitute a covered collection ; calculate the perception score index for each candidate service vehicle within the coverage set, and determine the candidate service vehicle with the highest perception score index as the execution vehicle; after each allocation, update the task load and computation potential factor of the execution vehicle.

[0022] Preferably, the roadside unit generates a resource allocation strategy based on the current system state, including the number of quantization bits, bandwidth allocation ratio, and edge computing allocation ratio. It then performs steps to extract semantic features from vehicles and adaptively quantize them according to the resource allocation strategy. Specifically, this includes: acquiring sensor data from the allocated sub-regions; stacking multiple frames of bird's-eye view intensity maps into a spatiotemporal tensor; inputting the spatiotemporal tensor into a semantic feature encoder, sequentially performing precision expansion, channel expansion, spatial pooling, and bottleneck compression to obtain semantic features; and quantizing the semantic features according to the number of quantization bits to obtain the quantized semantic features to be uploaded.

[0023] Preferably, the quantized semantic features are uploaded to the roadside unit via the vehicle-to-infrastructure communication link, and the roadside unit performs multi-source edge fusion and feeds back the fused perception results to the target vehicle. Specifically, this includes: the roadside unit allocating uplink bandwidth to each executing vehicle according to the resource allocation strategy, and allocating edge computing resources for the corresponding fusion task; each executing vehicle uploading quantized semantic features; the roadside unit performing multi-source edge fusion; and feeding back the fused perception results to the target vehicle.

[0024] Furthermore, the present invention also provides a semantic value-driven resource allocation apparatus for vehicle cooperative perception, including a processor and a memory for storing executable instructions of the processor, the processor being configured to execute the semantic value-driven resource allocation method for vehicle cooperative perception as described above by executing the executable instructions.

[0025] Compared with existing technologies, the method of this invention constructs a sub-region semantic value model that includes dynamic targets, information age, and blind spot risk. It combines PSI (Power Segment Sensor) for vehicle selection, adaptive quantization bits, and joint allocation of bandwidth and edge computing. This enables roadside units to prioritize scheduling sub-regions more valuable for blind spot coverage of target vehicles in each sensing time slot, achieving a trade-off between information value, system latency, and energy consumption. This improves the timeliness of sensing updates for blind spot-covering areas and reduces the risk of missed detections due to long-term neglect. Attached Figure Description

[0026] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0027] Figure 1 This is a flowchart of the semantic value-driven resource allocation method for vehicle cooperative perception according to the present invention.

[0028] Figure 2 This is a detailed flowchart of the semantic value-driven resource allocation method for vehicle cooperative perception according to the present invention.

[0029] Figure 3 A schematic diagram of a vehicle collaborative perception application scenario;

[0030] Figure 4 This is a schematic diagram of the occlusion scene and integrity factor. Detailed Implementation

[0031] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0032] This invention provides a semantic value-driven resource allocation method for vehicle cooperative perception. The method is applied to a vehicle-road cooperative system including a target vehicle, service vehicles, roadside units (Roadside Units), and Regions of Interest (RoIs). Roadside units are deployed at the center of an intersection. The target vehicle travels near the intersection, and multiple service vehicles continuously enter and leave the coverage area of ​​the Roadside Units. Buildings at the four corners of the intersection form static occlusions, while buses or large vehicles form dynamic occlusions and do not participate in cooperative perception. The road area within a predetermined distance in front of the target vehicle is designated as the RoI and divided into multiple non-overlapping sub-regions. Each sub-region maintains a dynamic target state, blind spot state, and the last update time.

[0033] Specifically, such as Figures 1 to 4As shown, the semantic value-driven resource allocation method for vehicle cooperative perception includes the following steps:

[0034] S1: Within each sensing time slot, the roadside unit first calculates the semantic value of the sub-region and filters it to obtain the effective task set;

[0035] Specifically, upon entering the online phase, the roadside unit first determines the RoI (Route of Interest) based on the target vehicle's location and direction of travel, and maintains the dynamic target state, blind spot state, and last update time for each sub-region. During each online sensing time slot, the roadside unit acquires or estimates the target vehicle's location and direction of travel, the location and speed of each serving vehicle, available onboard computing frequencies, wireless channel status, current task load, and obstacle locations, serving as state inputs for subsequent task selection and resource allocation.

[0036] Set time slot The set of candidate sub-regions within is The service vehicle collection is For each sub-region Semantic value is constructed based on dynamic target indicator variables, Age of Information (AoI), and blind spot indicator variables:

[0037]

[0038] in, Subregion In terms of the semantic value of time slot t, Subregion Does a dynamic target exist in time slot t? Subregion Last valid update slot, This indicates the Information Age (AoI) of the sub-region. Indicates the AoI scaling factor. This indicates whether the sub-area is located in the blind spot of the target vehicle. This represents the blind zone weight. The formula indicates that increasing the AoI only improves the semantic value of a region when a dynamic target is present; blind zone regions gain additional value due to reduced perception coverage of the target vehicle.

[0039] Let service vehicle i and sub-region i be the service vehicle i and sub-region i. The distance is the corresponding distance term, the effective sensing radius is r, and the value threshold is the corresponding threshold. It can cover a sub-region. The service vehicles constitute the coverage set, that is, the set of service vehicles whose distance from the sub-region does not exceed the effective sensing radius; the effective task set is defined as:

[0040]

[0041] Only sub-regions with semantic value not lower than the threshold and covered by at least one service vehicle are included in the collaborative perception task set of the current time slot, thus avoiding the consumption of bandwidth and computing resources for low-value or imperceptible regions.

[0042] S2: Calculate the perception score index of candidate service vehicles for each valid sub-region, and select the service vehicle with the highest perception score index as the execution vehicle.

[0043] Specifically, for each valid sub-region From the covered set Select one vehicle to perform the service. Candidate vehicles Pair of sub-regions The perceptual rating index is defined as:

[0044]

[0045] in, For distance factor, As an integrity factor, To calculate the potential factor.

[0046] The distance factor can be defined as:

[0047]

[0048] The integrity factor indicates whether there are buildings, buses, or other large vehicles obstructing the view between the candidate vehicle and the sub-region; it can be determined by whether the line connecting the candidate vehicle to the center of the sub-region intersects with an obstacle. Specifically, under line-of-sight conditions... =1, under non-line-of-sight conditions = ,in This indicates the degradation of perceived quality caused by obstacles.

[0049] The potential factor is defined as follows:

[0050]

[0051] in, The base computing ratio for service vehicle i represents the normalized level of its available onboard computing power relative to the nominal maximum onboard computing power; The number of tasks currently assigned. This is the load penalty coefficient. It is a superlinear penalty exponent. This is used to calculate the lower bound of potential. Updated after each sub-region is assigned. Furthermore, the Perception Scoring Index (PSI) is recalculated before subsequent sub-region allocation, thereby avoiding the concentration of tasks on a small number of service vehicles.

[0052] For each valid sub-region Select the candidate service vehicle with the largest PSI within the coverage set as the service vehicle to be executed:

[0053]

[0054] The task assignment results are represented by binary indicator variables:

[0055]

[0056] in, Assign indicator variables to binary tasks. Subregion Assigned to service vehicle i, This indicates that the data has not been allocated; this variable is used for subsequent data upload volume, information value, and resource allocation models.

[0057] S3: The roadside unit generates a resource allocation strategy based on the current system state, including the number of quantization bits, bandwidth allocation ratio, and edge computing allocation ratio, performs semantic feature extraction from vehicles, and performs adaptive quantization according to the resource allocation strategy.

[0058] Specifically, the service vehicle performs semantic feature extraction on the sensor data of the assigned sub-region. In a preferred embodiment, multiple frames of bird's-eye view BEV intensity maps can be stacked into a spatiotemporal tensor and input into a semantic feature encoder. The encoder may include processing modules such as precision expansion, channel expansion, spatial pooling, and bottleneck compression, or it may be replaced by image semantic slicing, point cloud voxel features, target-level features, or multimodal fusion features.

[0059] Assume the vehicle is executed The number of quantization bits is ,like Subregion Whether to assign to a vehicle , The number of semantic feature elements corresponding to each sub-region indicates the number of vehicles. The amount of uploaded semantic data is:

[0060]

[0061] A higher quantization bit count results in higher semantic feature fidelity, but also increases the amount of uploaded data and encoding computational overhead. To describe the impact of quantization precision on semantic discriminability, a discriminant gain function is introduced:

[0062]

[0063] in, To determine the efficiency coefficient, This represents the maximum number of quantization bits. This function indicates that as... Increasing the semantic discriminability improves but gradually saturates; dividing by Used for normalization according to the maximum quantization precision. It can be obtained through offline calibration and remains fixed during the online resource allocation phase.

[0064] In the optimization model, The average number of quantized bits for continuous operations; when When the value is not an integer, the vehicle uses mixed-precision quantization with adjacent integer bit widths to make the average number of quantized bits for the semantic feature elements equal to 1. .

[0065] S4: The quantized semantic features are uploaded to the roadside unit through the vehicle-road communication link. The roadside unit performs multi-source edge fusion and feeds back the fused perception results to the target vehicle.

[0066] Specifically, the roadside unit performs multi-source edge fusion based on the quantized semantic features uploaded by each service vehicle, and feeds back the fused perception results to the target vehicle.

[0067] The following models for information value, total system latency, and total system energy consumption are established to construct the joint resource allocation optimization problem shown in equation (29); among them, the federated training related terms are only included in the offline federated training stage. Considering the impact of the total system bandwidth on the semantic feature transmission capability, a bandwidth scaling factor is introduced. :

[0068]

[0069] in, This indicates the total available uplink bandwidth of the system. Indicates the reference bandwidth. This represents the minimum scaling factor under low bandwidth conditions. When ≥ hour, =1; when < hour, A value less than 1 indicates that insufficient bandwidth will limit the effective transmission of semantic features; the square root form is used to characterize the diminishing marginal returns of increasing bandwidth. This scaling factor describes the overall impact of the system's total bandwidth and is not equivalent to the subsequent vehicle bandwidth allocation ratio.

[0070] The system information value model for time slot t is given by the following formula, which is jointly determined by the semantic value of the sub-region, the task allocation relationship, the discrimination gain, and the bandwidth scaling factor:

[0071]

[0072] In ordinary online sensing time slots, for the execution of vehicles bandwidth allocation ratio This is used to calculate the uplink speed, communication latency, and communication power consumption. If the model parameters uploaded during offline federated training aggregation rounds are considered, then the vehicle... The total amount of uplink data is:

[0073]

[0074] in, The amount of data uploaded for the semantic features of the current sensing time slot. For federated training, aggregate the indicator variables for each round. This refers to the amount of data uploaded for a single model parameter upload. In a typical online sensing time slot, This item degenerates into Only in offline federated training aggregation rounds, Only then will the amount of model parameters uploaded be included in the total uplink data volume. (Vehicle) The uplink speed is:

[0075]

[0076] in, For transmission power, For channel coefficients, Let be the noise power spectral density. Based on the uplink data volume and uplink rate, the communication delay and communication energy consumption are shown in equations (14) and (15), respectively:

[0077]

[0078]

[0079] when When =0, let =0 and =0; when When >0, bandwidth allocation ratio Take the positive value.

[0080] In offline federated training aggregation rounds, roadside units also need to broadcast the aggregated global model to participating vehicles. Let... Given the downlink broadcast rate, the downlink broadcast delay is:

[0081]

[0082] The bandwidth allocation ratio must meet the requirement that the total allocation ratio does not exceed the system's available bandwidth.

[0083]

[0084] In the computational model, the vehicle-side needs to perform semantic feature extraction and adaptive quantization encoding. Its workload varies with the number of allocated sub-regions and the number of quantization bits. The semantic feature workload is defined as follows:

[0085]

[0086] in, This represents the workload of semantic feature extraction and quantization encoding on the vehicle side. This represents the number of valid sub-regions currently assigned to service vehicle i, and is equal to the final number of tasks after task allocation. ; This represents the baseline computational cost required to process a single subregion. and These represent the fixed feature extraction overhead ratio and the variable coding overhead ratio, which increases with the number of quantization bits, respectively. .

[0087] In offline federated training aggregation rounds, the local training load can also be included in the total local computing workload on the vehicle side:

[0088]

[0089] in, This indicates the number of local training rounds during offline training. This represents the training computation density. The total local computation workload on the vehicle side consists of semantic feature workload and training workload:

[0090]

[0091] It should be noted that the above contains Model parameter upload, downlink broadcast, local training, and federated aggregation terms are only used in offline federated training aggregation rounds. During normal online sensing time slots or online deployment phases, The system does not perform federated training, federated aggregation, or upload model parameters; during the online phase, it only uploads the quantized semantic features required for the current collaborative perception task, and the trained actor network outputs a resource allocation strategy based on the current system state.

[0092] The local computation latency and energy consumption at the vehicle end are shown in equations (21) and (22), respectively:

[0093]

[0094]

[0095] in, Indicates vehicle Computational frequency used for local semantic processing This indicates the energy consumption coefficient corresponding to the vehicle's computing chip.

[0096] The roadside unit edge layer performs multi-source fusion of the quantized semantic features uploaded by each vehicle and allocates computing power according to the edge computing ratio. Its edge fusion latency is:

[0097]

[0098] when When =0, let =0; when When >0, the edge computing allocation ratio Take the positive value.

[0099] Equation (23) represents the latency required for the roadside unit to perform edge fusion of the semantic features uploaded by vehicle i; the uploaded semantic data volume and the unit fusion computation coefficient in the numerator determine the fusion computation requirements, while the edge computing ratio and the available edge computing frequency of the roadside unit in the denominator determine the allocable fusion computing capacity. The edge computing resource allocation ratio satisfies:

[0100]

[0101] For each service vehicle, local processing, uploading, and edge fusion are performed sequentially; the processing links of different vehicles can be parallelized, so the system latency is taken as the maximum value of the processing link latency of each vehicle; in the offline federated training aggregation rounds, the roadside unit aggregation latency and the global model downlink broadcast latency shown in equation (16) also need to be taken into account. Online perception or online deployment stage The aggregation latency and downlink broadcast latency are both 0. The roadside unit feeds back the fused perception results to the target vehicle. The amount of feedback data is relatively small compared to the amount of semantic feature upload, so its feedback latency is not modeled separately in this embodiment. The roadside unit aggregation latency, total system latency, and total system energy consumption are as follows:

[0102]

[0103] in, This represents the aggregation calculation density of the model.

[0104]

[0105] Since the roadside unit has a stable power supply, its energy consumption is not considered in this embodiment. The total system energy consumption only includes the local computing energy consumption and uplink communication energy consumption of each service vehicle.

[0106]

[0107] Equation (25) represents the aggregated delay of the roadside unit, Equation (26) represents the total system delay, and Equation (27) represents the total system energy consumption. To simultaneously consider sensing quality, real-time performance, and energy consumption, a system utility function is constructed. :

[0108]

[0109] Equation (28) represents the system utility. The three utility weights sum to 1, corresponding to the three objectives of information value, latency, and energy consumption, respectively. The information value normalization reference term, maximum permissible system latency, and maximum permissible system energy consumption in Equation (28) are used to normalize the three objectives. This equation adopts an exponential weighted form after normalizing information value, latency, and energy consumption. In Equation (28), the greater the information value, the higher the system utility; the greater the latency or energy consumption, the lower the system utility. Therefore, maximizing system utility involves weighing these three factors.

[0110] Optimization variables include task assignment variables. Quantization bit count Bandwidth ratio and edge computing ratio The constraints include allocating each valid sub-region to only one serviceable vehicle, quantization bit range, total bandwidth constraints, edge computing resource constraints, latency limits, energy consumption limits, and minimum information value requirements. For vehicles actually undertaking upload or edge fusion tasks, and Take a positive value; vehicles that have not undertaken the corresponding task can take 0.

[0111] Furthermore, this invention defines the joint resource allocation for each time slot as the following principal optimization problem:

[0112]

[0113] In Equation (29), the main objective is to maximize system utility while satisfying constraints on unique task allocation, quantization accuracy range, total bandwidth, edge computing resources, latency, energy consumption, and minimum information value. Discrete task allocation determines which service vehicle will perform the sensing data, while continuous resource allocation determines the accuracy of uploaded semantic information, transmission resources, and edge fusion computing resources.

[0114] Due to the main optimization problem It contains both discrete task allocation variables and continuous resource allocation variables, making direct solution complex. Therefore, it is decomposed into a task allocation subproblem. and the continuous resource allocation subproblem The first subproblem For perception task allocation, given the semantic value of a sub-region, the coverage set, and the status of candidate service vehicles, determine the execution service vehicle corresponding to each valid sub-region:

[0115]

[0116]

[0117] Preferably, A sequential greedy algorithm based on PSI is used to solve the problem: first, the effective task set is filtered using equations (1) and (2), and then the vehicles are selected sequentially for each effective sub-region according to equations (3) to (7); for the current sub-region, the candidate service vehicle with the highest PSI is selected, and the task load and calculation potential factor of the vehicle are updated after each allocation. It should be noted that... exist Previously used for building and in information value As a weight for regional contribution, The target will not be multiplied repeatedly. .

[0118] exist After obtaining the task allocation results, the second subproblem For continuous resource allocation, optimize the number of quantization bits, bandwidth ratio, and edge computing ratio:

[0119]

[0120]

[0121] Equation (31) is in The continuous resource allocation subproblem is established after the task allocation results have been obtained and fixed. Instead of changing the choice of service vehicle, the quantization bit count, bandwidth ratio, and edge computing ratio are optimized based on the task allocation.

[0122] In another embodiment, the present invention also provides a semantic value-driven resource allocation apparatus for vehicle cooperative perception, including a processor and a memory for storing executable instructions of the processor, the processor being configured to perform the semantic value-driven resource allocation method for vehicle cooperative perception as described in the above embodiment by executing the executable instructions.

[0123] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A semantic value-driven resource allocation method for vehicle cooperative perception, characterized in that, The method includes: Within each sensing time slot, the roadside unit first calculates the semantic value of the sub-region and filters it to obtain the effective task set; For each valid sub-region, calculate the perception score index of candidate service vehicles, and select the service vehicle with the highest perception score index as the execution vehicle. The roadside unit generates a resource allocation strategy based on the current system state, including the number of quantization bits, bandwidth allocation ratio, and edge computing allocation ratio, performs semantic feature extraction from vehicles, and performs adaptive quantization according to the resource allocation strategy. The quantized semantic features are uploaded to the roadside unit via the vehicle-to-infrastructure communication link. The roadside unit performs multi-source edge fusion and feeds back the fused perception results to the target vehicle.

2. The semantic value-driven resource allocation method for vehicle cooperative perception according to claim 1, characterized in that, The step of calculating the semantic value of a sub-region and filtering it to obtain a set of effective tasks within each sensing time slot specifically includes: Set time slot The set of candidate sub-regions within is The service vehicle collection is For each sub-region Semantic value is constructed based on dynamic target indicator variables, information age, and blind spot indicator variables: in, Subregion In terms of the semantic value of time slot t, Subregion Does a dynamic target exist in time slot t? Subregion Last valid update slot, This indicates the information age of the sub-region. Indicates the AoI scaling factor. This indicates whether the sub-area is located in the blind spot of the target vehicle. Indicates the weight of the blind zone; Based on each service vehicle and sub-area The distance between them determines the coverage relationship. Sub-regions with semantic value not lower than the threshold and covered by at least one service vehicle are identified as valid tasks, forming a set of valid tasks.

3. The semantic value-driven resource allocation method for vehicle cooperative perception according to claim 1, characterized in that, The step of calculating the perception score index of candidate service vehicles for each valid sub-region and selecting the service vehicle with the highest perception score index as the execution vehicle specifically includes: For each valid sub-region From the covered set Select one vehicle to perform the service; candidate vehicles Pair of sub-regions The perceptual rating index is defined as: in, For distance factor, As an integrity factor, To calculate the potential factor; Will be with sub-regions The distance does not exceed the effective sensing radius. The service vehicles constitute a covered collection ; calculate the perception score index for each candidate service vehicle within the coverage set, and determine the candidate service vehicle with the highest perception score index as the execution vehicle; after each allocation, update the task load and computation potential factor of the execution vehicle.

4. The semantic value-driven resource allocation method for vehicle cooperative perception according to claim 1, characterized in that, The roadside unit generates a resource allocation strategy based on the current system state, including the number of quantization bits, bandwidth allocation ratio, and edge computing allocation ratio. It then performs steps to extract semantic features from vehicles and adaptively quantize them according to the resource allocation strategy. Specifically, this includes: acquiring sensor data from the allocated sub-regions; stacking multi-frame bird's-eye view intensity maps into a spatiotemporal tensor; inputting the spatiotemporal tensor into a semantic feature encoder, sequentially performing precision expansion, channel expansion, spatial pooling, and bottleneck compression to obtain semantic features; and quantizing the semantic features according to the number of quantization bits to obtain the quantized semantic features to be uploaded.

5. The semantic value-driven resource allocation method for vehicle cooperative perception according to claim 1, characterized in that, The quantized semantic features are uploaded to the roadside unit via the vehicle-to-infrastructure communication link. The roadside unit performs multi-source edge fusion and feeds back the fused perception results to the target vehicle. Specifically, the roadside unit allocates uplink bandwidth to each executing vehicle according to the resource allocation strategy and allocates edge computing resources for the corresponding fusion task; each executing vehicle uploads the quantized semantic features; the roadside unit performs multi-source edge fusion and feeds back the fused perception results to the target vehicle.

6. A semantic value-driven resource allocation device for vehicle cooperative perception, characterized in that, The apparatus includes a processor and a memory for storing executable instructions of the processor, the processor being configured to perform the semantic value-driven resource allocation method for vehicle cooperative perception as described in any one of claims 1 to 5 by executing the executable instructions.

Images