An AI task load-based computing resource scheduling method and system

Abstract

The application provides a kind of based on AI task load's computing resource scheduling method and system, it is related to data processing technical field, the method includes: according to AI task load feature determines the initial importance of each task, the relative relationship of vehicle trajectory and sensor position is combined to calculate collision risk coefficient, the task emergency degree priority sequence is obtained by fusing both;According to the sequence, the task is matched to the vehicle-mounted heterogeneous computing resource pool, and the initial allocation mapping table is generated;With the three sensors with the highest resource quota as the core source, the situation attention area is drawn, the distribution dispersion and morphological change rate of each sub-block in the adjacent time window are analyzed after gridding, the region situation evolution adjustment value is calculated, and the resource ratio is corrected;Real-time monitoring of resource core load, when the high-priority task is overloaded, the low-priority task resource is migrated to the high-priority task.The application improves the response timeliness and resource utilization efficiency of key AI tasks in dynamic driving scenarios.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for scheduling computing resources based on AI task load. Background Technology

[0002] In vehicle-mounted intelligent perception systems, AI tasks from multiple sensor nodes, such as image recognition and point cloud processing, are typically distributed to heterogeneous computing resource cores within the domain controller, such as GPUs, DSPs, and FPGAs, for processing. A common scheduling method is to statically or semi-statically allocate computing resources based on the initial priority or urgency of the tasks, combined with a simple load balancing strategy.

[0003] For example, one scheduling method calculates a risk coefficient for each sensor node based on the distance to or relative speed of obstacles in front of the vehicle, and then allocates a fixed quota of computing resources to different tasks based on this coefficient. However, in real-world driving scenarios, the situation around the vehicle is constantly changing. Taking a highway ramp merging scenario as an example, before a vehicle merges into the main road, the sensor nodes to its left and right rear may face a high risk of instantaneous collision due to the rapid approach of vehicles in adjacent lanes. At this time, the urgency of these nodes' tasks increases significantly. However, under existing scheduling methods, the system mostly completes resource allocation based solely on the initial risk coefficient, and the resource allocation ratio remains unchanged over a long period of time. If the situation remains fixed, the high-risk situation in the rear side area may quickly subside after a vehicle merges into the main road, while the area directly in front of the vehicle becomes more critical, for example, due to the sudden deceleration of the vehicle in front. However, due to the lack of a real-time correction mechanism for the dynamic evolution of the surrounding space, such as the movement, spread, or contraction of risk areas, the existing scheduling scheme is unable to adjust the computing resource quotas corresponding to different sensor nodes in a timely manner. This may result in core information source nodes that should have high computing power, such as the visual sensor directly in front, being in a state of insufficient resources after the situation changes, causing processing delays or decreased accuracy of critical AI tasks, thereby affecting the overall response timeliness of the system to some extent. Summary of the Invention

[0004] This invention provides a computing resource scheduling method and system based on AI task load, which improves the response timeliness and resource utilization efficiency of key AI tasks in dynamic driving scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a method for scheduling computing resources based on AI task load, the method comprising:

[0007] Step 1: Based on the initial AI task load feature vector, determine the initial urgency of the task carried by each sensor node. Combine the relative positional relationship between the physical spatial coordinates around the vehicle and the current driving trajectory of the vehicle to determine the collision risk coefficient of each sensor node relative to the vehicle. Based on the initial urgency and collision risk coefficient, calculate the spatiotemporal hazard weight value to obtain the task urgency priority sequence.

[0008] Step 2: Based on the task urgency priority sequence, match the preset heterogeneous computing resource pool within the vehicle domain controller to obtain the initial resource allocation mapping table;

[0009] Step 3: Based on the initial resource allocation mapping table, identify the three sensor nodes allocated the highest computational resource quota as core information sources; dynamically define the situational awareness area around the vehicle based on the spatial location of the three core information sources; decompose the situational awareness area into several situational sub-blocks; calculate the regional situational evolution adjustment value based on the distribution dispersion and morphological change rate of each situational sub-block between adjacent time windows, wherein the distribution dispersion is used to characterize the degree of dispersion of the spatial coordinate distribution of the target object in the same situational sub-block within adjacent time windows, the morphological change rate is used to characterize the relative change ratio of the three-dimensional envelope volume of the target cluster in the same situational sub-block within adjacent time windows, and the regional situational evolution adjustment value is used to characterize the degree of dynamic change of the overall environmental situation around the vehicle; correct the resource allocation ratio in the initial resource allocation mapping table with the regional situational evolution adjustment value to obtain the corrected resource allocation mapping table.

[0010] Step 4: Based on the revised resource allocation mapping table, monitor the current load rate of each computing resource core in real time; when the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, obtain a resource migration instruction based on the priority difference between adjacent tasks in the task urgency priority sequence; according to the resource migration instruction, dynamically allocate the computing resource cores occupied by low-priority tasks to high-priority tasks to obtain the updated computing resource scheduling scheme.

[0011] Secondly, a computing resource scheduling system based on AI task load includes:

[0012] The evaluation module is used to determine the initial urgency of the tasks carried by each sensor node based on the initial AI task load feature vector, and to determine the collision risk coefficient of each sensor node relative to the vehicle by combining the relative positional relationship between the physical spatial coordinates around the vehicle and the current driving trajectory of the vehicle. Based on the initial urgency and the collision risk coefficient, the spatiotemporal hazard weight value is calculated to obtain the task urgency priority sequence.

[0013] The mapping module is used to match the preset heterogeneous computing resource pool within the vehicle domain controller according to the priority sequence of task urgency to obtain the initial resource allocation mapping table;

[0014] The correction module is used to identify the three sensor nodes allocated the highest computing resource quota as core information sources based on the initial resource allocation mapping table; dynamically define the situational awareness area around the vehicle based on the spatial location of the three core information sources; decompose the situational awareness area into several situational sub-blocks; calculate the regional situational evolution adjustment value based on the distribution dispersion and morphological change rate of each situational sub-block between adjacent time windows, wherein the distribution dispersion is used to characterize the degree of dispersion of the spatial coordinate distribution of the target object in the same situational sub-block within adjacent time windows, the morphological change rate is used to characterize the relative change ratio of the three-dimensional envelope volume of the target cluster in the same situational sub-block within adjacent time windows, and the regional situational evolution adjustment value is used to characterize the degree of dynamic change of the overall environmental situation around the vehicle; and correct the resource allocation ratio in the initial resource allocation mapping table with the regional situational evolution adjustment value to obtain the corrected resource allocation mapping table.

[0015] The migration module is used to monitor the current load rate of each computing resource core in real time according to the revised resource allocation mapping table. When the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, a resource migration instruction is obtained based on the priority difference between adjacent tasks in the task urgency priority sequence. According to the resource migration instruction, the computing resource core occupied by the low-priority task is dynamically allocated to the high-priority task to obtain the updated computing resource scheduling scheme.

[0016] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0017] The above-described solution of the present invention has at least the following beneficial effects:

[0018] By combining initial urgency and collision risk coefficient to generate a task urgency priority sequence, and dynamically defining the situational concern area based on three core information sources with the highest resource allocation, the resource allocation ratio is corrected by block decomposition and regional situational evolution adjustment values. Finally, by combining real-time load rate monitoring and priority difference-driven resource migration, adaptive scheduling for dynamic driving scenarios can be achieved in the vehicle heterogeneous computing resource pool, improving the computing resource guarantee capability of high-priority AI tasks, reducing processing latency after situational transition, and optimizing the overall system's resource utilization efficiency and real-time response. Attached Figure Description

[0019] Figure 1This is a flowchart illustrating a computing resource scheduling method based on AI task load provided by an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of a computing resource scheduling system based on AI task load provided by an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] like Figure 1 As shown, an embodiment of the present invention proposes a computing resource scheduling method based on AI task load, the method comprising the following steps:

[0023] Step 1: Based on the initial AI task load feature vector, determine the initial urgency of the task carried by each sensor node. Combine the relative positional relationship between the physical spatial coordinates around the vehicle and the current driving trajectory of the vehicle to determine the collision risk coefficient of each sensor node relative to the vehicle. Based on the initial urgency and collision risk coefficient, calculate the spatiotemporal hazard weight value to obtain the task urgency priority sequence.

[0024] Step 2: Based on the task urgency priority sequence, match the preset heterogeneous computing resource pool within the vehicle domain controller to obtain the initial resource allocation mapping table;

[0025] Step 3: Based on the initial resource allocation mapping table, identify the three sensor nodes allocated the highest computational resource quota as core information sources; dynamically define the situational awareness area around the vehicle based on the spatial location of the three core information sources; decompose the situational awareness area into several situational sub-blocks; calculate the regional situational evolution adjustment value based on the distribution dispersion and morphological change rate of each situational sub-block between adjacent time windows, wherein the distribution dispersion is used to characterize the degree of dispersion of the spatial coordinate distribution of the target object in the same situational sub-block within adjacent time windows, the morphological change rate is used to characterize the relative change ratio of the three-dimensional envelope volume of the target cluster in the same situational sub-block within adjacent time windows, and the regional situational evolution adjustment value is used to characterize the degree of dynamic change of the overall environmental situation around the vehicle; correct the resource allocation ratio in the initial resource allocation mapping table with the regional situational evolution adjustment value to obtain the corrected resource allocation mapping table.

[0026] Step 4: Based on the revised resource allocation mapping table, monitor the current load rate of each computing resource core in real time; when the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, obtain a resource migration instruction based on the priority difference between adjacent tasks in the task urgency priority sequence; according to the resource migration instruction, dynamically allocate the computing resource cores occupied by low-priority tasks to high-priority tasks to obtain the updated computing resource scheduling scheme.

[0027] In this embodiment of the invention, a task urgency priority sequence is generated by combining the initial urgency and the collision risk coefficient. Based on the three core information sources with the highest resource allocation, the situational concern area is dynamically defined. Then, the resource allocation ratio is corrected by using block decomposition and regional situational evolution adjustment values. Finally, combined with real-time load rate monitoring and priority difference-driven resource migration, adaptive scheduling for dynamic driving scenarios can be achieved in the vehicle heterogeneous computing resource pool. This improves the computing resource guarantee capability of high-priority AI tasks, reduces processing latency after situational transition, and optimizes the overall system's resource utilization efficiency and real-time response.

[0028] In a preferred embodiment of the present invention, the multi-source raw data streams collected by each sensor node in the sensor array connected to the vehicle edge computing node within a preset time window are obtained, and the multi-source raw data streams are input into a feature extraction network to obtain the initial AI task feature vectors of each sensor node; simultaneously, the physical spatial coordinates around the vehicle and the vehicle's current driving trajectory information are obtained, and the initial AI task feature vectors, position coordinates, and driving trajectory information are concatenated and fused to obtain the initial AI task load feature vector, which may include:

[0029] In this embodiment of the invention, a preset sliding time window length is used. The preset criteria are based on the minimum decision cycle of 100ms for advanced driver assistance systems (ADAS), a single-frame sampling cycle of 20ms for LiDAR, 33ms for camera, and 50ms for millimeter-wave radar. A 100ms window can accommodate multiple frames of data from multiple sensors, preventing data loss due to an excessively short window and data lag due to an excessively long window, thus meeting the requirements for real-time perception of dynamic road conditions. The sensor data stream sampling frequency is 30Hz, and the vehicle positioning and trajectory data sampling frequency is 10Hz, based on the standard sampling frequencies of mainstream in-vehicle perception hardware. This ensures time alignment between perception and positioning data as integer multiples, avoiding time shifts. The single-sensor node task feature map vector dimension... The preset basis is that 128-dimensional features can fully cover five types of load features: task data volume, latency, fluctuation, noise, and complexity, while avoiding high-dimensional redundancy and low-dimensional underfitting, and adapting to the computing power of edge chips; vehicle position and trajectory fusion feature dimensions The preset basis is that a 32-dimensional vector can fully carry core parameters such as three-dimensional coordinates, heading angle, vehicle speed, trajectory curvature, and driving vector, and the dimension matches the task feature dimension fusion requirements; a unified feature value mapping range is also provided. The invalid data judgment threshold is set as follows: a single frame data missing rate >10%, a value exceeding the sensor range ±15%, or a timestamp offset >20ms is considered an invalid frame. The preset basis is to tolerate slight data jitter under complex vehicle operating conditions, and to remove seriously abnormal data to ensure the validity of features.

[0030] by Using a sliding window, with the window's start time as the global time zero point, multi-source raw data from the entire vehicle's sensor array is synchronously collected. This includes high-definition image data from front-view, rear-view, left-view, and right-view cameras, 3D point cloud data from the main LiDAR and blind-spot LiDAR, and ranging and speed measurement data from front and rear millimeter-wave radars. Using the vehicle's IMU inertial navigation timestamp as the reference clock, time-series calibration is performed on all sensor data. The timestamps of all data within the window are aligned to the reference time axis, and asynchronous data frames with timestamp offsets greater than 20ms are removed. This ensures that all perceived data within the same time window represents the road condition situation at the same moment. The collected raw data stream is verified frame by frame. Based on a preset invalid data judgment threshold, invalid data with missing frames, values ​​exceeding the range, abrupt changes, or transmission distortion are removed. Slightly missing data is filled by interpolation of adjacent frames to ensure that each sensor has valid and complete data frames within each time window. Data from different modal sensors undergoes unified structural processing, with image data being uniformly scaled and cropped into a 256×256 resolution three-channel image matrix. Laser point cloud data undergoes denoising, ground point removal, and effective target point clustering, and is regularized into a fixed-dimensional point set feature matrix. Radar ranging data undergoes filtering and smoothing, outputting a stable time-series numerical sequence. Finally, all modal data are unified into a standardized data format.

[0031] A lightweight convolutional feature extraction network adapted for vehicle edge devices is adopted. The specific implementation process and computational logic are as follows: Network input binding: Standardized and regularized data from a single sensor node within a single time window are sequentially input into the lightweight feature extraction network. Each sensor node infers independently without interference, ensuring independent representation of task features for each node. Multi-level feature inference computation: The network consists of a three-level structure: a basic feature extraction layer, a load feature mapping layer, and a dimensionality compression output layer. The basic feature extraction layer outputs basic features such as data shape, texture, and numerical distribution. The load feature mapping layer mines implicit AI task load features such as data latency fluctuations, data volume, task processing overhead, and noise interference intensity. The dimensionality compression output layer filters, reduces dimensionality, and removes redundancy from high-dimensional features. Feature map output rules: After the network's forward inference is completed, the output feature tensor is flattened, uniformly outputting a one-dimensional feature vector with 128 dimensions. This vector is the initial AI task feature vector of the current sensor node within the current time window. Vehicle physical space coordinates are acquired: GPS positioning data and IMU inertial compensation data are fused to output the vehicle's real-time high-precision three-dimensional spatial coordinates in the global geodetic coordinate system. ,in The horizontal axis represents the geographical coordinates pointing eastward. The vertical axis represents the northward geographical coordinates. For vehicle altitude coordinates, the coordinate update frequency is 10Hz. GPS positioning drift error is compensated by IMU acceleration integration to ensure coordinate temporal stability and strict time alignment with the 100ms sensing time window. A continuous temporal position sequence is constructed by backtracking and collecting vehicle 3D coordinates at five equally spaced historical moments (20ms, 40ms, 60ms, and 80ms) from the current sensing window as the endpoint, thus constructing a five-frame continuous temporal position sequence. To eliminate minor jitter interference in the height dimension, only the planar driving coordinates are retained for trajectory fitting, reflecting the short-term driving trend of the vehicle. The driving trajectory curve fitting and coefficient numerical range constraints are addressed by employing a quadratic polynomial least squares fitting algorithm to model the trajectory of five frames of position sequences, constructing the short-term driving trajectory equation for the vehicle. The fitting formula is as follows: The quadratic curvature coefficient of the trajectory This value characterizes the curvature of the vehicle's trajectory and determines the degree of steering aggression. Its range is [-0.05, 0.05], pre-defined based on the premise that this range fully covers the normal driving and steering conditions of passenger cars and commercial vehicles. >0 indicates a left turn in the trajectory. <0 indicates a right turn in the trajectory. ≈0 represents approximately straight-line travel; deviations outside this range are considered extreme abnormal trajectories or positioning noise and are directly discarded; the trajectory's first-order slope coefficient... Characterizes the overall driving angle and heading trend of the vehicle trajectory, reflecting the degree of tilt when the vehicle is traveling straight. The value range is [-1.0, 1.0]. The preset basis is the matching range of the vehicle's heading deviation on a normal road. A larger absolute value of the slope indicates a more significant deviation from the reference path; a slope close to 0 indicates the vehicle is traveling straight along the reference road. The trajectory offset constant... This represents the overall positional offset of the trajectory curve relative to the origin, without any trend meaning. It is only used for trajectory equation fitting and compensation, and its value range is [-5.0m, 5.0m]. The preset basis is the vehicle displacement limit within a single 100ms short-term window. The vehicle position offset within the short-term window will not exceed this range; if it does, it is judged as abnormal positioning drift data. The final coefficients are solved through least squares iteration. , , The solution results are truncated by interval constraints, and the coefficients are forcibly constrained within the above range to eliminate trajectory distortion caused by single-point positioning errors and sudden noise. A smooth, continuous real-time driving trajectory curve of the vehicle that conforms to the actual driving conditions is obtained by fitting.

[0032] Quantitative analysis of core driving parameters across all dimensions: real-time vehicle speed is calculated by the ratio of the Euclidean distance between the temporal coordinates of two adjacent frames to the time interval, with the speed value range constrained as follows. It covers all normal operating conditions on urban roads and highways, and is used to determine operating conditions such as high-speed merging, low-speed following, and idling. The heading angle is calculated based on the true north direction of the world coordinate system as 0°, combined with the tangent direction of the fitted trajectory, to determine the angle between the vehicle's current driving direction and the true north reference, with a value range of [value missing]. It is used to determine the vehicle's driving posture, such as going straight, turning left, turning right, and changing lanes; the trajectory curvature is calculated based on the real-time curvature parameters obtained by fitting a quadratic trajectory equation, and the curvature value range is [value missing]. A larger curvature value indicates a larger turning radius and a more abrupt driving path, while a curvature value close to 0 indicates that the vehicle is approaching a straight line, used to quantify the smoothness of the vehicle's driving trajectory. For driving state determination, a preset curvature threshold of 0.05 is used; curvature < 0.05 is classified as straight-line driving, and curvature ≥ 0.05 is classified as turning. Combined with a vehicle speed threshold of 5 km / h, a vehicle speed < 5 km / h is classified as idling stationary driving, thus classifying driving conditions. For short-term future driving path vector prediction, based on the fitted trajectory curve, two extrapolations are made to predict the path in the next 100ms and 200ms. The predicted vehicle position coordinates at each moment are used to construct the vehicle's short-term future driving vector, predicting the vehicle's subsequent driving trend and providing a preliminary prediction basis for sensor node collision risk and regional situation evolution calculation. The six core parameters mentioned above—vehicle three-dimensional spatial coordinates, real-time vehicle speed, heading angle, trajectory curvature, driving condition identifier, and future predicted driving vector—are uniformly normalized and dimensionally compressed to construct a 32-dimensional vehicle spatiotemporal state feature vector. This completely completes the structured and quantitative transformation of position and trajectory information, matching the input requirements for subsequent cross-feature fusion.

[0033] The single-sensor AI task feature vector and the vehicle spatiotemporal state feature vector are directly concatenated along the channel dimension, with the task feature first and the spatiotemporal feature second. The resulting concatenated vector is the initial fusion vector. The total dimension is calculated as follows: ,in The original fused feature dimensions after splicing; =128 represents the feature dimension of the sensor AI task; =32 represents the vehicle's spatiotemporal feature dimension; calculated as follows: =160 dimensions; To distinguish the scheduling weight ratios of task load characteristics and vehicle spatiotemporal characteristics, a dual-weight adaptive fusion formula is introduced, the formula is as follows: ,in This is the intermediate feature vector after weighted fusion; =0.8 is the AI ​​task load feature weight, which is preset based on the fact that task load is the core decision basis for resource scheduling and has the highest weight. =0.2 is the weight of the vehicle's spatiotemporal trajectory features, which is preset based on the spatiotemporal situation as an auxiliary correction parameter; This is the initial task feature vector for the 128-dimensional sensor. This is a 32-dimensional vehicle spatiotemporal feature vector; to eliminate the difference in the dimensions of the two types of features, [the following is done]... All 160-dimensional feature values ​​are normalized dimension by dimension, and all values ​​are uniformly mapped to the interval [0,1]. After the dimension-by-dimensional normalization calculation is completed, the 160-dimensional standardized feature vector is obtained. This vector is the initial AI task load feature vector.

[0034] In a preferred embodiment of the present invention, step 1 above may include:

[0035] In this embodiment of the invention, step 110 involves analyzing the initial AI task load feature vector, extracting the latency tolerance feature representing the timeliness of task processing and the data volume feature representing the scale of data processing, and performing normalization mapping on the latency tolerance feature and the data volume feature to obtain the initial importance assessment value of the task carried by each sensor node; specifically, this includes: the 160-dimensional initial AI task load feature vector is obtained by splicing and fusing sensor task load features and vehicle spatiotemporal features, with each dimension having a clear physical semantic distinction, and the core feature purification is completed according to the physical attribute discrimination rules of the vehicle AI task load features, traversing step 1... The 60-dimensional feature data, combined with the physical meaning, numerical variation patterns, and correlation with task load of each dimension, was categorized and screened. Two types of core features strongly correlated with the fundamental urgency of the sensor AI task were selected: latency tolerance features and data volume scale features. All other spatiotemporal auxiliary features unrelated to the task's own load, such as position offset features, trajectory curvature features, and heading angle features, were determined to be redundant features and removed to avoid irrelevant parameters interfering with the accuracy of task urgency assessment. The selection criterion for latency tolerance features was that the numerical value of this type of feature dimension could directly reflect the completion of a single frame sensor AI task from data acquisition to inference. The maximum allowable latency deviation, with its numerical fluctuation negatively correlated with the task's real-time sensitivity, is a core static indicator for assessing task urgency. For all latency tolerance-related dimension data obtained through comprehensive screening, dimensionality compression is performed using mean aggregation to eliminate random noise in single-dimensional features. This yields the original value of the latency tolerance feature for the current sensor node task. A lower value indicates a shorter allowable processing latency, stricter real-time constraints, and a higher basic urgency of the task; a higher value indicates a larger tolerable latency margin and relatively relaxed real-time requirements. The selection criterion for data volume scale features is that the value of this type of feature dimension can be directly... This reflects the number of effective sensor sampling points, the number of data frames to be processed, and the data storage volume within the current 100ms preset time window. The fluctuation of this value is positively correlated with the scale of task computing resource consumption. It is a core static indicator for evaluating the task's computing load level. For all data volume-related dimensions obtained through traversal and filtering, the same mean aggregation method is used to complete the dimension compression, and the original value of the data volume characteristic of the current sensor node task is calculated. The larger this value is, the larger the volume of sensor data to be processed within the current time window, the higher the AI ​​inference computing overhead, and the heavier the task load; the smaller this value is, the lower the task data processing pressure and the lower the computing power consumption.

[0036] For the original values ​​of latency tolerance and data volume characteristics corresponding to all sensor node sets, global maximum / minimum value statistics and normalization mapping are performed respectively. Since the original values ​​of the two types of characteristics have different dimensions and inconsistent distribution ranges, direct weighted calculations are not possible. Therefore, normalization is needed to map all characteristic values ​​to a standard dimensionless range of 0 to 1. For the normalized latency tolerance and data volume standard values ​​of a single sensor node, a step-by-step weighted fusion operation is performed to obtain the initial importance assessment value of the AI ​​task carried by each sensor node. The initial importance assessment value is used to statically and objectively characterize the urgency of the basic computing power requirements of the sensor task itself. It is determined solely by the task's real-time requirements and data load scale, and does not fluctuate with changes in the vehicle's external driving conditions. The formula for calculating the initial importance assessment value is as follows: ,in This is the initial urgency assessment value for the sensor node task, ranging from 0 to 1. The larger the value, the higher the basic urgency of the task itself, the more basic computing resources required, and the higher the basic priority of the task. The standard value of the delay tolerance feature after normalization processing for a single sensor node; Standard values ​​representing the data volume characteristics after normalization processing for a single sensor node; The delay tolerance feature weight coefficient is fixed at 0.6, reflecting the scheduling logic that prioritizes real-time performance over data load in intelligent driving perception tasks. The weight coefficient for data volume features is fixed at 0.4 and is used to assist in quantifying the computational power consumption level of the task; the weight strictly satisfies the constraint relationship. To ensure the output results converge stably within the standard range of 0 to 1, the specific weighted fusion calculation process is as follows: First, the normalized delay tolerance standard value corresponding to the current sensor node to be calculated is read. This value is multiplied by a fixed weighting coefficient of 0.6 to obtain the urgency component corresponding to the real-time requirement. This component is used to characterize the basic urgency brought about by the task's real-time constraints. Second, the normalized data volume standard value corresponding to the same sensor node is read. This value is multiplied by a fixed weighting coefficient of 0.4 to obtain the urgency component corresponding to the data load. This component is used to characterize the urgency of the load brought about by the task's data computing power consumption. Third, the real-time urgency component and the data load urgency component are linearly summed to obtain the initial urgency evaluation value of the task within the current time window of the sensor node. The above step-by-step calculation is performed on all sensor nodes in the vehicle one by one to finally obtain the standardized initial urgency evaluation value corresponding to each sensor node.

[0037] Step 111: Extract the physical spatial coordinates of the vehicle's surroundings from the initial AI task load feature vector. Calculate the relative distance and relative azimuth between the physical spatial coordinates and the vehicle's current trajectory. Based on the relative distance and relative azimuth, predict the expected intersection time between the target detected by each sensor node and the vehicle's path. Map the collision risk coefficient of each sensor node relative to the vehicle based on the expected intersection time. Specifically, this includes: extracting the physical spatial three-dimensional coordinates of the detection area corresponding to each sensor node from the structured spatiotemporal feature parameters contained in the initial AI task load feature vector. These coordinates are the center point coordinates of the sensor's effective detection range within the current 100ms preset time window. Each sensor node uniquely corresponds to a set of fixed detection coordinates. Simultaneously read the real-time three-dimensional spatial coordinates of the vehicle body in the world coordinate system, as well as all continuous coordinate point data of the vehicle's short-term trajectory obtained by least-squares fitting of five frames of temporal position sequence from the previous process. Unify the world coordinate system benchmark of all spatial data. Perform temporal alignment calibration on the sensor detection coordinates, vehicle body coordinates, and vehicle trajectory coordinates to eliminate the time difference of multi-source data acquisition and coordinate benchmark deviation, and establish the spatial correspondence between the vehicle's real-time path and the location of each sensor detection area.

[0038] Two core relative spatial parameters of the target position detected by each sensor node relative to the current vehicle trajectory are calculated one by one: the shortest relative Euclidean distance and the relative azimuth angle. The specific calculation process is as follows: For the shortest relative Euclidean distance calculation, the continuous short-term vehicle trajectory curve obtained from the previous process is first discretized into a finite number of ordered trajectory coordinate points according to a fixed step size, forming a complete set of trajectory coordinate points. This set contains all the spatial position information of the vehicle's short-term travel path. The center of the detection area of ​​a single sensor node is set as a spatial fixed point, and the coordinates of this fixed point are fixed. Each trajectory point in the set of trajectory coordinate points is traversed, and the calculation is performed point by point. The Euclidean distance is calculated by sequentially determining the straight-line distance between the spatial fixed point and each trajectory point. After traversing all trajectory points, a set of distance values ​​corresponding to all trajectory points of the sensor fixed point is obtained. All distance values ​​in this set are compared, and the smallest value is selected. This result is the shortest relative Euclidean distance between the sensor-detected target and the vehicle's trajectory. The shortest relative Euclidean distance characterizes the spatial proximity between the sensor's detection area and the vehicle's preset driving path. The smaller the shortest relative Euclidean distance, the closer the detection area is to the vehicle's driving path, and the higher the potential collision threat. The relative azimuth angle calculation process relies on... The unified world coordinate system completes the construction of dual vectors and the solution of the included angle, which is divided into three stages: vector construction, vector operation, and angle constraint. In the first stage, the vehicle's positive travel vector is constructed. The real-time spatial coordinates of the vehicle body are used as the vector starting point. Combined with the heading tangent direction of the vehicle's current fitted trajectory, a unit length of the vehicle's positive travel vector is constructed forward. This vector strictly fits the vehicle's current instantaneous travel direction and serves as the zero-degree reference for the azimuth angle. In the second stage, the target's azimuth vector is constructed. Similarly, the real-time spatial coordinates of the vehicle body are used as the vector starting point, and the spatial coordinates of the center of the current sensor node's detection area are used as the vector ending point. Connecting the two points forms a complete target azimuth vector. The first stage is used to characterize the spatial orientation of the detection area relative to the vehicle. The third stage is vector angle calculation. By calculating the spatial angle between the vehicle's forward travel vector and the target's azimuth vector, the initial azimuth angle is obtained. The initial angle is then standardized and constrained to a range of 0° to 180°. Here, 0° represents the target being directly in front of the vehicle's travel path, and the closer the angle is to 0°, the higher the degree of forward danger. 90° represents the target being directly to the side of the vehicle. 180° represents the target being directly behind the vehicle. This quantifies the azimuth danger level of different detection areas.

[0039] Based on the shortest relative Euclidean distance and relative azimuth angle obtained from the solution, combined with the vehicle's real-time driving speed, the estimated intersection time between the sensor-detected target and the vehicle's driving path is calculated. The estimated intersection time refers to the time required for the vehicle's driving path and the sensor-detected danger zone to spatially intersect while the vehicle maintains its current trajectory and speed. It is a core time-series parameter for quantifying instantaneous situational risk. The complete calculation process is as follows: Danger zone effectiveness screening and effective danger distance determination are performed. The danger attributes of the detected area are distinguished according to the numerical range of the relative azimuth angle. 90° is set as the critical threshold for azimuth risk. If the relative azimuth angle is less than 90°, the sensor-detected area is determined to be within the vehicle's forward coverage area and belongs to the effective danger zone. If a vehicle is traveling along its current trajectory and there is a possibility of path intersection, the shortest relative Euclidean distance obtained earlier is directly assigned as the effective danger distance. If the relative azimuth angle is greater than or equal to 90°, it is determined that the sensor detection area is located to the side or rear of the vehicle. The vehicle's current forward trajectory will not intersect with this area, and there is no risk of a frontal collision. The effective danger distance is directly set to infinity. The estimated intersection time is calculated for different scenarios. For effective danger areas where the effective danger distance is a finite value, the calculation is completed using the physical motion time relationship between distance and speed. The effective danger distance is taken as the length of the dangerous path the vehicle is about to travel, and the vehicle's current real-time speed is taken as the constant speed. The estimated intersection time is obtained by calculating the ratio between the two. The corresponding calculation formula is as follows: ,in The estimated time of rendezvous is in seconds. Effective danger distance, in meters; The vehicle's current real-time speed is expressed in meters per second. If the vehicle speed is... (Unit: km / h), then it needs to be converted first. = / 3.6 Substituted into the calculation, under the premise of constant driving speed, the smaller the effective danger distance and the faster the driving speed, the shorter the estimated intersection time obtained by the solution, which means that the instantaneous situational risk of the current sensor detection area is higher; for non-dangerous areas with infinite effective danger distance, the estimated intersection time is directly determined to be infinite, which means that there is no path intersection risk in the sensor area and the situation is always safe.

[0040] Based on a pre-defined monotonically decreasing mapping rule, the predicted intersection times of different magnitudes are uniformly mapped to a standard range of 0 to 1, resulting in a normalized collision risk coefficient. This enables a unified quantitative comparison of the situational risks of different sensor nodes. The pre-defined monotonically decreasing mapping function is based on the significant temporal decay characteristics of intelligent vehicle driving situational risks. The shorter the intersection time between the vehicle and the detection area, the more exponentially the probability of an instantaneous collision increases; conversely, the longer the intersection time, the collision risk gradually decays to a risk-free state. Linear mapping cannot adapt to the non-linear abrupt changes in vehicle situational risks. Therefore, an exponentially monotonically decreasing function is used as the core function for risk mapping. This perfectly matches the risk evolution patterns of dynamic driving scenarios such as high-speed merging, lane changing, deceleration of the vehicle in front, and approaching vehicles from the side. It can distinguish the real-time risk differences between different sensor nodes. The collision risk coefficient calculation formula used is as follows: ,in This is the collision risk coefficient corresponding to the sensor node, with a value range from 0 to 1. The larger the value, the higher the collision risk and the stronger the urgency of the mission situation. It is a natural constant; The risk mapping adjustment coefficient is fixed at 0.2. This value is calibrated based on the vehicle's 100ms real-time decision update cycle and is used to adapt to the rate of change of the vehicle's dynamic situation, ensuring that the risk update frequency is synchronized with the vehicle's perception and scheduling frequency. The k value is obtained by fitting the curve of collision risk with TTC (time of collision) under typical driving scenarios (40km / h on urban roads and 120km / h on highways). The estimated collision time is expressed in seconds. When the estimated collision time is close to 0 seconds, the collision risk coefficient is close to 1, corresponding to an extremely high risk situation. When the estimated collision time is greater than 20 seconds, the collision risk coefficient is close to 0, corresponding to a risk-free and stable situation. It can fully cover the risk range of vehicles on regular urban roads and highways under all operating conditions.

[0041] Step 112 involves weighted fusion calculation of the initial urgency assessment value and the collision risk coefficient to obtain a spatiotemporal hazard weight value characterizing the spatiotemporal urgency of the task. Specifically, this includes: for each independent sensor node, reading the initial urgency assessment value and collision risk coefficient generated for that node one by one. The initial urgency assessment value is a static task load parameter quantified based on the task data processing scale and task delay tolerance characteristics. The collision risk coefficient is a dynamic situational risk parameter derived from the vehicle's driving trajectory, relative spatial position, and path intersection duration. Both types of parameters have undergone prior normalization processing, uniformly converging to a dimensionless standard numerical range of 0 to 1. The parameters have consistent dimensions, uniform numerical ranges, and are of the same level. Comparability is ensured; no additional scaling, offset correction, or dimension conversion is required, and the data can be directly substituted into the weighted fusion algorithm for quantitative calculation. The specific calculation process of the dynamic weighted fusion algorithm is as follows: The weighted fusion weight ratio parameters are determined, and two-dimensional fixed fusion weights are set: a basic task importance weight factor and a dynamic collision risk weight factor. To achieve a balance between the inherent computing power requirements of the task and the dynamic situation risk of the vehicle, and to ensure equal reference priority, the basic task importance weight factor λ is set to 0.5, and the dynamic collision risk weight factor is set to 1−λ=0.5. Both sets of weights strictly satisfy the global weight normalization constraint, and the sum of the weight values ​​is always 1, ensuring that the fusion result will not exhibit numerical divergence or weight imbalance. The problem of imbalance is addressed by: 1) Calculating the static importance-weighted contribution component: Retrieving the initial importance assessment value S corresponding to the current sensor node, and performing a point-to-point multiplication operation between the initial importance assessment value S and the basic task importance weighting factor λ to obtain the static importance-weighted contribution component of the current sensor task. This component is used to quantify the contribution of the task's inherent data load and real-time constraints to the overall task urgency, fully preserving the static characteristics of the sensor task's basic computing power requirements. 2) Calculating the dynamic risk-weighted contribution component: Retrieving the collision risk coefficient R corresponding to the same sensor node, and performing a point-to-point multiplication operation between the collision risk coefficient R and the dynamic collision risk weighting factor (1−λ) to obtain the current... The dynamic risk-weighted contribution component of the sensor task is used to quantify the dynamic impact of the vehicle's real-time driving situation and spatial collision risk on the overall task urgency, fully absorbing the situational characteristics brought about by changes in the vehicle's dynamic operating conditions. The fourth step completes the linear fusion and summarization calculation of the two components. The static urgency-weighted contribution component and the dynamic risk-weighted contribution component are algebraically summed. Through the linear fusion of the two feature components, the limitations of single-dimensional evaluation are offset, and a spatiotemporal hazard weight value that takes into account both static computing power requirements and dynamic spatial risks is obtained, achieving a spatiotemporal coupled quantitative assessment of task urgency. Based on the above step-by-step dynamic weighted fusion calculation logic, a dedicated calculation formula for the spatiotemporal hazard weight value is adopted. ,in This is the spatiotemporal hazard weight value for sensor node tasks, ranging from 0 to 1. The larger the value, the higher the spatiotemporal overall urgency of the corresponding sensor task and the higher the priority of resource scheduling. The basic task importance weighting factor is fixed at 0.5 and is used to balance and adjust the proportion of static task load characteristics in the fusion result. The initial importance assessment value of the sensor node task calculated in step 110; The dynamic collision risk weight factor is fixed at 0.5 and is used to balance and adjust the proportion of dynamic spatial situation risk characteristics in the fusion result. The collision risk coefficient of the sensor node is calculated in step 111.

[0042] Step 113: Based on the spatiotemporal hazard weight values, the tasks carried by each sensor node are sorted in descending order to obtain a task urgency priority sequence that includes the correspondence between task identifiers and spatiotemporal hazard weight values. Specifically, this includes: for the AI ​​perception tasks corresponding to all sensor nodes on the vehicle, retrieving the unique task identifier for each sensor node. This task identifier is a preset fixed code. The task identifiers of different sensor nodes are independent of each other, without repetition, and can uniquely distinguish various perception tasks such as forward-looking, rear-looking, side-looking, LiDAR, and millimeter-wave radar. At the same time, matching the unique spatiotemporal hazard weight value calculated in step 112 within the current 100ms perception time window for each task identifier, establishing a one-to-one mapping relationship between task identifiers and spatiotemporal hazard weight values, and preventing data confusion issues such as one-to-many and many-to-one relationships. After completing the mapping and matching of all sensor tasks, a unified dataset of vehicle sensor task weights is generated. This dataset completely contains the identity information and spatiotemporal urgency quantification values ​​of all on-board perception tasks at the current moment, realizing the systematic integration of vehicle perception task data. The dataset is used as the sorting object, and the unified sorting rule is to sort in descending order of spatiotemporal hazard weight values. The specific sorting logic is as follows: the spatiotemporal hazard weight values ​​corresponding to all tasks in the dataset are compared across the entire domain, and sorted in descending order of weight values. The larger the spatiotemporal hazard weight value, the stronger the comprehensive spatiotemporal urgency of the sensor task after integrating the inherent computing power load and dynamic collision risk. The higher the real-time guarantee and safety guarantee requirements of the task during vehicle operation, the higher the corresponding computing resource scheduling priority. The smaller the spatiotemporal hazard weight value, the lower the comprehensive urgency of the task, and the lower the scheduling priority. In the special case where there are equal weight values, the original device encoding order of the sensors is maintained by default to avoid sorting disorder and missing task order problems, and to ensure that the sorting results are stable, unique and reproducible. After completing the descending sort of tasks across the entire domain, each group of task identifiers and corresponding spatiotemporal hazard weight values ​​are structured and bound and archived in the order of the final sorting to form an ordered structured data sequence, and finally generate a complete task urgency priority sequence.

[0043] Improve the utilization rate of onboard heterogeneous computing resources, while enhancing the real-time performance and accuracy of intelligent vehicle environmental perception, and ensuring the safety and stability of vehicle driving decisions.

[0044] In a preferred embodiment of the present invention, step 2 above may include:

[0045] In this embodiment of the invention, step 220 involves analyzing the task urgency priority sequence, extracting the identification information of each task in the sequence and its corresponding spatiotemporal hazard weight value, and mapping the computational resource requirement specifications corresponding to each task based on the spatiotemporal hazard weight value. Specifically, this includes: completely traversing all sorted entries within the task urgency priority sequence, reading the unique task identifier of each sensor sensing task corresponding to each entry in sorting order, and the standardized spatiotemporal hazard weight value bound to that task, completely retaining the identification information of all tasks and the quantized weight value in the range of 0 to 1, and establishing the task identifier and spatiotemporal hazard weight value mapping. The one-to-one correspondence of repetitive values ​​ensures that all perception task data in the vehicle is complete, without omissions, misalignments, or missing information. Based on the scheduling logic that higher task spatiotemporal urgency, higher dynamic protection level, and larger required computing resource margins are achieved, a linear forward mapping calculation method is used to convert the spatiotemporal hazard weight values ​​in the 0-1 interval into the minimum guaranteed computing power and minimum guaranteed memory for the corresponding task. All basic constraint parameters are set fixed and conform to the numerical range of the vehicle domain controller's operating conditions. The vehicle domain controller is integrated within the vehicle edge computing node and adapts to the computing power and memory overhead characteristics of multi-sensor parallel inference in intelligent driving. The specific calculation formula is as follows:

[0046] in This is the computing power requirement threshold for a single task, representing the minimum effective computing power capacity required for the stable operation of the task, in TOPS. This is the memory usage threshold for a single task, representing the minimum remaining memory capacity required for the stable operation of the task, in MB. This is the general basic computing power threshold for vehicle-mounted perception tasks. It is the fixed minimum basic computing power required for normal inference of all perception tasks. It is a system-preset fixed parameter with a value range of 0.5 TOPS, used to ensure the minimum computing power for regular lightweight perception inference tasks. This is a general basic memory threshold for vehicle perception tasks. It is the fixed minimum basic memory required for data caching and feature calculation of all perception tasks. It is a system preset fixed parameter with a value range of 256MB, used to carry the overhead of basic image and point cloud data caching and shallow feature calculation. The maximum computing power limit that can be allocated to a single perception task is set to 4.0 TOPS, which is the capping threshold for computing power resource allocation for a single perception task to avoid excessive competition for computing power resources by a single task. The maximum memory limit that can be allocated to a single perception task is set to 2048MB. This is the capping threshold for memory resource usage by a single perception task to prevent memory overflow from causing system lag. This represents the spatiotemporal hazard weight value for sensor node tasks, ranging from 0 to 1.

[0047] The specific calculation process is as follows: For each sensor sensing task, firstly, the unique spatiotemporal hazard weight value corresponding to that task is read and substituted into the computing power requirement calculation formula. Using a fixed base computing power of 0.5 TOPS as the baseline, the weight is linearly expanded within the range of 0.5 TOPS to 4.0 TOPS, yielding the minimum computing power requirement threshold suitable for the task's urgency. Simultaneously, the same weight value is substituted into the memory requirement calculation formula. Using a fixed base memory of 256MB as the baseline, the weight is linearly expanded within the range of 256MB to 2048MB, yielding the minimum memory usage threshold suitable for the task's urgency. When the spatiotemporal hazard weight value approaches 0, the task's computing power requirement threshold approaches 0.5 TOPS, and the memory usage threshold approaches 256MB, suitable for low-urgency, low-risk sensing tasks. Quantifying operational requirements: When the spatiotemporal hazard weight value approaches 1, the task's computing power requirement threshold approaches 4.0 TOPS, and the memory usage threshold approaches 2048MB. This adapts to the high-intensity computing demands of core perception tasks with high urgency, high collision risk, and high real-time requirements. The larger the spatiotemporal hazard weight value, the higher the overall urgency of the task, and the larger the calculated computing power requirement threshold and memory usage threshold, corresponding to the need for higher-specification computing resources. The smaller the weight value, the lower the task's resource requirements, achieving a precise positive correlation between task urgency and resource requirements. The task identifier, spatiotemporal hazard weight value, calculated computing power requirement threshold, and memory usage threshold of a single task are uniformly bound and integrated to form a complete, independent, and quantifiable computing resource requirement specification for the current perception task.

[0048] Step 221: Traverse the pre-defined heterogeneous computing resource pool within the vehicle domain controller, and collect the real-time operating status parameters of each computing resource core, including the current available computing power capacity, the current remaining memory, and the resource core type identifier. Specifically, this includes: pre-collecting all heterogeneous computing resource cores within the vehicle domain controller that can be used to support AI perception task inference, data processing, and feature computation; constructing a heterogeneous computing resource pool; the resource pool covers all idle, schedulable, and reusable computing cores of the vehicle system, including computing resources with different architectures, computing power levels, and computational characteristics, ensuring no blind spots or omissions in resource scheduling; and conducting real-time parameter collection for each independent computing resource core within the resource pool, synchronously collecting three types of key operating status parameters for each core. The specific collection definitions and specifications are as follows: collect the current available computing power capacity; read the rated total computing power and currently occupied computing power of a single computing resource core in real time; and calculate the remaining schedulable computing power through difference calculation. The calculation formula is as follows: ,in This represents the current available computing power capacity. To calculate the maximum rated computing power of the core; This parameter represents the real-time computing power of a computing core that is currently occupied by other tasks. It indicates the computing capacity of the core that can currently accommodate additional tasks. The remaining memory is collected, and the total memory capacity and currently occupied memory capacity of a single computing resource core are read in real time. The difference is used to calculate the available free memory for adding new tasks. The calculation formula is as follows: ,in This represents the current amount of memory remaining. The total memory capacity is set for the computing core. This parameter represents the real-time memory capacity currently occupied by the computing core. It indicates the space capacity that the computing core can currently support for new task data caching and computation storage. It collects resource core type identifiers and reads the inherent architectural attribute identifiers of each computing resource core to distinguish the type differences between general computing cores, accelerated computing cores, and real-time inference computing cores. It is used for fine-grained adaptation and allocation of high-priority tasks, ensuring that high-urgency tasks are matched with high-performance inference cores first. The unique core identifier, current available computing power capacity, current remaining memory, and resource core type identifier of each computing resource core are bound one by one and summarized into a real-time running status dataset of the heterogeneous computing resource pool.

[0049] Step 222 involves matching and comparing the computing resource requirement specifications with the operating status parameters to filter out computing resource cores whose current available computing power capacity is greater than the computing power requirement threshold and whose current remaining memory is greater than the memory usage threshold, thus constructing a candidate computing resource core set that meets the computing power requirements. Specifically, this includes: retrieving the quantified computing power requirement threshold and memory usage threshold from the computing resource requirement specifications of each sensing task, setting these two thresholds as the minimum hard constraints for stable task operation; any computing core that cannot simultaneously meet both constraints is deemed unable to support the real-time inference operation of the current sensing task; traversing the real-time operating parameters of each independent computing resource core in the heterogeneous computing resource pool, and sequentially performing computing power matching and memory matching judgments. The specific judgment logic is as follows: the computing power matching judgment condition is to determine whether the current available computing power capacity of the computing resource core is greater than or equal to the task's computing power requirement threshold, and the judgment formula is... When this condition is met, it means that the remaining computing power of the computing core can fully cover the inference computation overhead of the current task, and there will be no inference latency, frame rate drop, or task stuttering issues caused by insufficient computing power; if this condition is not met, it is determined that the computing power resources are insufficient; the memory matching judgment condition is to determine whether the current remaining memory of the computing resource core is greater than or equal to the task memory usage threshold, and the judgment formula is as follows. When this condition is met, it means that the remaining memory of the computing core can fully meet the caching and computational storage needs of the current task's image data, point cloud data, and feature matrix data, without memory overflow, data loss, or task interruption. If this condition is not met, the memory resources are deemed insufficient. Computing resource cores that simultaneously meet both the computing power matching condition and the memory matching condition are determined to be valid and usable resource cores and are allowed to be included in the candidate computing resource core set. Computing resource cores that only meet one condition or do not meet both conditions are determined to be invalid resource cores and are directly excluded from the schedulable resource range. After completing the full-domain resource traversal comparison, all resource cores included in the candidate computing resource core set have the hardware conditions to stably support the operation of the perception task.

[0050] Step 223: According to the order of task identifiers in the task urgency priority sequence, assign each task to the final computing resource core in the candidate computing resource core set based on its running status parameters, obtaining an initial resource allocation mapping table that includes the correspondence between task identifiers and computing resource core identifiers; specifically, this includes: reading the complete task urgency priority sequence, strictly following the predetermined order of task identifiers within the priority sequence, adhering to the core rule of allocating resources to high-urgency tasks first and low-urgency tasks later, determining the execution order of task resource allocation, ensuring that core perception tasks with higher spatiotemporal danger weights and stronger real-time and security requirements preemptively seize high-quality resources; and sequentially retrieving individual perception tasks to be allocated according to the predetermined order. Based on the computational resource requirements of the task, the computational resource cores in the candidate set of computational resource cores are further adapted and matched to the final running state. The computational cores with sufficient current computing power and memory reserves and whose resource core types match the operational attributes of the perception task are preferred as the final allocation cores to ensure the stability of task operation and maximize resource utilization efficiency. After completing the resource binding of a single task, the correspondence between the unique identifier of the current task and the unique identifier of the final allocated computational resource core is recorded. The resource allocation and binding correspondence of all perception tasks in the priority sequence are completed in turn. After all tasks are allocated, the correspondence between all task identifiers and computational resource core identifiers is summarized, structured and regularized, and finally an initial resource allocation mapping table containing all correspondences is generated.

[0051] It adapts to the real-time load status of heterogeneous in-vehicle resources, avoids resource idleness and over-occupancy, and improves the overall utilization efficiency of in-vehicle computing resources and the stability of multi-task parallel scheduling.

[0052] In a preferred embodiment of the present invention, step 3 above may include:

[0053] In this embodiment of the invention, step 330 involves extracting the sensor node identifiers and associated computing resource quota values ​​from the initial resource allocation mapping table, sorting the computing resource quota values ​​in descending order, and extracting the top three sensor node identifiers as core information sources to obtain a core information source set. Specifically, this includes: fully reading all standardized structured entries within the initial resource allocation mapping table, performing data validity checks on each entry, identifying and removing invalid or dirty data entries with missing sensor node identifiers, empty resource quota parameters, out-of-bounds quota values, or disordered mapping relationships, and retaining only standard valid data that is formatted correctly, has complete parameters, and has a unique and valid core mapping relationship between the task and resources. For each selected valid data entry, two sets of strongly correlated, quantifiable, and uniquely corresponding core correlation parameters are extracted: the unique identifier of the sensor node and the original value of the computational resource quota for the corresponding perception task. Simultaneously, a data-by-data verification mechanism is established. Based on the extraction of these two sets of core correlation parameters, the quantitative calculation of the resource quota ratio of all sensor nodes in the vehicle is completed, and a standardized statistical table of the vehicle's sensor resource quota is constructed. The specific calculation process and rules are as follows: The total effective resource quota value across the entire domain is summarized; the original values ​​of the computational resource quota of all valid sensor nodes are iterated and summed across the entire domain to obtain the total resource quota value for the vehicle's perception tasks. The calculation formula is as follows: ,in This represents the total computing resource quota for all valid sensor nodes in the vehicle. This represents the total number of currently active sensor nodes. For the first The original computing resource quota values ​​corresponding to each sensor node are calculated; the resource quota ratio is calculated for each node, and the proportion of the resource quota of a single sensor node in the total resource quota of the entire vehicle is solved based on the total resource quota of the entire domain. This proportion is the core quantitative indicator characterizing the sensor perception weight, task urgency, and resource allocation priority. The calculation formula is as follows: ,in For the first The percentage of computing resource quota for each sensor node; For the first The original resource quota values ​​for each sensor node; The total resource quota value for all sensors in the vehicle represents the total resource quota value. The higher the resource quota percentage, the greater the proportion of onboard heterogeneous computing resources occupied by the sensor's perception task, the higher the task's temporal and spatial urgency priority, the greater the amount of environmental perception data and inference computation it carries, and the higher its weight in influencing the vehicle's intelligent driving environmental perception and decision output. A structured statistical table of the total vehicle sensor resource quota is constructed by binding the unique identifier of each extracted sensor node, the corresponding original resource quota value, and the quantified resource quota percentage to form a structured data entry. All entries are then uniformly collected to generate the statistical table of the total vehicle sensor resource quota.

[0054] Using the vehicle sensor resource quota statistics table as the sole basic dataset, the core sorting benchmark is uniformly set as the sensor node resource quota percentage. A global descending sorting operation is performed, strictly following the order of resource quota percentage from largest to smallest, to rearrange all effective vehicle sensor nodes. After sorting, sensor nodes with higher rankings have higher resource occupancy ratios, stronger perception task urgency, and more prominent perception core status, belonging to the dominant information source nodes in the vehicle perception system; nodes with lower rankings have lower resource occupancy ratios and weaker task urgency, belonging to auxiliary perception information source nodes, and are considered as having lower rankings in the overall vehicle perception system. After the resource quota ratio descending sorting results are stably generated, the first, second, and third ranked sensor nodes are strictly extracted and marked as the core information sources for real-time vehicle environmental perception. All other sensor nodes are marked as auxiliary information sources. The three sets of core information source nodes are uniformly collected, deduplicated, and normalized. The node information is solidified and stored in strict order of resource quota ratio from high to low, and finally a standardized core information source set with a standardized structure, clear hierarchy, and unique data is formed.

[0055] Step 331: Obtain the three-dimensional position coordinate data of each core source in the core source set in the physical space around the vehicle. Taking the current position of the vehicle as the reference origin, calculate the spatial vector relationship between the three-dimensional position coordinate data of the three core sources and the reference origin. Use the three-dimensional position coordinates of the three core sources as spatial boundary constraints to dynamically define and generate the situational awareness region around the vehicle, including the core sources, and obtain the spatial boundary parameters of the situational awareness region. Specifically, this includes: traversing all three core source nodes within the core source set, completing the original coordinate acquisition based on the vehicle's unified world coordinate system, and completing the deviation calibration by combining the vehicle's real-time attitude parameters. The specific acquisition and calibration process is as follows: Original three-dimensional coordinate acquisition of core sources: For each core source node, call the sensor installation external parameters pre-stored in the vehicle sensor calibration database, combine with real-time positioning data, and acquire the uncalibrated three-dimensional original coordinates of the core source in the original world coordinate system. Define the original coordinate vector as... ,in The original horizontal coordinates of the core information source The original vertical coordinates of the core information source The core signal source uses the original vertical coordinates, which directly reflect the sensor hardware installation location. However, these coordinates are not adapted to real-time vehicle attitude changes and therefore exhibit attitude deviations. Vehicle attitude parameter collection involves real-time reading of the global attitude angle parameters output by the vehicle's inertial measurement unit, including the vehicle's pitch angle. Roll angle of the vehicle body Vehicle yaw angle The three types of angle parameters are the core reference parameters for attitude calibration, used to characterize the vehicle's current attitude deflection state relative to the horizontal ground. A three-dimensional attitude rotation matrix is ​​constructed to complete coordinate calibration. A standard three-dimensional spatial rotation matrix is ​​built using the vehicle's three-axis attitude angles to correct the attitude deflection of the original coordinates, eliminating spatial coordinate offset errors caused by vehicle tilt and rotation. The calibration calculation formula is as follows: ,in This is the calibrated core source three-dimensional coordinate column vector. ; The Z-axis yaw rotation matrix; This is the pitch and rotation matrix along the Y-axis; This is the X-axis roll rotation matrix; This is the column vector obtained by transposing the original coordinate row vector, where the specific expressions for the rotation matrices in each dimension are as follows:

[0056] The output calibration coordinates are used to perform global attitude correction of the original coordinates through matrix multiplication, resulting in accurate three-dimensional coordinates of the core information source that perfectly match the vehicle's current driving attitude and are free from deflection errors. The three core information sources each output independent calibration coordinates, which serve as the sole valid coordinate data source for subsequent spatial calculations.

[0057] A unified spatial reference is established across the entire domain. Coordinate alignment and spatial vector relationship calculation of core information sources are completed, enabling the conversion from absolute coordinates to relative vehicle body coordinates. This characterizes the spatial distribution of each core information source relative to the vehicle body. Specifically, the implementation and calculation process involves setting a unified spatial reference origin, uniformly setting the geometric center of the vehicle body as the origin of the three-dimensional spatial calculation reference, with the coordinates of the reference origin fixed at [value missing]. All subsequent spatial calculations, spatial distance calculations, and spatial region constructions are based solely on this origin, ensuring a unified and unbiased spatial calculation benchmark across the entire domain. Coordinate benchmark translation and alignment transforms the absolute world coordinates calibrated from the three core information sources into relative coordinates relative to the vehicle's reference origin, completing the coordinate benchmark normalization and alignment process. The conversion calculation formula is as follows: ,in The first The three-axis relative coordinates of each core information source relative to the vehicle reference origin; For the first Absolute three-dimensional coordinates of each core information source after calibration; Using the absolute coordinates of the origin as the reference point, with a value always equal to 0, the relative spatial positions of the three core information sources are obtained after transformation. , , The calculation involves determining the relative spatial vector and Euclidean distance of the core signal sources. Using the vehicle's reference origin as the vector starting point and the relative coordinates of each core signal source as the vector ending point, a three-dimensional spatial offset vector is constructed for each core signal source relative to the vehicle body. Simultaneously, the spatial straight-line distance between each core signal source and the vehicle body is calculated using the following formula: ,in For the first The three-dimensional Euclidean spatial distance between each core information source and the vehicle's reference origin represents the spatial interval between the core sensing information source and the vehicle itself, providing a basic scale basis for the subsequent situational area envelope range. Through vector and distance calculation, the spatial orientation, spatial distance, distribution angle, and spatial coverage range of the three core information sources relative to the vehicle body are fully and clearly defined.

[0058] A three-dimensional convex hull iterative growth algorithm is used to dynamically construct the situation region, expand the safety margin, and analyze boundary parameters. The algorithm fully refines the solution of the plane point-normal equation, the entire iterative growth process, and the basis and values ​​of the margin threshold parameters. This enables the generation of a continuous, closed three-dimensional situation region from discrete three-point spatial locations. The specific calculation and implementation process is as follows: Input the original convex hull control points, and then calibrate the relative spatial locations of the three core information sources. , , As the original constraint control points of the 3D convex envelope algorithm, the three points together define the spatial distribution boundary of the current vehicle core perception system, which is the basic constraint boundary of the situation area. Based on solving the basic envelope plane using vector cross product and dot normal equations, any two of the three points are selected to construct spatial edge vectors, and vectors are constructed respectively. ,vector The plane normal vector is solved by the cross product of two vectors. The cross product formula is: , in Let be the normal vector of the target plane. These are the three-axis components of the normal vector, based on the solved plane normal vector, combined with any point in the plane. Substituting the equations into the standard space plane point normal form, we obtain the unique fundamental envelope plane equation. The complete point normal form calculation formula is as follows: Expanding and rearranging, we obtain the standard general form of the plane equation. The constant term This equation forms a core basic envelope plane composed of three points, which accurately locks the basic spatial plane boundary of the core information source distribution.

[0059] A 3D convex envelope iterative growth algorithm is executed to construct the minimum closed convex geometry. Using the basic triangular plane as the initial seed surface, a closed iterative process involving boundary surface iterative growth, visibility determination, surface reconstruction, and redundant surface removal generates a complete closed 3D minimum convex envelope region. Detailed iterative steps are as follows: Initialize the seed facet... , , The resulting triangular plane serves as the initial seed face, recording the coordinates of all boundary edges, vertices, and plane normals of the face, thus constructing the initial face set and boundary edge set. Boundary edge traversal is performed, iterating through all currently unclosed boundary edges one by one. Using a single boundary edge as the growth reference, all spatial control points are traversed, and the spatial vertex furthest from the current boundary edge plane and satisfying the convex hull condition is selected as the growth expansion point. Visibility is then used for selection, determining the visibility of the expansion point using the spatial point-plane distance formula. Valid expansion points located outside the current face and at a distance greater than zero are retained, while indented points on the inner side are discarded. Coplanar points; new facet generation and iterative update: construct new triangular facets using the two endpoints of the boundary edges and valid extension points, solve the plane equation and normal vector of the new facet, add the new facet to the facet set, close the corresponding boundary edges, and update the global boundary edge set; iterative termination determination: repeatedly perform boundary retrieval, visibility determination, and new facet generation process until all boundary edges are closed and there are no unclosed boundaries, and finally generate the minimum three-dimensional convex hull geometry that is free of depressions and gaps, spatially continuous, and fully closed. This geometry is the minimum three-dimensional spatial range that can completely enclose the three core source points.

[0060] To avoid issues such as missed detection of edge targets and insufficient sensing range caused by the boundary of the situation area aligning with the core information source location, a uniform outward expansion process is performed on the minimum convex hull geometry, with a preset fixed safety expansion margin threshold. The preset threshold is designed to adapt to typical perception scenarios on urban roads and expressways for intelligent driving, balancing full coverage of near-range obstacles with filtering of invalid areas at long distances. A 2-meter safety margin completely covers boundary errors caused by sensor installation deviations, vehicle vibrations, and small-scale attitude shifts, while avoiding large-scale invalid spatial calculations. This balances perception integrity with computational efficiency. The full-domain expansion calculation formula is as follows: ,in These are the boundary points of the expanded situational area; These are the original convex hull boundary points; A preset fixed safety expansion margin threshold is set. Let be the unit outward normal vector corresponding to the boundary point, obtained by normalizing the plane normal vector. After uniform expansion across the entire region, a three-dimensional spatial region is finally formed that adapts to the current driving situation of the vehicle, fully covers the perception range of the core information source, and has safety redundancy. This region is defined as the situational awareness region at the current moment. All three-dimensional boundary points of the situational awareness region after expansion are traversed, and the global spatial limit boundary parameters are solved by extreme value comparison. The calculation formula is as follows: ,in These are the maximum and minimum boundary coordinates of the X-axis of the area of ​​concern. These are the maximum and minimum boundary coordinates of the Y-axis of the area of ​​concern. , These are the maximum and minimum boundary coordinates of the Z-axis of the area of ​​concern. , , The three-axis coordinates of all boundary points in the situation area are used as the set. The six sets of limit boundary parameters obtained from the solution are structured and collected to form the spatial boundary parameter set of the situation area of ​​concern.

[0061] Step 332: Extract the spatial boundary parameters of the area of ​​interest, analyze the geometric range parameters of the area of ​​interest, and calculate the number of grids in the three-dimensional space based on the preset grid resolution threshold. Decompose the area of ​​interest into uniformly segmented grids based on the number of grids, resulting in several spatially non-overlapping sub-blocks with unique spatial indexes. Specifically, this includes: fully retrieving the final output spatial boundary limit parameters of the entire area of ​​interest, including the maximum and minimum boundary coordinates on the X-axis, Y-axis, Z-axis, and Z-axis. Through the difference calculation of boundary coordinates in each dimension, solve for the actual extension span of the area of ​​interest in each dimension of three-dimensional space, quantifying the three-dimensional spatial size of the current area of ​​interest. The specific calculation formula is as follows: ,in The horizontal X-axis spatial span of the area of ​​interest, in meters; The longitudinal Y-axis spatial span of the area of ​​interest, in meters; This represents the vertical Z-axis spatial span of the area of ​​interest, in meters. It provides the actual coverage area of ​​the area of ​​interest in three-dimensional space at the current moment. It includes a built-in fixed standardized 3D grid resolution threshold, used to limit the single-side length of a standard cube for a single situational sub-block. This is a core parameter balancing the accuracy of environmental situational assessment with the consumption of onboard computing power. The preset 3D grid resolution threshold is [value missing]. The preset value is based on the fact that in intelligent driving vehicle environmental perception scenarios, a 0.5-meter grid scale can accurately identify small obstacles around the vehicle, nearby traffic participants, and subtle changes in the local road structure, perfectly matching the perception accuracy and environmental change granularity of vehicle sensors. Simultaneously, this scale effectively avoids the problem of lost local situational details and insufficient assessment accuracy caused by excessively large grid sizes, and also avoids the defects of redundant grid numbers, increased computing power, and decreased real-time performance caused by excessively small grid sizes, perfectly matching the performance requirements of real-time scheduling and environmental assessment accuracy requirements of vehicle embedded platforms. Based on a fixed resolution threshold, the theoretical grid division number is calculated dimension by dimension based on the actual spatial span of each dimension and the preset resolution threshold. To thoroughly ensure that the entire spatial area of ​​concern is covered without blind spots, residues, or omissions, and to avoid the problem of fragmented boundary spaces being unable to be gridded for analysis, the calculation results of the grid number in all dimensions are uniformly rounded up. The specific calculation formula is as follows: ,in The total number of grid divisions along the X-axis in the area of ​​interest; The total number of grid divisions along the Y-axis in the area of ​​interest; The total number of grid divisions along the Z-axis in the area of ​​interest; A fixed 3D grid resolution threshold is preset for the system, representing the standard side length of a single side of a cube representing a single situational awareness sub-block. Based on the total number of grid divisions determined by the X, Y, and Z axes, the overall irregular 3D situational awareness area is subjected to uniform 3D grid splitting across the entire domain. Using a preset resolution of 0.5 meters as the standard, the area is sequentially divided row-by-row, column-by-column, and layer-by-layer, decomposing the complete large-scale 3D situational awareness space into several uniformly sized cubic situational awareness sub-blocks. After grid splitting, a unique spatial index identifier is assigned to each situational awareness sub-block using a 3D hierarchical ordered coding rule. The coding rule adopts a three-axis dimensional combination coding method, with the coding format as follows: ,in X-axis grid number, Y-axis grid number, The Z-axis grid number is used, and the numbers of each dimension are sequentially increased from 0. This encoding method enables each spatial situation sub-block to be uniquely identified, located, and independently traced within the global situation area. Any sub-block can be quickly matched with its corresponding three-dimensional spatial location through the spatial index ID. Finally, all the cube units that have been split and encoded are collected to form a set of global situation sub-blocks.

[0062] Step 333: Obtain the set of spatial coordinates of target objects in each situational sub-block with a unique spatial index identifier within the current time window and the previous adjacent time window. Calculate the dispersion of the set of spatial coordinates of target objects in the situational sub-block with the same spatial index identifier within the two time windows to obtain the distribution dispersion of each situational sub-block. Specifically, for each situational sub-block corresponding to a unique spatial index identifier, retrieve the three-dimensional spatial coordinate sets of all obstacles, traffic participants, and road targets detected within the sub-block under the two consecutive time-series nodes of the previous adjacent time window and the current time window. Strictly match them one-to-one according to the spatial index. The spatial mean is used to characterize the overall center position of the target cluster within the block at a single time moment and is the benchmark parameter for calculating the distribution deviation. The mean is calculated for the data of the two time windows respectively. The calculation formula is as follows: ,in This represents the mean of the three-dimensional coordinate space of the target in the situational sub-block within the corresponding time window; Represents the previous adjacent historical time window, This serves as the time sequence identifier for the current time window. This represents the total number of valid target objects detected in this sub-block within the current time window. This represents the 3D spatial coordinates of the i-th valid target within the corresponding time window. The degree of clustering and dispersion of the target distribution at a single time moment is quantified by calculating the mean squared Euclidean distance between all target coordinates and the spatial mean. A larger variance indicates a more scattered target distribution, while a smaller variance indicates a more concentrated target distribution. The calculation formula is as follows: ,in This represents the total number of valid targets in the corresponding time window. The variance of the target coordinate distribution within the corresponding time window is used; by calculating the absolute difference between the variances of the distributions of two consecutive time windows, the dynamic change amplitude of the target distribution state within a time period is quantified, yielding the final distribution dispersion. The calculation formula is as follows: ,in This represents the distribution dispersion corresponding to the current situation sub-block. The larger the distribution dispersion value, the more drastic the changes in the aggregation and dispersion state of target objects within that spatial block, and the higher the complexity of the local spatial situation distribution. A value close to zero indicates that the target distribution state within the block is stable, with no significant situation fluctuations. The distribution dispersion dataset of all situation sub-blocks in the entire domain is obtained by calculating it one by one.

[0063] Step 334: Extract the contour boundary feature data of each situational sub-block with a unique spatial index within the current time window and the previous adjacent time window. Compare the contour boundary feature data of situational sub-blocks with the same spatial index within the two time windows, calculate the relative change ratio of the contour range of situational sub-blocks with the same spatial index within the two time windows, and obtain the morphological change rate of each situational sub-block. Specifically, this includes: for each situational sub-block with a unique spatial index, extracting the complete contour boundary information within the previous adjacent time window and the current time window, including the overall cluster outer contour boundary points and boundary orientation of all target objects within the block. The maximum envelope range is integrated to form complete temporal contour boundary feature data, ensuring that the dimensions of the temporal contour data before and after are consistent, the comparison benchmark is consistent, and the spatial blocks are completely corresponding. Based on the contour boundary feature data, a three-dimensional convex envelope volume calculation method is used to calculate the overall three-dimensional envelope volume of the target cluster within the situational sub-block under the historical time window and the current time window, respectively. This is used to quantify the overall spatial occupancy of the target within the block. The complete calculation process for the three-dimensional envelope volume of the target cluster under a single time window is as follows: construct the three-dimensional contour point set of the target cluster and complete the preprocessing; retrieve the outer contour three-dimensional boundary points of all target objects within the corresponding situational sub-block under the current time window, and integrate them. A set of original 3D contour points for the target cluster is formed. This set undergoes standardization preprocessing to remove duplicate coordinate points, spatially collinear outliers, and distant outliers, retaining only valid 3D coordinate points that accurately reflect the outer contour boundary of the target cluster. This results in a well-organized and valid contour point set, providing raw data for convex hull construction. Based on this valid contour point set, a closed 3D convex hull geometry is constructed. A global spatial traversal is performed on the preprocessed valid 3D contour point set, selecting all boundary vertices on the outermost edge of the set. Through spatial triangular facet splicing, a three-dimensional convex hull geometry without depressions or gaps and completely closed over the entire domain is constructed. This convex hull geometry is capable of completely enclosing the current... The smallest closed solid geometry of all target objects within a sub-block can represent the overall spatial occupancy and three-dimensional shape of the target cluster. The three-dimensional convex hull geometry is decomposed into several spatial tetrahedral micro-elements. Taking the local three-dimensional coordinate origin of the situation sub-block as a unified reference vertex, each independent triangular facet on the surface of the three-dimensional convex hull geometry is combined with the reference vertex to form several non-overlapping tetrahedral computational micro-elements that completely fill the interior space of the convex hull. This achieves the discretization and decomposition of complex solid geometry, providing a basic unit for volume integral calculation. For single tetrahedral micro-element volume quantization calculation, arbitrarily select a triangular facet on the convex hull surface, assuming the coordinates of the three spatial vertices of this triangular facet are... The reference origin is Construct three spatial vectors respectively The volume of a single tetrahedral infinitesimal element is calculated using the spatial vector mixed product algorithm. The formula is as follows: ,in The solid volume of a single tetrahedral element; Using the 3D spatial vector pointing from the origin to the three vertices of the triangular facet, the overall solid envelope volume is solved by global summation. This involves traversing all tetrahedral infinitesimal elements obtained from the decomposition of the 3D convex envelope geometry, summing the volumes of all individual elements, and finally obtaining the overall solid envelope volume of the target cluster within the current time window. The calculation formula is as follows: ,in This represents the overall 3D envelope volume of the target cluster within the current time window. This represents the total number of tetrahedral infinitesimals obtained by decomposing the current convex hull geometry. For the first The volume values ​​of each tetrahedral element are calculated; following the complete standardized calculation process described above, the volume of the target cluster's three-dimensional envelope corresponding to the previous adjacent time window is calculated independently. Target cluster 3D envelope volume corresponding to the current time window This involves quantifying the target space occupancy size across two temporal dimensions; using the volume at historical moments as a benchmark, calculating the relative change in the contour volume at the current moment relative to historical moments, reflecting the degree of expansion, contraction, and reconstruction of the overall block shape. The calculation formula is as follows: ,in The rate of change of the shape of the current sub-block is always greater than or equal to 0; if =0 and >0, then =1 (or a larger value); if both are 0, then =0. The larger the value of the morphological change rate, the more drastic the changes in the overall shape and spatial occupancy of the target cluster within the block, and the higher the degree of local scene reconstruction. A value close to zero indicates that the scene shape is stable and there is no obvious change in situation. The entire sub-blocks are traversed one by one to complete the calculation of the morphological change rate of all blocks.

[0064] Step 335 involves weighted fusion calculation of the distribution dispersion of each situational sub-block and the morphological change rate of each situational sub-block to assess the local situational complexity increment of each situational sub-block. The local situational complexity increments of all situational sub-blocks are then aggregated to calculate the regional situational evolution adjustment value, which characterizes the drastic degree of overall environmental change. Specifically, this includes: after calculating the distribution dispersion of each situational sub-block, further spatial normalization of the distribution dispersion is performed using a preset reference scale. Using 2.5 meters (half the vehicle length) as a baseline, calculate the normalized distribution dispersion. = / ,in The distribution dispersion corresponding to the current situation sub-block is converted into a dimensionless parameter. Then, the normalized distribution dispersion and the morphological change rate are weighted and fused. The distribution dispersion characterizes the discrete change characteristics of the target distribution, while the morphological change rate characterizes the morphological change characteristics of the scene contour. These two features are independent of each other and jointly determine the complexity of the local situation. A balanced weighted fusion strategy is adopted to achieve deep coupling of the two features. To ensure the balanced contribution of the two features and the objectivity and reliability of the fusion result, the weight factor is calculated using a feature normalization weight calculation method. The specific calculation process is as follows: the basic contribution coefficient of the distribution dispersion is set as... The basic contribution coefficient of morphological change rate is Since the impact of distribution dispersion and morphological change rate on local situational complexity is equivalent and the importance of feature dimensions is consistent, the basic contribution coefficients are set to be equal, i.e. Based on the principle of weight normalization, the formulas for calculating the weight factors of the two types of features are derived respectively.

[0065] In the formula The weighting factor corresponding to the dispersion of the distribution; The weighting factor corresponding to the rate of morphological change; Substituting the values ​​into the calculation, we can obtain the result. The two sets of weighting factors strictly satisfy the normalization constraint. This achieves a balanced fusion of distribution and morphological features with equal weight, without dimensional bias. After solving for the weight factors, the weighted fusion calculation formula is used to calculate the local situation complexity increment. The calculation formula is as follows: ,in For the first The local situational complexity increment of each situational sub-block; For the first The distribution dispersion of each sub-block of the situation; For the first The rate of change in the shape of each sub-block; This is the distribution dispersion weighting factor obtained after normalization calculation; The normalized morphological change rate weighting factor is used; the local situation complexity increments of all spatial sub-blocks are summed across the entire domain to obtain the quantitative value of the overall environmental situation evolution. The calculation formula is as follows: ,in This is a regional situation evolution adjustment value, used to uniformly quantify the overall drastic changes in the entire surrounding environment of vehicles; This represents the total number of all situational sub-blocks within the situational awareness area. This represents the increment of local situational complexity corresponding to a single situational sub-block. The larger the value of the regional situational evolution adjustment, the higher the disorder of the distribution of targets in the vehicle's surrounding environment, the more frequent the reconstruction of the scene form, and the more complex the overall perception situation.

[0066] Step 336: Based on the regional situation evolution adjustment value, determine the resource demand compensation coefficient for the sensor node tasks corresponding to each situation sub-block. Using the resource demand compensation coefficient, dynamically scale and correct the computational resource quota ratio of each sensor node task within the corresponding situation attention area in the initial resource allocation mapping table, and update the mapping relationship between task identifiers and core computational resource identifiers to obtain the corrected resource allocation mapping table. Specifically, this includes: adaptively generating a resource compensation multiplier based on the drastic changes in the overall environmental situation, realizing an adaptive adjustment logic of proactively increasing resources in complex scenarios and maintaining baseline resources in stable scenarios. The calculation formula is as follows: ,in This is the resource requirement compensation coefficient corresponding to the sensor node task, and its value is always greater than 0. The situation compensation adjustment coefficient is preset to a fixed value. =0.8. The value of this parameter is based on the dynamic scheduling response characteristics of the vehicle-mounted heterogeneous computing resources. It can effectively balance the over-compensation and under-compensation of resources caused by situational fluctuations. It can ensure that there is sufficient incremental adjustment range of resources in highly complex dynamic scenarios to meet the needs of perception computing power reinforcement, while avoiding the problems of excessive adjustment coefficient leading to resource surge, system load exceeding the limit, or excessive adjustment coefficient leading to compensation effect failure, thus ensuring that the compensation ratio is reasonable and controllable. The regional situation evolution adjustment value is used when the environmental situation is stable and without fluctuations. This value approaches 0, the compensation coefficient approaches 1, and there is no additional increase or decrease in resource quotas. When the environmental situation changes drastically, the compensation coefficient is greater than 1, achieving dynamic resource increment compensation. For all sensor node tasks participating in sensing operations within the situation monitoring area, the initial computational resource quota ratio fixed in the initial resource allocation mapping table is retrieved, and combined with the real-time solved resource demand compensation coefficient, a full-domain dynamic scaling correction is completed. The calculation formula is as follows: ,in Calculate the resource quota ratio for the corrected sensor task; The initial fixed computational resource quota ratio is stored in the initial resource allocation mapping table; The resource requirement compensation coefficient corresponding to the sensor node task is used to achieve positive dynamic adaptation of resource quotas to environmental situation complexity through scaling calculation. Based on the corrected resource quota ratio of all sensor tasks, the binding mapping relationship between the corresponding perception task identifier and the computing resource core identifier is updated synchronously and iteratively, replacing the initial static fixed mapping result. After completing the resource quota correction and mapping update of all tasks in the entire domain, full table data verification, deduplication, normalization, and sorting are carried out to remove abnormal data, correct the ratio deviation, unify the data format, and finally generate the corrected resource allocation mapping table.

[0067] To improve the dynamic adaptability and overall utilization efficiency of onboard heterogeneous computing resources, and enhance the adaptability of intelligent driving systems to dynamic and changing driving environments and their ability to ensure driving safety.

[0068] In a preferred embodiment of the present invention, step 4 above may include:

[0069] In this embodiment of the invention, step 440 involves real-time collection of computing power occupancy data for each computing resource core in the corrected resource allocation mapping table, and calculation of the current load rate of each computing resource core based on the computing power occupancy data. Specifically, this includes: using the corrected and updated resource allocation mapping table as the sole matching basis, locking all computing resource cores participating in the vehicle perception task computation; for each valid computing resource core, real-time collection of raw hardware computing power operation data within a unit scheduling cycle, specifically including four types of core raw data: the current occupied computing power quota of the computing resource core, the maximum rated computing power quota of the computing resource core, the effective computation time within the cycle, and the idle standby time. This ensures a complete record of the real-time computing power occupancy status of each computing resource core, guaranteeing full coverage, no omissions, and no delays in data collection. The load rate is used to characterize the current computing power saturation level of a single computing resource core and is a core quantitative indicator for determining whether resources are overloaded and whether dynamic resource migration is necessary. It is calculated based on the ratio of the currently occupied computing power to the maximum rated computing power, using the following formula: ,in For the first The current load rate of each computing resource core, ranging from 0 to 100%; For the first The real-time occupied computing power of each computing resource core within the current scheduling cycle, in units of computing power units; For the first Each computing resource core has a factory-rated maximum computing power, representing the maximum computing capacity that the resource core can handle. The unit is computing power unit. The load rate is calculated for each computing resource core bound to the perception task. A one-to-one correspondence is established between the unique identifier of the computing resource core and the corresponding load rate value, forming a real-time load rate statistics table for the entire computing resource core, and completely retaining the real-time load status data of each resource core.

[0070] Step 441: Compare the current load rate of the computing resource cores corresponding to high-priority tasks with a preset safety threshold. When the current load rate of the computing resource cores corresponding to high-priority tasks exceeds the preset safety threshold, trigger the resource reallocation process and determine the high-priority tasks that trigger the resource reallocation process. Specifically, this includes: retrieving a pre-generated and fixed task urgency priority sequence, batch filtering out all intelligent driving perception tasks marked as high priority within the sequence, aggregating them into a high-priority task set, further retrieving the currently effective resource allocation mapping table, and relying on the binding correspondence between the unique identifier of the perception task and the unique identifier of the computing resource core stored in the mapping table, sorting the tasks... Each selected high-priority task undergoes line-by-line matching and occupancy status verification. For a single high-priority task, all valid binding entries in the resource allocation mapping table are traversed to retrieve the computing resource core identifier that corresponds one-to-one with the high-priority task identifier. The occupancy status of the computing resource core within the current scheduling cycle is verified to confirm that the current computing resource core is exclusively occupied by a single high-priority task, and there are no anomalies such as multiple tasks sharing or crowding out, no idle unbinding, or cross-task incorrect binding. Through line-by-line matching, verification, and association processing, a pairing relationship is established between each high-priority awareness task and its own exclusively bound computing resource core. Finally, all tasks to be tested for load are compiled. The target resource core set ensures that the pairing relationship is real and effective, the occupancy status is accurate and traceable, and the detection objects are comprehensive and without omission. A unified preset safety threshold of 85% for the core load of computing resources is set. This threshold is the critical safety load standard for the stable operation of vehicle-mounted heterogeneous computing resources. When the core load rate of the resource is below this threshold, it means that the current computing power is sufficient, the resource operation status is stable, and no resource scheduling intervention is required. When the core load rate of the resource is above this threshold, it means that the current computing power is saturated, the computing load is too high, and it is easy to cause problems such as increased processing latency of perception tasks, decreased inference accuracy, and system lag and overload. The current load calculated in step 440 is retrieved for each high-priority task corresponding to the computing resource core. The load rate is compared one by one with a preset safety threshold of 85%. If the current load rate of the computing resource core corresponding to a high-priority task is greater than the preset safety threshold, it is determined that the computing resource core is in an overloaded state and cannot stably support the real-time computing needs of the corresponding high-priority perception task. The dynamic reallocation process of vehicle computing resources is immediately triggered. At the same time, the high-priority task bound to the overloaded resource core is locked and marked as the target high-priority task for triggering the resource reallocation process, which will serve as the core optimization object for subsequent resource migration optimization. If the load rate is less than or equal to the preset safety threshold, it is determined that the resource operation status is normal and no resource reallocation operation is required. The current resource binding relationship remains unchanged.

[0071] Step 442: In the task urgency priority sequence, extract the priority values ​​of the high-priority task that triggers the resource reallocation process and its adjacent low-priority tasks, calculate the priority difference between the high-priority task and its adjacent low-priority task, and obtain the resource migration instruction including the source task identifier and the target task identifier based on the priority difference; specifically, this includes: retrieving the pre-built and solidified task urgency priority sequence, which is arranged in descending order of task urgency, with each sensing task bound to a unique task identifier and a standardized priority quantification value, and using the resource reallocation trigger finally locked in step 441. The high-priority task in the task matching process is used as the positioning base point. A search and matching process is performed in the global priority sequence using the task's unique identifier to determine its position and sequence index within the overall sequence. Based on this base point index, the next-ranking task is searched downwards. Tasks with the next-ranking task, lower urgency level, stable computing power usage, and the ability to release resources are defined as adjacent low-priority tasks. After locating the high- and low-priority paired tasks, the priority values ​​are standardized and extracted. The high-priority task at the base point and the adjacent low-priority tasks are then retrieved and stored in the task urgency priority sequence. The standard quantitative values ​​are standardized parameters with unified evaluation dimensions, calculation rules, and quantification ranges. No secondary conversions or bias corrections are performed throughout the extraction process. Data validity verification is conducted simultaneously during extraction to confirm that the priority values ​​for both task types are complete, valid, non-empty, without abnormal offsets, and without sequence mismatch issues. This strictly ensures that the two sets of priority values ​​originate from the same evaluation system, with unified data dimensions, consistent comparison benchmarks, and genuine and valid hierarchical differences, providing accurate raw data support for subsequent priority difference quantification calculations. The degree of difference in urgency between the two task types is quantified by calculating the difference between the high and low priority task values. The larger the difference, the more significant the difference in urgency between the two types of tasks, and the higher the rationality and necessity of resource migration. Based on the calculated priority difference, the adaptability and execution priority of resource migration are determined. A priority difference greater than zero proves that the urgency of the high-priority task is significantly higher than that of the adjacent low-priority task, satisfying the preconditions for resource migration. The high-priority task occupying overloaded computing resources is marked as the target task for resource migration and bound with a corresponding unique target task identifier. The low-priority tasks that are adjacent in rank, have lower urgency, and can release computing resources are marked as the source tasks for resource migration and bound with a corresponding unique source task identifier. The source task identifier and the target task identifier are integrated to form a complete resource migration instruction that can be directly scheduled and executed.

[0072] Step 443: According to the resource migration instruction, interrupt the execution state of the adjacent low-priority task on the current computing resource core, release the computing resource core occupied by the adjacent low-priority task, and dynamically allocate the released computing resource core to the high-priority task. Update the binding relationship between the task identifier and the computing resource core identifier to obtain the updated computing resource scheduling scheme. Specifically, this includes: parsing the source task identifier contained in the resource migration instruction, locating the adjacent low-priority task that is currently running normally, issuing a run interrupt instruction to the low-priority task, terminating the low-priority task's operation process, data caching, and resource call behavior on the currently bound computing resource core in real time, suspending all computing power occupation operations, ensuring that the task interruption process is smooth, without data conflicts, and without system anomalies. After the low-priority task process is interrupted, completely reclaim all hardware resources such as computing power quota, cache resources, and scheduling resources that the task previously exclusively occupied, clear the temporary binding relationship between the low-priority task and the current computing resource core, and freeing the corresponding computing resource core from the execution state. The overload state is converted to an idle and available state, effectively releasing computing resources. According to the target task identifier in the resource migration instruction, the newly released idle computing core resources are dynamically allocated to high-priority tasks that are currently overloaded. This makes up for the current computing power gap of high-priority tasks, reduces the real-time load rate of the corresponding resource cores of high-priority tasks, eliminates problems such as computing power overload, operation lag, and perception latency, and ensures the continuous, stable, and efficient operation of high-priority perception tasks. The old binding relationship between low-priority tasks and their corresponding computing cores, and between high-priority tasks and overloaded computing cores, is completely abolished. A standardized binding relationship between high-priority tasks and newly allocated idle computing resources is re-established. The binding relationship between global task identifiers and computing core identifiers is iteratively updated. All updated binding relationships are verified, standardized, and solidified. Abnormal binding data is removed, and resource allocation deviations are corrected. Finally, an updated computing resource scheduling scheme that adapts to the real-time environment, computing power load status, and provides hierarchical protection for high and low priority tasks is generated.

[0073] Reduce the latency of high-priority tasks, improve the accuracy of perception and reasoning, and avoid the problems of idle resources and load imbalance.

[0074] like Figure 2 As shown, embodiments of the present invention also provide a computing resource scheduling system based on AI task load, comprising:

[0075] The evaluation module is used to determine the initial urgency of the tasks carried by each sensor node based on the initial AI task load feature vector, and to determine the collision risk coefficient of each sensor node relative to the vehicle by combining the relative positional relationship between the physical spatial coordinates around the vehicle and the current driving trajectory of the vehicle. Based on the initial urgency and the collision risk coefficient, the spatiotemporal hazard weight value is calculated to obtain the task urgency priority sequence.

[0076] The mapping module is used to match the preset heterogeneous computing resource pool within the vehicle domain controller according to the priority sequence of task urgency to obtain the initial resource allocation mapping table;

[0077] The correction module is used to identify the three sensor nodes allocated the highest computing resource quota as core information sources based on the initial resource allocation mapping table; dynamically define the situational awareness area around the vehicle based on the spatial location of the three core information sources; decompose the situational awareness area into several situational sub-blocks; calculate the regional situational evolution adjustment value based on the distribution dispersion and morphological change rate of each situational sub-block between adjacent time windows, wherein the distribution dispersion is used to characterize the degree of dispersion of the spatial coordinate distribution of the target object in the same situational sub-block within adjacent time windows, the morphological change rate is used to characterize the relative change ratio of the three-dimensional envelope volume of the target cluster in the same situational sub-block within adjacent time windows, and the regional situational evolution adjustment value is used to characterize the degree of dynamic change of the overall environmental situation around the vehicle; and correct the resource allocation ratio in the initial resource allocation mapping table with the regional situational evolution adjustment value to obtain the corrected resource allocation mapping table.

[0078] The migration module is used to monitor the current load rate of each computing resource core in real time according to the revised resource allocation mapping table. When the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, a resource migration instruction is obtained based on the priority difference between adjacent tasks in the task urgency priority sequence. According to the resource migration instruction, the computing resource core occupied by the low-priority task is dynamically allocated to the high-priority task to obtain the updated computing resource scheduling scheme.

[0079] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0080] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0081] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A computational resource scheduling method based on AI task load, characterized in that, The method includes: Step 1: Based on the initial AI task load feature vector, determine the initial urgency of the task carried by each sensor node. Combine the relative positional relationship between the physical spatial coordinates around the vehicle and the current driving trajectory of the vehicle to determine the collision risk coefficient of each sensor node relative to the vehicle. Based on the initial urgency and collision risk coefficient, calculate the spatiotemporal hazard weight value to obtain the task urgency priority sequence. Step 2: Based on the task urgency priority sequence, match the preset heterogeneous computing resource pool within the vehicle domain controller to obtain the initial resource allocation mapping table; Step 3: Based on the initial resource allocation mapping table, identify the three sensor nodes allocated the highest computing resource quota as core information sources; dynamically define the situational awareness area around the vehicle based on the spatial location of the three core information sources; decompose the situational awareness area into several situational sub-blocks; obtain the set of spatial coordinates of target objects in each situational sub-block with a unique spatial index in the current time window and the previous adjacent time window; calculate the dispersion of the set of spatial coordinates of target objects in the situational sub-block with the same spatial index in the two time windows to obtain the distribution dispersion of each situational sub-block; extract the contour boundary feature data of each situational sub-block with a unique spatial index in the current time window and the previous adjacent time window, and compare the contour boundary feature data of the situational sub-block with the same spatial index in the two time windows. The relative change ratio of the contour range of situational sub-blocks with the same spatial index identifier within two time windows is calculated to obtain the morphological change rate of each situational sub-block. The distribution dispersion of each situational sub-block is weighted and fused with the morphological change rate of each situational sub-block to evaluate the local situational complexity increment of each situational sub-block. The local situational complexity increments of all situational sub-blocks are aggregated to calculate the regional situational evolution adjustment value, which represents the degree of drastic change in the overall environment. Based on the regional situational evolution adjustment value, the resource demand compensation coefficient of the sensor node task corresponding to each situational sub-block is determined. Using the resource demand compensation coefficient, the computational resource quota ratio of each sensor node task in the corresponding situational concern area in the initial resource allocation mapping table is dynamically scaled and corrected to update the mapping relationship between task identifier and core computational resource identifier, resulting in the corrected resource allocation mapping table. Step 4: Based on the revised resource allocation mapping table, monitor the current load rate of each computing resource core in real time; when the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, obtain a resource migration instruction based on the priority difference between adjacent tasks in the task urgency priority sequence; according to the resource migration instruction, dynamically allocate the computing resource cores occupied by low-priority tasks to high-priority tasks to obtain the updated computing resource scheduling scheme.

2. The computational resource scheduling method based on AI task load according to claim 1, characterized in that, Before step 1, the multi-source raw data streams collected by each sensor node in the sensor array connected to the vehicle edge computing node within a preset time window are obtained, and the multi-source raw data streams are input into the feature extraction network to obtain the initial AI task feature vector of each sensor node; at the same time, the physical space coordinates around the vehicle and the current driving trajectory information of the vehicle are obtained, and the initial AI task feature vector, position coordinates and driving trajectory information are spliced ​​and fused to obtain the initial AI task load feature vector.

3. The computational resource scheduling method based on AI task load according to claim 2, characterized in that, Based on the initial AI task load feature vector, the initial importance of the task carried by each sensor node is determined. Combining the relative positional relationship between the physical spatial coordinates around the vehicle and the current driving trajectory of the vehicle, the collision risk coefficient of each sensor node relative to the vehicle is determined. Based on the initial urgency and collision risk coefficient, the spatiotemporal hazard weight values ​​are calculated to obtain the mission urgency priority sequence, including: The initial AI task load feature vector is analyzed to extract the delay tolerance feature representing the timeliness of task processing and the data volume feature representing the scale of data processing. The delay tolerance feature and the data volume feature are normalized and mapped to obtain the initial importance assessment value of the task carried by each sensor node. Extract the physical space coordinates of the vehicle's surroundings from the feature vector of the initial AI task load, calculate the relative distance and relative azimuth between the physical space coordinates and the vehicle's current driving trajectory, predict the expected intersection time between the target detected by each sensor node and the vehicle's driving path based on the relative distance and relative azimuth, and obtain the collision risk coefficient of each sensor node relative to the vehicle based on the expected intersection time. The initial urgency assessment value and the collision risk coefficient are weighted and fused to obtain the spatiotemporal hazard weight value that characterizes the spatiotemporal urgency of the mission. Based on the spatiotemporal hazard weight values, the tasks carried by each sensor node are sorted in descending order to obtain a task urgency priority sequence that includes the correspondence between task identifiers and spatiotemporal hazard weight values.

4. The computational resource scheduling method based on AI task load according to claim 3, characterized in that, Based on the task urgency priority sequence, a preliminary resource allocation mapping table is obtained by matching the preset heterogeneous computing resource pool within the vehicle domain controller, including: The task urgency priority sequence is analyzed, the identification information of each task in the sequence and the corresponding spatiotemporal hazard weight value are extracted, and the computing resource requirements of each task are mapped according to the spatiotemporal hazard weight value. Traverse the preset heterogeneous computing resource pools within the vehicle domain controller and collect the operating status parameters of each computing resource core in real time, including the current available computing power capacity, the current remaining memory, and the resource core type identifier; By matching and comparing the computing resource requirements with the operating status parameters, computing resource cores that are currently available computing power capacity greater than the computing power requirement threshold and currently have a memory balance greater than the memory usage threshold are selected, and a set of candidate computing resource cores that meet the computing power requirements is constructed. Based on the order of task identifiers in the task urgency priority sequence, each task is sequentially assigned to the final computing resource core in the candidate computing resource core set, resulting in an initial resource allocation mapping table that includes the correspondence between task identifiers and computing resource core identifiers.

5. The computational resource scheduling method based on AI task load according to claim 4, characterized in that, The computing resource requirements specifications include computing power requirement thresholds and memory usage thresholds.

6. The computational resource scheduling method based on AI task load according to claim 4, characterized in that, Based on the initial resource allocation mapping table, the three sensor nodes allocated the highest computing resource quotas are identified as the core information sources; The spatial location of the vehicle's surroundings is dynamically defined based on the spatial position of three core information sources; The area of ​​interest in the situation is decomposed into several sub-blocks, including: Extract the sensor node identifiers and associated computing resource quota values ​​from the initial resource allocation mapping table, sort the computing resource quota values ​​in descending order, extract the top three sensor node identifiers and mark them as core information sources, thus obtaining the core information source set. The system acquires the three-dimensional position coordinate data of each core source in the core source set in the physical space around the vehicle. Taking the current position of the vehicle as the reference origin, it calculates the spatial vector relationship between the three-dimensional position coordinate data of the three core sources and the reference origin. The three-dimensional position coordinates of the three core sources are used as spatial boundary constraints to dynamically define and generate the situational concern area around the vehicle, including the core sources, and obtains the spatial boundary parameters of the situational concern area. The spatial boundary parameters of the area of ​​interest are extracted, and the geometric range parameters of the area of ​​interest are obtained by parsing. Combined with the preset grid resolution threshold, the number of grids in the three-dimensional space is calculated for the geometric range parameters. Based on the number of grids, the area of ​​interest is decomposed into uniform blocks to obtain several spatially non-overlapping sub-blocks with unique spatial index identifiers.

7. The computational resource scheduling method based on AI task load according to claim 6, characterized in that, Based on the revised resource allocation mapping table, the current load rate of each computing resource core is monitored in real time; when the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, a resource migration instruction is obtained based on the priority difference between adjacent tasks in the task urgency priority sequence. Based on the resource migration instructions, the computing resource cores occupied by low-priority tasks are dynamically allocated to high-priority tasks, resulting in an updated computing resource scheduling scheme, including: The computing power occupancy data of each computing resource core in the corrected resource allocation mapping table is collected in real time, and the current load rate of each computing resource core is calculated based on the computing power occupancy data. The current load rate of the computing resource cores corresponding to high-priority tasks is compared with a preset security threshold. When the current load rate of the computing resource cores corresponding to high-priority tasks is greater than the preset security threshold, the resource reallocation process is triggered, and the high-priority task that triggers the resource reallocation process is determined. In the task urgency priority sequence, extract the priority values ​​of the high-priority task that triggers the resource reallocation process and the adjacent low-priority task, calculate the priority difference between the high-priority task and the adjacent low-priority task, and obtain the resource migration instruction including the source task identifier and the target task identifier based on the priority difference. According to the resource migration instruction, the execution state of adjacent low-priority tasks in the current computing resource core is interrupted, the computing resource core occupied by the adjacent low-priority tasks is released, and the released computing resource core is dynamically allocated to high-priority tasks. The binding relationship between task identifier and computing resource core identifier is updated to obtain the updated computing resource scheduling scheme.

8. A computing resource scheduling system based on AI task load, the system implementing the method as described in any one of claims 1 to 7, characterized in that, include: The evaluation module is used to determine the initial importance of the tasks carried by each sensor node based on the initial AI task load feature vector, and to determine the collision risk coefficient of each sensor node relative to the vehicle by combining the relative positional relationship between the physical space coordinates around the vehicle and the current driving trajectory of the vehicle. Based on the initial urgency and collision risk coefficient, the spatiotemporal hazard weight value is calculated to obtain the mission urgency priority sequence; The mapping module is used to match the preset heterogeneous computing resource pool within the vehicle domain controller according to the priority sequence of task urgency to obtain the initial resource allocation mapping table; The correction module is used to identify the three sensor nodes allocated the highest computing resource quota as core information sources based on the initial resource allocation mapping table; The situational awareness area around the vehicle is dynamically defined based on the spatial location of three core information sources. This area is then decomposed into several sub-blocks. Based on the distribution dispersion and morphological change rate of each sub-block within adjacent time windows, a regional situational evolution adjustment value is calculated. The distribution dispersion characterizes the degree of dispersion of the spatial coordinate distribution of target objects within the same sub-block in adjacent time windows, the morphological change rate characterizes the relative change in the three-dimensional envelope volume of the target cluster within the same sub-block in adjacent time windows, and the regional situational evolution adjustment value characterizes the degree of dynamic change in the overall environmental situation around the vehicle. The resource allocation ratio in the initial resource allocation mapping table is then corrected using the regional situational evolution adjustment value to obtain a corrected resource allocation mapping table. The migration module is used to monitor the current load rate of each computing resource core in real time according to the revised resource allocation mapping table. When the current load rate of the computing resource core corresponding to a high-priority task exceeds the preset safety threshold, a resource migration instruction is obtained based on the priority difference between adjacent tasks in the task urgency priority sequence. According to the resource migration instruction, the computing resource core occupied by the low-priority task is dynamically allocated to the high-priority task to obtain the updated computing resource scheduling scheme.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images