A multi-vehicle cooperative perception method based on risk prediction

Abstract

The application provides a multi-vehicle cooperative perception method based on risk prediction, comprising: acquiring dynamic information reported by each vehicle in a scheduling area; constructing a space-time graph based on the dynamic information; inputting the space-time graph into a space-time graph attention network model to predict collision risks between each target in the scheduling area and generate a risk matrix, wherein elements in the risk matrix represent quantitative collision risk values of specified target pairs within a prediction time window; calculating a global priority for the perception data in the scheduling area according to the risk matrix, wherein the calculation of the global priority is positively correlated with the risk values in the risk matrix; and dynamically allocating a communication time slot to a selected vehicle in a time division multiplexing manner based on the global priority to broadcast its perception data. The method improves the safety and real-time performance of multi-vehicle cooperative perception, optimizes the efficiency of communication resource allocation, and effectively adapts to complex and dynamic application scenarios of Internet of Vehicles.

Description

Technical Field

[0001] This invention relates to the field of vehicle coordination and perception technology, and specifically provides a multi-vehicle cooperative perception method based on risk prediction. Background Technology

[0002] Cooperative perception technology effectively compensates for the limited range and susceptibility to obstruction of single-vehicle intelligent perception by sharing information among multiple vehicles, significantly improving driving safety. However, frequent information interaction between vehicles consumes a large amount of bandwidth resources, leading to increased channel conflicts and severely affecting the timeliness of information. Therefore, determining when and where to share which information, while ensuring transmission efficiency and improving road safety, is a key challenge facing vehicle-to-everything (V2X) cooperative perception. Summary of the Invention

[0003] To address the aforementioned problems, this invention provides a multi-vehicle cooperative perception method based on risk prediction, comprising: The system acquires dynamic information reported by each vehicle within the scheduling area, and the dynamic information includes at least structured motion parameters obtained by the vehicle end from motion estimation of the perceived target. Based on the dynamic information, a spatiotemporal graph is constructed. The spatiotemporal graph is then input into a spatiotemporal graph attention network model to predict the collision risk between targets within the scheduling area and generate a risk matrix. The elements in the risk matrix represent the quantified collision risk value of a specified target pair within the prediction time window. Based on the risk matrix, a global priority is calculated for the sensing data within the scheduling area, wherein the calculation of the global priority is positively correlated with the risk value in the risk matrix; Based on the global priority, communication time slots are dynamically allocated to selected vehicles in a time-division multiplexing manner to broadcast their sensing data.

[0004] Preferably, the structured motion parameters obtained by the vehicle-side motion estimation of the perceived target include: Acquire continuous frame point cloud data perceived by this vehicle; Target detection is performed on the continuous frame point cloud data to obtain the detection results for each frame; Based on the detection results between two frames, target matching is performed by calculating the affinity matrix to associate the detection boxes of the same target in different frames; For each successfully matched target pair, the transformation matrix of its point cloud features between the two frames is calculated using the Iterative Closest Point (ICP) algorithm. The transformation matrix is ​​used as an observation and input into an unscented Kalman filter (UKF) to estimate the state vector of the perceived target.

[0005] Preferably, the affinity matrix is ​​constructed based on one or more of the following affinities: semantic affinity, size affinity, orientation affinity, relative velocity affinity, and position affinity.

[0006] Preferably, the construction of the spatiotemporal graph based on dynamic information includes: Dynamic targets within the scheduling area are used as graph nodes; Node features are constructed, which integrate the kinematic features of the target, the degree centrality features of the nodes, and the road space features extracted with the target as the center; Construct edge features, which are collision risk values ​​calculated based on the relative motion parameters and interaction intensity between two connected targets.

[0007] Preferably, the road spatial features include: road dynamic features and vehicle group perception features; The node feature construction includes: generating a local occupancy density map and velocity field centered on the target vehicle based on the position and speed information of vehicles within the scheduling area; Features are extracted from the local occupancy density map and velocity field using a convolutional neural network (CNN), and the extracted features are fused to obtain the road dynamic features; Dynamically filter the neighboring vehicle groups based on the target vehicle's speed, direction of movement, and lane information; Calculate the statistical characteristics of the adjacent vehicle group, the statistical characteristics including at least one of the group density, the average speed of the group vehicles, the average acceleration, and the average steering angle; Based on the relative motion relationship between the target vehicle and its neighboring vehicles, the characteristics reflecting the consistency of traffic flow direction and the group interaction intensity characteristics reflecting the degree of interaction are calculated. By integrating the statistical features, directional consistency features, and group interaction intensity features, the vehicle group perception features are obtained.

[0008] Preferably, the spatiotemporal graph attention network model includes a spatial attention layer and a temporal attention layer; The spatial attention layer is used to aggregate information about neighboring nodes based on node features and edge features; The time attention layer is used to aggregate the feature information of nodes at different time steps.

[0009] Preferably, the spatiotemporal graph is input into a spatiotemporal graph attention network model to predict the collision risk between targets within the scheduling area, generating a risk matrix including: The spatiotemporal graph is processed by the spatial attention layer and the temporal attention layer to obtain updated node features containing temporal context information; The updated node features are decoded into predicted trajectories of each target within the scheduling area; Based on the predicted trajectory, calculate the collision time between any two targets; Based on the collision time, a quantified collision risk value between the two targets is determined to generate the risk matrix.

[0010] Preferably, calculating the global priority for the sensing data within the scheduling area based on the risk matrix includes: For each sensed object, calculate the global priority of the sensed object. : ; in For the set of vehicles of the demand-aware object o, The collision risk values ​​for vehicles at o and v. To send the communication coverage of vehicle s, and These are the weighting coefficients.

[0011] Preferably, based on the global priority, communication time slots are dynamically allocated to selected vehicles in a time-division multiplexing manner to broadcast their sensing data, including at least: For each perceived object o, the set of vehicles that can perceive it... In the middle, select to make the utility value The largest vehicle, acting as the sender, selects data packets from its set of sensed objects. To maximize the overall priority and obtain the optimal data packet. ; All sending tasks are prioritized by group. The data are sorted in descending order, and a greedy algorithm is used to allocate conflict-free communication time slots. Each time slot allows only one vehicle to send data, and the time slot length is dynamically adjusted according to the amount of data to be sent. ,in Channel rate; After the vehicle sends a data packet, a demand update instruction is broadcast to vehicles within the coverage area. If the perceived object o is the demand vehicle of vehicle v, o is removed from the local demand list. The central node updates its local priority based on the updated requirement list.

[0012] Preferably, based on the global priority, dynamically allocating communication time slots to selected vehicles in a time-division multiplexing manner to broadcast their sensing data further includes: If a new vehicle is detected, a sudden change in collision risk, or a change in network topology, the global priority calculation must be re-executed before entering the next scheduling cycle.

[0013] The beneficial effects of this invention are as follows: 1. Effectively solves the core contradiction of "information transmission redundancy" and "insufficient timeliness of high-risk information" in existing multi-vehicle collaborative perception: By positively correlating the global priority of perception data with the quantified collision risk value between targets, the key perception data in high collision risk scenarios is prioritized for transmission, avoiding unnecessary data from occupying bandwidth resources. This not only ensures that vehicles can obtain dangerous information in emergency driving scenarios in a timely manner, significantly improving driving safety, but also greatly reduces transmission redundancy and communication load.

[0014] 2. Significantly improved accuracy and comprehensiveness of collision risk prediction: Based on the structured motion parameters in dynamic information, a spatiotemporal graph is constructed. Combined with the characteristics of the spatiotemporal graph attention network model, it can capture the interaction and correlation of perceived targets in the spatial dimension (such as the relative position and motion direction coordination between vehicles) and explore the trajectory evolution pattern in the temporal dimension (such as the temporal change of the target's motion state). This enables a quantitative assessment of the collision risk of targets within the prediction time window, overcoming the shortcomings of existing risk assessment methods that are difficult to take into account the spatiotemporal dimensions and have incomplete characterization of interaction relationships. The generated risk matrix provides a reliable basis for priority allocation.

[0015] 3. Significantly optimized communication resource utilization and transmission stability: The time-division multiplexing method is used to dynamically allocate communication time slots, and vehicles are selected for data broadcasting by global priority screening. On the one hand, this avoids channel conflicts caused by multiple vehicles transmitting at the same time, ensuring the integrity and stability of data transmission. On the other hand, the time slot allocation is deeply adapted to the priority of sensing data and the vehicle communication coverage, realizing dynamic optimization of "who should transmit and when to transmit", improving the utilization efficiency of limited bandwidth resources and adapting to complex vehicle networking scenarios such as dense traffic and dynamic changes in vehicle topology. Attached Figure Description

[0016] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a schematic diagram of a multi-vehicle cooperative perception method based on risk prediction according to an embodiment of the present invention; Figure 2 This is a dynamic information extraction process according to an embodiment of the present invention; Figure 3 This is a dynamic risk assessment process according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating a multi-vehicle cooperative perception method based on risk prediction according to an embodiment of the present invention. Detailed Implementation

[0017] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0018] like Figure 1-4 As shown, the present invention provides a multi-vehicle cooperative perception method based on risk prediction, comprising: Step S1: Obtain dynamic information reported by each vehicle within the scheduling area. The dynamic information includes at least structured motion parameters obtained by the vehicle end for motion estimation of the perceived target.

[0019] In this embodiment, step S11 first involves dividing and preprocessing the original point cloud into regions, and then extracting dynamic point clusters through ground point removal and background filtering.

[0020] Step S12: Then, based on PointPillars, generate 3D boundary detection results and locate the geometric and semantic features of the target object on the feature map.

[0021] Step S13: For point cloud feature data of two consecutive frames, use affinity matrix to achieve cross-frame object matching and construct motion trajectory segments of dynamic targets; Step S14: Calculate the initial values ​​of inter-frame pose changes using the ICP point cloud registration algorithm, and combine this with unscented Kalman filtering (UKF) to perform optimal motion state estimation, obtaining the motion estimation parameters of the dynamic object. This model achieves end-to-end parsing from raw sensing data to structured motion parameters.

[0022] Specifically, the Randomization and Rectification Algorithm (RANSAC) and High Definition Map (HD) are used to remove ground and background points from the raw point cloud data, retaining only the detected dynamic objects.

[0023] A 3D object detector is used to detect objects on the remaining points, identifying and extracting point cloud regions of all dynamic objects on the road. The results are represented by labeled 3D bounding boxes. The point cloud features of the detected objects are obtained by locating their feature maps using these 3D bounding boxes. The 3D object detector is based on PointPillars, and its output includes each detected 3D bounding box and associated features. Specifically, object detection is performed on two consecutive frames of point cloud data. The output of the 3D detector for each frame is a detection list D, where each object in the detection list is represented as... .

[0024] To obtain point cloud feature clustering for target objects, corresponding features are extracted from the feature map obtained by the PointPillars algorithm in object detection. For each detected object, the 3D coordinates of its bounding box are mapped to the coordinate system of the feature map, and the feature vectors of the mapped region are extracted to obtain a set of features. .

[0025] To match target objects in two consecutive frames for motion estimation, an affinity matrix is ​​introduced into object matching, using a multi-attribute-based affinity matrix to associate the same objects in two consecutive frames.

[0026] Affinity attributes include the following: semantic affinity, size affinity, orientation affinity, relative velocity affinity, and position affinity. The predicted object velocity vector v is added to the detected object (D), and the detected object state is updated. Semantic affinity is defined based on the object's category as follows: ,in and It is the semantic label of the detected object.

[0027] Affinity is set to 0 for objects of different categories, and 1 for objects of the same category.

[0028] For each target object, assume the object The size is Size affinity is defined as .

[0029] In order to utilize the motion characteristics of objects, Defined as directional affinity, it assesses the similarity between objects based on their rotation angle.

[0030] Utilizing relative velocity affinity To measure the relative motion between objects.

[0031] Similarity is assessed based on the spatial location of objects, using Measure location affinity.

[0032] coefficient , , , The sensitivity adjustment parameters were set to 0.5, 0.3, 0.4, and 0.5. The overall affinity is expressed as follows: , , , Weighting coefficients: ; Construct an affinity matrix for the two objects based on their overall affinity. Using affinity matrix Matching is performed, and the motion of the object is estimated based on the bounding boxes and point cloud features in two consecutive frames, respectively using... , and , express, Represents the object ID. Represents a timestamp.

[0033] Traditional motion estimation methods based on motion models often assume that the target's motion is stationary and predictable, ignoring uncertainties and external disturbances during the object's motion, making it difficult to reflect the actual trajectory of vehicles and objects. Another motion estimation algorithm for point cloud data, Iterative Closest Point (ICP) registration, registers point cloud data to estimate the target's displacement between different time frames. ICP registration performs well in handling objects in static environments, but its performance is limited when the target is dynamically changing. To overcome the limitations of traditional motion models and ICP registration methods, this method combines the ICP registration algorithm with physical model-based target tracking to effectively estimate the trajectory of road objects in dynamic environments. By accurately aligning the object's point cloud feature data through ICP and combining it with physical models (such as Kalman filtering or particle filtering) for motion prediction, tracking accuracy can be improved while reducing the computational burden of ICP and enhancing real-time performance. Specifically: For each matched object pair, the transformation matrix of the object's point cloud features is calculated using the ICP (Iterative Closest Point) registration algorithm. The ICP algorithm iteratively optimizes to minimize the matching error between the source and target point clouds, thus obtaining the object's precise displacement (translation and rotation) between two frames. The source and target point cloud features are respectively... and The ICP algorithm calculates the transformation matrix by minimizing the objective function. : ; The transformation matrix T = [R|t], where R is the rotation matrix and t is the translation vector. Next, this transformation matrix is ​​applied to the changes over a unit of time, performing time-difference scaling.

[0034] ; The calculated unit transformation matrix can accurately describe the motion of a target object in two consecutive frames. As input for subsequent motion prediction, it is used as the observation value input into the unscented Kalman filter, and the target's state is corrected through the UKF update step.

[0035] Define the state vector in a UKF (United Kingdom Functions) system. In a UKF, the state vector typically includes the vehicle's position. and speed Information such as the vehicle's state vector is: ; Next, we define a state transition model to describe how the vehicle transitions from one state to another. Using the CTRV model, the state transition equation is: ; in It is the time step. It is process noise.

[0036] The UKF observation model is defined, which maps the vehicle state to the observation space. The transformation matrix provided by ICP contains translation and rotation information, which is used as the observation input for the UKF. The observation model is as follows: ; in, , and It is a translation vector. , and It is an element of the rotation matrix R.

[0037] The core of UKF is to update state predictions using observations. In this process, the state at the current time is first... Covariance Generate sigma points, propagate these sigma points to the next time step through the state transition function to obtain the predicted sigma points, and calculate the predicted state mean. Covariance Matrix : ; in and These are the weighting coefficients, and Q is the process noise covariance matrix.

[0038] During the observation update process, the propagated sigma points are first mapped to the observation space to obtain the corresponding observation values. Next, the predicted observations and covariance matrix are calculated: ; Where R is the observation noise covariance matrix, representing the ICP measurement error. Next, the covariance matrix between the state and the observation is calculated. And use this matrix to calculate the Kalman gain. : ; Kalman gain is used to correct the state prediction and updates the vehicle's state using the following formula: ; Finally, update the state covariance matrix: .

[0039] Step S2: Construct a spatiotemporal graph based on the dynamic information, input the spatiotemporal graph into the spatiotemporal graph attention network model to predict the collision risk between each target in the scheduling area, and generate a risk matrix. The elements in the risk matrix represent the quantified collision risk value of a specified target pair within the prediction time window.

[0040] In this embodiment, during the trajectory prediction stage, a spatiotemporal heterogeneous graph is first constructed based on vehicle motion state, node degree centrality, and road dynamic features. A spatial attention mechanism is used to analyze the interaction patterns between vehicles and road structural constraints, and a temporal attention layer captures the temporal evolution of traffic conditions. Finally, a GRU decoder generates the future trajectory sequence of vehicles. In the risk assessment stage, the spatiotemporal distribution of the predicted trajectories is used as input. Collision risks between vehicle pairs are quantified through probabilistic modeling, and collision probabilities are dynamically assessed based on trajectory uncertainty analysis. Both stages achieve feature sharing and decision optimization through end-to-end joint training, enhancing trajectory prediction accuracy and risk assessment reliability. Specifically: Step S21, Graph Construction: For the dynamic road objects detected by the infrastructure, the object set constructed from the detection results is represented as a graph. : ; in Let represent the set of road objects in frame t. Representation diagram The link relationships within.

[0041] To avoid feature blurring and redundant computation caused by too many connections between nodes, Delaunay triangles are used to connect the nodes. Considering the connectivity of the road scene, road semantic features are used when connecting nodes to filter out edges that inappropriately connect vehicles.

[0042] Step S22, Node Encoding: Encode the kinematic information of road objects, such as their position, direction of movement, and speed. Furthermore, the degree centrality of a vehicle in the graph often reflects its importance within the entire road network; degree centrality can effectively represent a vehicle's influence and interaction frequency within the traffic network.

[0043] First, LSTM is used to capture the vehicle's kinematic features, and then a multilayer perceptron (MLP) is used to fuse the vehicle's kinematic features (such as position, velocity, and velocity angle) with the node degree centrality features. The representation process is as follows: ; in , , These represent the forget gate, input gate, and output gate, respectively. It represents the current state of the candidate cells. It is the current state of the cell. This represents the hidden state at the current time step. W and b are the weight matrix and bias term, respectively. Input Next, MLP is used to fuse node degree centrality features with kinematic features, which is defined as: ; It is the extracted vehicle kinematics time-series features. This is a node degree centrality feature. This fusion method not only effectively represents the dynamic behavior of vehicles but also reveals their relative importance within the entire transportation network, enabling the model to better understand the complex interactions between vehicles.

[0044] Step S23, Road Spatial Coding: This mainly includes the dynamic characteristics of the road and the perception characteristics of the vehicle group.

[0045] Road dynamics reflect changes in the road itself and how vehicles move on it. Local occupancy density maps and velocity fields can be used to describe road dynamics. These maps provide information on vehicle distribution and movement trends on the road and can capture inertial behaviors of vehicles on specific roads (such as acceleration, deceleration, lane changing, etc.).

[0046] Local occupancy density maps represent the distribution of vehicles within a specific area of ​​a road, reflecting the degree of traffic congestion. A two-dimensional grid is used to represent the vehicle density in different areas, dividing the road area into... The grid, each grid cell Indicates position The grid region. For each grid cell, the occupancy density of the region is calculated using the following formula. : ; in It is a grid cell area, As an indicator function, when the vehicle Location Falling on the grid When the coverage is within a certain range, the function value is 1; otherwise, it is 0. The occupancy density of the area can be obtained by weighted averaging of the coverage of all vehicles in the area.

[0047] The velocity field displays the vehicle speed distribution over a road area. Similar to the occupancy density map, it divides the road area into a two-dimensional grid, reflecting the flow of traffic on the road through the velocity values ​​assigned to each grid cell. In each grid cell... Within this area, calculate the average speed of all vehicles. : ; in It is a vehicle speed, It is a grid cell area, For indicator functions, representing vehicles Whether it is within this grid cell.

[0048] A convolutional neural network (CNN) is used as the feature extractor to extract spatial features from the occupancy density map and velocity field, including vehicle distribution and speed changes on the road. Finally, the features from the local occupancy density map and velocity field are combined to enable the model to better understand the inertial behavior of vehicles on the road.

[0049] ; in and These are the vehicle-centric occupancy density map and the velocity field, respectively.

[0050] Vehicle group perception features, by capturing the interaction between the target vehicle and surrounding vehicles, also reflect, to some extent, the driving behavior patterns of vehicles on the road. First, a group of neighboring vehicles related to the target vehicle is selected. A comprehensive selection strategy is designed, using the target vehicle as the center and setting a dynamic radius related to vehicle speed. Vehicles in specific directions are selected based on their movement direction, and vehicles in the target vehicle's lane and adjacent lanes are selected based on lane information. Finally, vehicles with potential collision risks to the target vehicle are prioritized using collision events.

[0051] Within the selected vehicle group, based on the current state of the target vehicles and their historical data, the group perception features of the vehicle group, such as group density, are extracted. Group movement characteristics (average speed of group vehicles) Average acceleration and average steering angle Consistency of traffic flow direction and the intensity of group interaction The consistency of traffic flow direction and the intensity of group interaction are represented as follows: ; in and These represent the travel direction angles of the target vehicle and other vehicles, respectively. , and These represent the relative speed, relative acceleration, and relative distance between the target vehicle and neighboring vehicles, respectively. , , These are weighting coefficients that control the relative influence of relative velocity, relative acceleration, and relative distance on the interaction intensity. Finally, a multilayer perceptron (MLP) is used to further fuse these features: ; Finally, the road dynamics and vehicle group characteristics are combined to obtain the road spatial characteristics: ; By integrating the above coding, the spatial state of the road itself, traffic flow dynamics, and the behavioral characteristics of vehicle groups are incorporated into the same spatial code, providing more comprehensive information for subsequent collision risk prediction.

[0052] Step S24, Edge Encoding: In graph neural network models, edge encoding is crucial for capturing the relationships between vehicles. To more accurately assess collision risks on the road, collision risk is used as an edge feature.

[0053] For collision risk, relative motion, distance, acceleration, and steering angle between vehicles, as well as driving behavior, are key factors in collision risk indicators. By encoding these factors, the model's focus on different nodes can be dynamically adjusted during information propagation, thereby enhancing the model's ability to identify potential collisions. Based on the above analysis, the collision risk between adjacent nodes is defined as: ; in It is the relative distance between vehicles. It is the relative speed between vehicles. It is the collision time (TTC). It is a nonlinear function of the interaction strength between vehicles. (Inter-vehicle interaction strength) The definition is as follows: ; and These represent the behavior patterns of the target vehicle and its neighboring vehicles, respectively, and can be classified using clustering algorithms. Indicates the similarity between two behavioral patterns. Represents the behavioral differences between two objects. , , It is the weighting coefficient.

[0054] The method described above, which uses collision risk as an edge encoding, can effectively capture the dynamic interaction relationships between vehicles and improve the graph neural network's ability to identify potential collision risks. This encoding method can enhance the model's sensitivity to high-risk areas, guide the network to focus on key interactions during the decision-making process, and optimize collision prediction and early warning performance.

[0055] Step S25: After obtaining the extracted features, the road spatial code is added to the node attributes to construct a spatiotemporal graph. The spatiotemporal graph is represented as follows: Each of the graphs Includes node features Sum of edge features .

[0056] Design a two-layer graph attention network, including a spatial attention layer and a temporal attention layer, to capture important information in both spatial and temporal dimensions.

[0057] For the graph at each time step The spatial attention layer updates node and edge information from a topological perspective. The updated node features are used as input to the temporal attention layer, which updates node information according to the relative importance of each time period and outputs the final time-aware node features.

[0058] The spatial attention layer first receives the original spatiotemporal graph input, where each node and edge contains the vehicle's state features and its collision risk information. In the spatial attention layer, the attention coefficient consists of two parts: one based on the relative spatial information between nodes, and the other related to edge features (collision risk). Specifically, the formula for calculating the spatial attention coefficient is: ; in For the set of neighboring nodes, and and These are the learned parameters. The topological properties of nodes can be captured by calculating the spatial attention coefficients. Then, the node properties are updated: ; in For activation function, For learnable parameters, the above update process is applied at each time step. This paper also employs a multi-head attention mechanism to calculate spatial attention coefficients, averaging the outputs of multiple attention heads to obtain the final updated features of the nodes. Edge attribute updates are as follows: ; The temporal attention layer is used to capture importance in the temporal dimension. Its input is the output of the spatial attention layer, which is the updated set of node features. The time attention coefficient is: ; in and These are learnable parameters. Update node attributes: ; Step S3: Calculate the global priority for the sensing data within the scheduling area based on the risk matrix, wherein the calculation of the global priority is positively correlated with the risk value in the risk matrix.

[0059] In this embodiment, the node feature representation is decoded into the future trajectory of the vehicle based on the GRU decoder. The intersection of the future trajectories of the vehicles is determined. If two vehicles within the prediction range have an intersection point, the time of the intersection point is taken as the collision time. If there is no obvious intersection point, the relative distance between the vehicles at the latest predicted time is calculated, and the collision time (TTC) is estimated. The collision time can be estimated using the following formula: ; in and These are the relative distance and relative speed of the vehicles, respectively. This refers to the safe distance between vehicles, determined based on factors such as vehicle size and speed. Using the collision time mentioned above, risk indicators can be defined to quantify potential collision risk. ; in The threshold value represents the safe critical time for a collision between vehicles.

[0060] in and These are the relative distance and relative speed of the vehicles, respectively. This refers to the safe distance between vehicles, determined based on factors such as vehicle size and speed. Using the collision time mentioned above, risk indicators can be defined to quantify potential collision risk. ; in The threshold value represents the safe critical time for a collision between vehicles.

[0061] Step S4: Based on the global priority, dynamically allocate communication time slots to the selected vehicles in a time-division multiplexing manner to broadcast their perception data.

[0062] In this embodiment, network bandwidth resources are limited, vehicles have timeliness requirements for perception results, and the number and size of perception results are uncertain. Transmitting all relevant perception results within the decision interval is not feasible; therefore, it is necessary to determine which objects each agent should transmit within the decision interval. Furthermore, to avoid conflicts and bandwidth waste caused by multiple vehicles broadcasting data simultaneously, this paper employs time-slot scheduling. The scheduler needs to determine which vehicle should transmit which data content in each time slot. For example, objects with a high collision risk require high-priority transmission. Vehicles need to determine the perception result objects they need to broadcast and their order based on the data scheduler, and then broadcast their data content.

[0063] A priority-based data scheduling method was designed to reduce redundancy, lower latency, and improve network resource utilization while ensuring vehicle perception quality and area coverage.

[0064] Based on the preliminary global risk assessment, the vehicle perception data is further processed to obtain perception results objects that are highly correlated with collision risk. ,in They represent The first in the vehicle perception results Bounding boxes and motion prediction information for each object , For object With object The relative collision risk, where c is the number of global objects.

[0065] Within the dispatch domain, vehicles share collision risk data, meaning all vehicles share it. The vehicle, based on local perception results and It is easy to obtain a list of local requirements. .

[0066] The following section details the priority-based multi-vehicle perception data broadcast scheduling design, which consists of three parts: global priority calculation, sender selection, and time slot allocation. The first step is the global priority calculation of the perception objects. Each vehicle maintains its local demand list in a timely manner. The sensed objects are sorted in descending order of collision risk, and the central node calculates the global priority. ; in For the set of vehicles of the demand-aware object o, The collision risk values ​​for vehicles at o and v. To send the communication coverage of vehicle s, and This represents the weighting coefficient. Combining collision risk with demand coverage ensures that high-risk and high-demand objects are prioritized for broadcasting, improving perceived safety and network resource utilization.

[0067] When selecting vehicles for transmission, it is necessary to consider maximizing the benefit of a single transmission. For each perceived object o, filter the set of vehicles. , The vehicle sensing objects in the context include o, and at least one vehicle within its communication range is in need of a vehicle. .calculate Utility value of each vehicle: ; in For vehicles and Euclidean distance, This is the distance attenuation factor. The sending vehicle with the highest utility value is selected. Select data packets from its set of perceived objects. Maximize the overall priority: ; For the perceived object Given the amount of data, solving the above formula yields the optimal data packet. .

[0068] To avoid channel conflicts caused by multiple vehicles broadcasting simultaneously, transmission tasks are prioritized by combination. Arrange the time slots in descending order and use a greedy algorithm to allocate them.

[0069] Only one vehicle is allowed to send data in each time slot, and the time slot length is dynamically adjusted according to the amount of data to be sent. ,in This refers to the channel rate.

[0070] After a vehicle sends a data packet, the vehicle's demand status and global priority weight are updated in a timely manner.

[0071] When vehicle s sends a request to object o, it broadcasts a demand update command to vehicles within its coverage area. If object o is a vehicle requesting a request from vehicle v, o is removed from its local demand list. The central node updates its local priority based on the updated demand list. ; This is the collection of vehicles whose needs are met as a result of this broadcast.

[0072] If a new vehicle is detected, a sudden change in collision risk, or a change in network topology, the global priority calculation must be re-executed to enter the next scheduling cycle.

[0073] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multi-vehicle cooperative perception method based on risk prediction, characterized in that, include: The system acquires dynamic information reported by each vehicle within the scheduling area, wherein the dynamic information includes at least structured motion parameters obtained by the vehicle end for motion estimation of the perceived target. Based on the dynamic information, a spatiotemporal graph is constructed. This graph is then input into a spatiotemporal graph attention network model to predict the collision risk between targets within the scheduling area, generating a risk matrix. Elements in the risk matrix represent the quantified collision risk value of a specified target pair within the prediction time window. The construction of the spatiotemporal graph based on dynamic information includes: using dynamic targets within the scheduling area as graph nodes; constructing node features, which integrate the kinematic features of the target, node degree centrality features, and road space features extracted centered on the target; and constructing edge features, which are collision risk values ​​calculated based on the relative motion parameters and interaction intensity between two connected targets. Based on the risk matrix, a global priority is calculated for the sensing data within the scheduling area, wherein the calculation of the global priority is positively correlated with the risk value in the risk matrix; Based on the global priority, communication time slots are dynamically allocated to selected vehicles in a time-division multiplexing manner to broadcast their perception data.

2. The method according to claim 1, characterized in that, The structured motion parameters obtained by estimating the motion of the perceived target from the vehicle end include: Acquire continuous frame point cloud data perceived by this vehicle; Target detection is performed on the continuous frame point cloud data to obtain the detection results for each frame; Based on the detection results between two frames, target matching is performed by calculating the affinity matrix to associate the detection boxes of the same target in different frames; For each successfully matched target pair, the transformation matrix of its point cloud features between the two frames is calculated using the Iterative Closest Point (ICP) algorithm. The transformation matrix is ​​used as an observation and input into an unscented Kalman filter (UKF) to estimate the state vector of the perceived target.

3. The method according to claim 2, characterized in that, The affinity matrix is ​​constructed based on one or more of the following affinities: semantic affinity, size affinity, orientation affinity, relative velocity affinity, and position affinity.

4. The method according to claim 1, characterized in that, The road spatial features include: road dynamic features and vehicle group perception features; The node feature construction includes: generating a local occupancy density map and velocity field centered on the target vehicle based on the position and speed information of vehicles within the scheduling area; Features are extracted from the local occupancy density map and velocity field using a convolutional neural network (CNN), and the extracted features are fused to obtain the road dynamic features; Dynamically filter the neighboring vehicle groups based on the target vehicle's speed, direction of movement, and lane information; Calculate the statistical characteristics of the adjacent vehicle group, the statistical characteristics including at least one of the group density, the average speed of the group vehicles, the average acceleration, and the average steering angle; Based on the relative motion relationship between the target vehicle and its neighboring vehicles, the characteristics reflecting the consistency of traffic flow direction and the group interaction intensity characteristics reflecting the degree of interaction are calculated. By integrating the statistical features, directional consistency features, and group interaction intensity features, the vehicle group perception features are obtained.

5. The method according to claim 1, characterized in that, The spatiotemporal graph attention network model includes a spatial attention layer and a temporal attention layer; The spatial attention layer is used to aggregate information about neighboring nodes based on node features and edge features; The time attention layer is used to aggregate the feature information of nodes at different time steps.

6. The method according to claim 5, characterized in that, The spatiotemporal graph is input into a spatiotemporal graph attention network model to predict the collision risk between targets within the scheduling region, generating a risk matrix including: The spatiotemporal graph is processed by the spatial attention layer and the temporal attention layer to obtain updated node features containing temporal context information; The updated node features are decoded into predicted trajectories of each target within the scheduling area; Based on the predicted trajectory, calculate the collision time between any two targets; Based on the collision time, a quantified collision risk value between the two targets is determined to generate the risk matrix.

7. The method according to claim 1, characterized in that, Calculating the global priority for the sensing data within the scheduling area based on the risk matrix includes: For each sensed object, calculate the global priority of the sensed object. : in For the set of vehicles of the demand-aware object o, The collision risk values ​​for vehicles at o and v. To send the communication coverage of vehicle s, and These are the weighting coefficients.

8. The method according to claim 7, characterized in that, Based on the global priority, communication time slots are dynamically allocated to selected vehicles in a time-division multiplexing manner to broadcast their sensing data, which includes at least: For each perceived object o, the set of vehicles that can perceive it... In the middle, select to make the utility value The largest vehicle, acting as the sender, selects data packets from its set of sensed objects. To maximize the overall priority and obtain the optimal data packet. ; All sending tasks are prioritized by group. The data are sorted in descending order, and a greedy algorithm is used to allocate conflict-free communication time slots. Each time slot allows only one vehicle to send data, and the time slot length is dynamically adjusted according to the amount of data to be sent. ,in For channel rate, To perceive The amount of data; After the vehicle sends a data packet, a demand update instruction is broadcast to vehicles within the coverage area. If the perceived object o is the demand vehicle of vehicle v, o is removed from the local demand list. The central node updates its local priority based on the updated requirement list.

9. The method according to claim 8, characterized in that, Based on the global priority, dynamically allocating communication time slots to selected vehicles in a time-division multiplexing manner to broadcast their sensing data also includes: If a new vehicle is detected, a sudden change in collision risk, or a change in network topology, the global priority calculation must be re-executed before entering the next scheduling cycle.

Images