Method for constructing open unmanned vehicle group based on perception contribution

Abstract

The application belongs to the field of unmanned driving, and proposes an open unmanned vehicle group construction method based on perception contribution, comprising: step 1. Perception contribution; step 2. Prediction of factors influencing perception contribution in an open scenario; step 3. Construction of an unmanned vehicle group model based on perception contribution; and step 4. Unmanned vehicle group formation method. The application improves the understanding ability of unmanned vehicles to the surrounding environment, improves the understanding ability of unmanned vehicles to the surrounding environment, provides effective protection for intelligent cooperation of unmanned vehicle groups, and thus can effectively help solve the current problems of single intelligent agent of unmanned vehicles, such as lane changing, avoidance, safety distance adjustment, and road traffic efficiency, and has important significance and application value for promoting the healthy and rapid development of unmanned vehicles in urban scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of unmanned driving, and in particular to an open unmanned vehicle group construction method based on perception contribution. BACKGROUND

[0002] Existing research on unmanned vehicle groups mainly focuses on closed scenarios, highway scenarios, and open scenarios. Researchers have proposed a semi-centralized vehicle group formation method in a closed scenario, which selects a leading node based on the relative speed and position between unmanned vehicles, and forms a vehicle group with unmanned vehicles having the same destination. The complexity of the road in the closed scenario is low, and fewer interference factors need to be considered. Researchers have proposed a semi-centralized unmanned vehicle group centered on a leading node in a highway scenario. The group defines five states and their transitions in the group. Then, based on these states, an unmanned vehicle group formation method is proposed, and an unmanned vehicle group model is established. Simulation results show that the method can establish a stable unmanned vehicle group with mutual connection between its members. However, when the leading node or the secondary leading node leaves the group or fails, the structure of the group becomes chaotic. Researchers have proposed a distributed group construction method based on sidelink consensus in a highway scenario. The distributed unmanned vehicle group has no central control center, and each group member has the ability to make independent decisions. The method first uses sidelink consensus to extract the state of the unmanned vehicle, and then constructs a distributed group based on sidelink consensus. However, the above methods are mainly for unmanned vehicle group formation in closed and highway scenarios, and are not suitable for direct application to open scenarios. Because the road topology in the open scenario is complex and there are human-driven vehicles, traffic lights, roadside obstacles, and pedestrians as interference factors. Researchers have proposed a group construction method in an open scenario. The method defines the pre-perception degree, vehicle activity degree, and movement similarity of unmanned vehicles, and selects a leading node based on this. The leading node selects corresponding secondary leading nodes, and other free nodes can join the group through them.

[0003] In summary, existing research on unmanned vehicle groups focuses on highway scenarios and open scenarios. The road topology in closed scenarios and highway scenarios is simple, and there are fewer external interference factors. The existing highway scenario unmanned vehicle group construction method is not suitable for open scenarios, but the unmanned vehicle group in an open scenario has only a semi-centralized structure, and still has the following problems:

[0004] 1) Load concentration: the existing unmanned vehicle group is characterized by a semi-centralized structure, with a leading node as the center. The network and computing load is concentrated on the leading node, which has a high requirement for its information processing capacity. If the leading node leaves or cannot communicate with other members, it will cause serious waste of computing resources and network resources.

[0005] 2) No consideration of perception cooperation: the existing unmanned vehicle group assumes that a single unmanned vehicle can obtain completely accurate perception results within its perception range. However, in an open scenario, interference factors and blind areas will cause a single unmanned vehicle to be unable to obtain completely correct perception results, and they do not establish cooperative perception between unmanned vehicles.

[0006] 3) Only consider stable communication links between vehicle nodes as the basis for construction: they only construct a vehicle group that can communicate with each other and maintain a stable communication link as much as possible, and do not consider sharing perception results through stable vehicle group links with other vehicle group members, nor do they consider whether the perception demand relationship between vehicle group members has an impact on vehicle group construction. SUMMARY

[0007] To solve the above problems, the present application proposes an open unmanned vehicle group construction method based on perception contribution.

[0008] The present application constructs an unmanned vehicle group model by maintaining perception contribution interaction between unmanned vehicle nodes. First, based on edge computing, a collaborative interaction mode of unmanned vehicle nodes is proposed: perception contribution, which helps to improve the environmental understanding ability of unmanned vehicles when the perception is limited; then, based on Transformer, the influencing factors of perception contribution are predicted; finally, the unmanned vehicle group model is constructed and solved by using a multi-objective optimization method. Through simulation experiments, the effectiveness of contribution to perception information and its reliability, the efficiency of perception information contribution, the ability to continuously contribute to perception information, and the rationality of the formed unmanned vehicle group are verified, which provides an important guarantee for the cooperative perception of unmanned vehicle group, thereby effectively helping to solve the current problems of single intelligent agent of unmanned vehicles such as lane changing, avoidance, safety distance adjustment, and road traffic efficiency, and has important significance and application value for promoting the healthy and rapid development of unmanned intelligent cooperation.

[0009] Technical scheme of the present application:

[0010] The open unmanned vehicle group construction method based on perception contribution comprises the following steps:

[0011] Step 1. Perception contribution

[0012] Step 1.1 Measurement of contribution;

[0013] Step 1.2 Measurement of required degree;

[0014] Step 1.3 Measurement of contribution degree.

[0015] Step 2. Prediction of perception contribution influencing factors in open scene

[0016] Step 2.1 Formalization of edge features of unmanned vehicle;

[0017] Step 2.2 Contribution influencing factor prediction algorithm;

[0018] Step 3. Construction of unmanned vehicle group model based on perception contribution

[0019] Step 3.1 Definition of unmanned vehicle group;

[0020] Step 3.2 Group contribution degree;

[0021] Step 3.3 Group property;

[0022] Step 3.4 Construction of unmanned vehicle group model;

[0023] Step 4. Forming of unmanned vehicle group

[0024] Step 4.1 Multi-objective optimization solution of unmanned vehicle group model;

[0025] Step 4.2 Unmanned vehicle group forming algorithm;

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] The present application discloses that unmanned vehicle groups are constructed through inter-vehicle perception contribution, which improves the understanding ability of unmanned vehicles to the surrounding environment, provides effective guarantee for intelligent cooperation of unmanned vehicle groups, and thus can effectively help solve the problems of current unmanned single agents, such as lane changing, avoidance, safety distance adjustment, and road passing efficiency, and has important significance and application value for promoting the healthy and rapid development of unmanned vehicles in urban scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 Flowchart of the present application;

[0029] Figure 2 Perception contribution between nodes;

[0030] Figure 3 TIFF algorithm framework;

[0031] Figure 4 Scene 1 schematic diagram;

[0032] Figure 5 Scene 2 schematic diagram;

[0033] Figure 6 Relationship between average contribution of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 1 and the maximum number of vehicles at different maximum speeds;

[0034] Figure 7 Relationship between average persistence of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 1 and the maximum number of vehicles at different maximum speeds;

[0035] Figure 8 Relationship between average accessibility of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 1 and the maximum number of vehicles at different maximum speeds;

[0036] Figure 9 Relationship between average real-time performance of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 1 and the maximum number of vehicles at different maximum speeds;

[0037] Figure 10 Performance of the node survival time, node stable time and vehicle group survival time of the CPF algorithm and its comparative algorithm in scenario 1;

[0038] Figure 11 Relationship between average contribution of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 2 and the maximum vehicle speed at different maximum numbers;

[0039] Figure 12 Relationship between average persistence of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 2 and the maximum vehicle speed at different maximum numbers;

[0040] Figure 13 Relationship between average accessibility of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 2 and the maximum vehicle speed at different maximum numbers;

[0041] Figure 14 Relationship between average real-time performance of the CPF algorithm and its comparative algorithm for the unmanned vehicle group in scenario 2 and the maximum vehicle speed at different maximum numbers;

[0042] Figure 15 Performance of the node survival time, node stable time and vehicle group survival time of the CPF algorithm and its comparative algorithm in scenario 2. DETAILED DESCRIPTION

[0043] The technical solutions provided by the present application will be further described below with reference to specific embodiments and the accompanying drawings. The advantages and features of the present application will be more apparent in the light of the following description.

[0044] As Figure 1 shown, the present application specifically includes the following five aspects:

[0045] Step 1. Perception contribution;

[0046] Step 1.1 Measurement of contribution;

[0047] Step 1.2 Measurement of demand;

[0048] Step 1.3 Degree of contribution;

[0049] Step 2. Prediction of perception contribution influencing factors in open scenarios

[0050] Step 2.1 Formalization of edge features of unmanned vehicles;

[0051] Step 2.2 Contribution influencing factor prediction algorithm;

[0052] Step 3. Construction of unmanned vehicle group model based on perception contribution

[0053] Step 3.1 Definition of unmanned vehicle group;

[0054] Step 3.2 Degree of contribution of vehicle group;

[0055] Step 3.3 Nature of vehicle group;

[0056] Step 3.4 Construction of unmanned vehicle group model;

[0057] Step 4. Formation method of unmanned vehicle group

[0058] Step 4.1 Multi-objective optimization solution of unmanned vehicle group model;

[0059] Step 4.2 Unmanned vehicle group formation algorithm;

[0060] Step 5. Simulation experiment verification.

[0061] Detailed as follows:

[0062] Step 1 Perception contribution

[0063] This step aims to give the perception contribution interaction mode, so that the degree of contribution between unmanned vehicles can be calculated. Perception contribution refers to the unloading of perception demand from vehicles that need to perceive the front area in advance or have a perception blind area to their edge vehicles. The edge vehicles perceive the results through their own sensors and give feedback to the former, thereby realizing collaborative perception. Edge vehicles refer to vehicles that are closer to the pre-perception area or are more likely to obtain correct perception results in perception contribution. As shown in Figure 2 , unmanned vehicles v j and v l are edge nodes of vehicle v i . These edge nodes v j and v l perform real-time perception near the data source and contribute their perception information to vi Therefore, v i The correct perception result can be obtained in the pre-perception area of the edge node or in the case of obstacle blocking.

[0064] (I) Measure of contribution

[0065] As Figure 2 shown, the grid of the area where the object cannot be perceived by the sensor is represented in yellow, the grid is represented in green when the sensor perceives that there is no obstacle in the area, and the grid is represented in red when the sensor perceives that there is an obstacle in the area.v l The degree of occlusion o l (x) is represented as follows:

[0066]

[0067] wherein, c and represent no obstacle, obstacle, and invalid perception, respectively. No obstacle means that there is no object at the position x that prevents the autonomous vehicle from passing through, and obstacle means that the sensor perceives the presence of other autonomous vehicles, manned vehicles, pedestrians, or obstacles at the position x, which prevents the autonomous vehicle from passing through. Invalid perception means that the sensor cannot obtain valid perception information about the position x due to interference or occlusion by obstacles. When the sensor of the autonomous vehicle perceives an obstacle or no obstacle, it is valid perception information, but when the sensor is interfered or occluded by obstacles, it is invalid perception data, based on which the autonomous vehiclev l The degree of occlusion o l (x) at the position x can further measure thev l contribution of the position x in the perception range, that is:

[0068] w l (x) = |2o l (x) - 1| (2)

[0069] Based on equation (2), the perception contribution of v l is the accumulation of the contributions of all grids in its perception range, which can be represented as:

[0070]

[0071] wherein, B l = {b i |i∈{1,2,…,m i}} is the set of m l (>0) grids in the perception range of the vehicle v i

[0072] (II) Measure of demand ​

[0073] Internal demand: unmanned vehicle v i The internal demand of unmanned vehicle v i is the demand for the perception information provided by other vehicles in the perception area that is blocked by obstacles or other interference factors within its perception range. is the sum of the demand of each position x in the grid set B i , that is:

[0074]

[0075] where, is the demand of v i at position x.

[0076] External demand: it is the perception data outside the perception range of unmanned vehicle v i , as shown in definition 3.2, if v i can obtain the perception results of the pre-perception area, i.e. the area similar to its driving direction and closer to it, in advance, it can better ensure the accuracy of the decision of v i . The perception demand of v i for the area outside the pre-perception area is 0. On the contrary, its perception demand for point x in the pre-perception area depends on its speed, relative distance and direction. Then the external perception demand of v i is:

[0077]

[0078] where, v i is the speed of v i , is the braking time of vehicle v i , θ is the angle between the moving direction of v i and the connecting line of position x, and l is the distance between v i and x, that is:

[0079]

[0080] where, (x i , y i ) and (x, y) are the coordinates of vehicle v i and x respectively.

[0081] Therefore, the perception demand of v i at position x, i.e. the product of internal and external perception demands, can be represented as:

[0082]

[0083] (III) Measurement of contribution degree

[0084] The contribution degree is used to measure the effectiveness of information transfer between vehicles through perception contribution. i and j The contribution degree of v j and the demand of v i can be calculated, expressed as:

[0085]

[0086] where B i and B l represent the grid set within the perception range of v i and v l respectively.

[0087] Step 2 Prediction of factors affecting perception contribution in open scenarios

[0088] In open scenarios, the perception contribution between autonomous vehicles is significantly affected by external interference factors. To address this problem, by analyzing the edge features of autonomous vehicles to analyze the degree of influence of external interference factors on the perception contribution relationship between autonomous vehicles, a perception contribution influence factor prediction method based on Transformer is proposed.

[0089] (I) Formalization of edge features of autonomous vehicles

[0090] The edge features of autonomous vehicles are the basic features of the input of the perception contribution influence factor prediction method, which refer to the perception features within the perception range of the vehicle at a specific time. The edge features of v i at time t are represented as follows:

[0091]

[0092] where, are the speed, acceleration, horizontal, coordinate and direction of v i at time t respectively; and represent the set of autonomous vehicles, manned vehicles, roadside obstacles, traffic signals and pedestrians within the perception range of v i at time t.

[0093] (1) The features of manned vehicle m i at time t are represented as:

[0094]

[0095] where v j , a j , x j , y j , ∈j respectively, the speed, acceleration, level, coordinate and direction of o j

[0096] (2) Roadside obstacles o k The feature representation of o at time t is:

[0097]

[0098] wherein, wherein l k and w k are the length and width of o k

[0099] (3) Traffic signal light l z The feature representation of l at time t is:

[0100]

[0101] wherein, represents the state of the signal light, 0, 1 and 2 represent red, green and yellow respectively.

[0102] (4) Pedestrian p u The feature representation of p at time t is:

[0103]

[0104] It can be obtained that the relative feature representation of v i and v j at t is:

[0105]

[0106] wherein, v j ∈ V i ∪ M i is an autonomous or manned vehicle, and λ respectively represent the relative speed, acceleration and distance between v i and v j . In addition, and are the movement directions of v i and v j at time t.

[0107] (B) Contribution influencing factor prediction method

[0108] (1) Contribution influencing factor prediction framework: taking the edge features of unmanned vehicles as input, predicting the contribution influencing factors based on the idea of Transformer, providing basis for the formation of vehicle groups. For example, Figure 3 ​​As shown, the contribution influence factor prediction framework is composed of five parts: the edge feature input layer of the unmanned vehicle, the edge feature extraction layer of the unmanned vehicle, the attention layer, the contribution influence factor output layer, and the regression of the contribution influence. The specific implementation is as follows:

[0109] Edge feature input layer of the unmanned vehicle: The input of the contribution influence factor prediction method is composed of the edge features and relative features within the perception range of the two vehicles at time t, which is expressed as:

[0110] γ=γ i (t)∪γ j (t) (15)wherein, γ i (t) and γ j (t) represent the features of v i and v j , that is

[0111]

[0112] Edge feature extraction layer of the unmanned vehicle: The task of this layer is to extract the edge features of the unmanned vehicle. Two layers of convolution kernels with k are used for convolution to convert the edge features of the unmanned vehicle into a distributed semantic representation, and combined with the position encoding module, position encoding is added for each time step so that the prediction framework can understand the time sequence.

[0113] Attention layer: The task of this layer is to capture the time dependence between the unmanned vehicles. The structure is consistent with the encoder part in the Transfrmer framework, which is composed of a self-attention mechanism and a feedforward neural network. The self-attention mechanism calculates the attention weight of each element in the sequence output by the edge feature extraction layer of the unmanned vehicle to other elements; the feedforward neural network passes the output of the self-attention mechanism through two fully connected networks and performs nonlinear activation on it. Through the self-attention layer, the importance of various elements of the edge features of the unmanned vehicle is effectively weighed, and the time dependence relationship is captured.

[0114] Contribution influence factor output layer: The task of this layer is to output the final prediction result. The specific implementation is to add a multi-layer perceptron after the output of the attention layer to predict the following 6 contribution influence factors:

[0115] 1) The contribution interruption probability χ ij , that is, the probability of interruption of the perception contribution relationship between v i and v j ;

[0116] 2) The contribution transmission throughput δ ij , that is, the number of perception information successfully transmitted between v i and v j through perception contribution per unit time;

[0117] 3) contribution reliability k ij , the probability that the information transfer between v i and v j is reliable in the future;

[0118] 4) contribution latency , the latency of v j receiving the perception result of v i contribution;

[0119] 5) contribution efficiency ξ ij , the time spent by v j sending unit test data to v i ;

[0120] 6) duration τ ij , the longest time that the perception contribution relationship between v i and v j is maintained.

[0121] Regression of contribution influencing factors: After the output prediction value of the contribution influencing factor output layer, compare the prediction with the true value, use the following formula as the loss function to calculate the loss value, and update the model parameters.

[0122]

[0123] where, is the learnable parameter of the prediction model, m is the number of training samples, γ i represents an input feature of the i-th training sample, Y(γ i ) represents the true value of the perception contribution influencing factor, represents the prediction value.

[0124] (2) Contribution influencing factor prediction algorithm: The specific steps of contribution influencing factor prediction are as follows:

[0125] Training phase:

[0126] 1) Initialize the weight parameters of the contribution influencing factor prediction framework, and divide the training set into several sub-training sets of the same size according to the step size (algorithm 1 lines 1-3);

[0127] 2) For each iteration of each sub-training set, input the training set into the contribution influencing factor prediction model to obtain the prediction value of the contribution influencing factor (algorithm 1 lines 4-7);

[0128] 3) Calculate the loss by the loss function , update the gradient by back propagation, and update the prediction model parameters (algorithm 1 lines 8-13);

[0129] 4) reach the maximum number of iterations, stop training, and output the trained contribution impact factor prediction model (lines 11-13);

[0130] Prediction phase:

[0131] 5) input the input data into the trained contribution impact factor prediction model, obtain the predicted value of the contribution impact factor, and output (algorithm 1 lines 14-17);

[0132] The contribution impact factor prediction algorithm is shown in Algorithm 1:

[0133]

[0134] Step 3: Construction of an autonomous vehicle group model based on perception contribution

[0135] (I) Definition of an autonomous vehicle group

[0136] Definition 3.3 Autonomous vehicle group: In an open scenario, members of an autonomous vehicle group can collaborate by contributing perception information to each other, effectively filtering external environmental disturbances through perception contribution. An autonomous vehicle group in an open scenario is represented as:

[0137] G i =(V i ,E i ,η,f) (19)

[0138] where,

[0139] 1) V i ={v i |i∈{1,2,…}} is the set of group members of G i , and v i is an autonomous vehicle;

[0140] 2) E i ={e i |i∈{1,2,…}} is the set of perception contribution relationships between group members in G i ;

[0141] 3) η={0,1} is the set of vehicle types, 0 and 1 representing group members and free vehicles, respectively;

[0142] 4) f:V→η represents a vehicle type mapping function.

[0143] (II) Group contribution degree

[0144] In the car group, the members of the car group enhance their own perception ability by effectively using the perception information obtained by the edge node. This enables them to filter out interference from blind areas, blockages of roadside obstacles, and obtain necessary perception information in their pre-perception area in advance. The car group contribution measures the overall perception ability of individuals to the surrounding environment. It is calculated by the sum of all contributions within the car group, i.e.:

[0145]

[0146] wherein v i and v l are members of the car group G i .

[0147] (Three) Car group properties

[0148] (1) Persistence: Persistence represents the persistence of the perception contribution relationship between the members of the car group G i , i.e.:

[0149]

[0150] wherein τ ij is the perception contribution duration between v i and v j , and N i is the number of autonomous vehicles in G i .

[0151] (2) Reachability: Reachability refers to the ability of the car group G i to form mutually reachable perception contribution relationships between group members in an open scenario. The greater the reachability of the car group, the higher the reliability of the perception contribution relationship between the members of the car group. Use the contribution interruption probability, perception contribution transmission throughput, and contribution reliability to measure, i.e.:

[0152]

[0153] wherein χ ih , δ ij , κ ij are the contribution interruption probability, perception contribution transmission throughput, and contribution reliability between v i and v j .

[0154] (3) Real-time: There is a time overhead when autonomous vehicles transmit perception information through perception contribution. Use the contribution delay and efficiency to measure the real-time of the car group, i.e.,

[0155]

[0156] wherein ξ ij and v respectively i and v j The contribution latency and efficiency between them.

[0157] (iv) Construction of the autonomous vehicle swarm model

[0158] Considering the contribution, continuity, reliability, and real-time performance of the autonomous vehicle swarm, the autonomous vehicle swarm model is as follows:

[0159]

[0160] Among them, G i It is a fleet of driverless vehicles in an open environment. This indicates a group that can join the vehicle group G. i The set of free nodes, l ij It is v i and v j The relative distance between them Indicates v j Belongs to v i Within the pre-sensory region, |∈ i -∈ j |=0 indicates v i and v j Having the same direction of motion, and A collection of idle vehicles in an open scene.

[0161] Step 4: Method for Forming an Autonomous Vehicle Swarm

[0162] (I) Multi-objective optimization solution of autonomous vehicle swarm model

[0163] The candidate vehicle selection method for autonomous vehicle swarms based on multi-objective optimization first transforms the autonomous vehicle swarm model into an optimization problem with four objectives, and incorporates the constraints of the swarm. Using particle swarm optimization to obtain the Pareto optimal solution helps determine the best set of vehicles to join the autonomous vehicle swarm.

[0164] Autonomous vehicle swarm model Π i Pareto optimal solution for:

[0165]

[0166] in, It is the Pareto optimal solution to a multi-objective optimization problem, i.e., G i The best collection of vehicles, It is the solution set of an n-dimensional 0-1 vector model of an autonomous vehicle swarm. It is an n×1 dimensional 0-1 vector, when When the i-th element is 1, it indicates agreement with the autonomous vehicle node v.i Join car group G i ,when When the i-th element is 0, it indicates rejection. Ultimately, those who agree to join the car group G... i set of autonomous vehicle nodes for:

[0167]

[0168] The algorithm for solving the autonomous vehicle swarm model based on multi-objective optimization is shown in Algorithm 3.2, and is described in detail below:

[0169] 1) In an open scenario, autonomous vehicles communicate with each other; some free vehicles that contribute the most within a certain range are given the ability to build an autonomous vehicle swarm. When these vehicles receive requests to join from other free vehicles, they perform an initial screening of the requests based on the vehicle's direction, position, and hop count; then, they count the number of vehicles requesting to join to determine the solution set. The size n; next, initialize the number of particles N, the inertia weight w, the learning factors ψ1 and ψ2, and the maximum number of iterations. (Line 1 of Algorithm 2);

[0170] 2) Generate a random binary vector and a random vector as the position and particle velocity, respectively. Evaluate each particle using an autonomous vehicle swarm model, assign each particle its individual best position to its current position, and classify all particles into non-dominated and dominated solutions based on Pareto dominance. Then, assign the set of all non-dominated solutions to... As the current Pareto frontier (Algorithm 2, lines 2-8);

[0171] 3) During the iteration phase, perform The next iteration; in each iteration, for each particle, from A non-dominated solution is randomly selected and assigned to the particle's local optimum position; then, Algorithm 1 updates the particle's velocity and position, evaluates it using an autonomous vehicle swarm model, and updates its individual optimal position; if a particle is a non-dominated solution and is not... If any solution dominates, then add it to In the middle; when v i Obtain the Pareto optimal solution hour, The 0-1 vector in each dimension indicates whether a free node can be added to group G. i (Algorithm 2, lines 9-27);

[0172] 4) Finally, A vehicle node with a median value of 1 is added to the set of nodes that agree to join the vehicle group. and output (lines 28-30 of Algorithm 2).

[0173]

[0174]

[0175] (II) Autonomous vehicle group formation algorithm

[0176] In an open scenario, autonomous vehicles communicate with each other; some free vehicles that contribute the most value within a certain range are given the ability to build autonomous vehicle groups, and others join autonomous vehicle groups by joining them; the autonomous vehicle group formation algorithm based on contributed perception (CPF) is shown in Algorithm 3.3, and is described as follows:

[0177] 1) Free vehicle v j Send a join request carrying its own perception (lines 1-4 of Algorithm 3);

[0178] 2) Node v i When receiving the request sent by v j , v i solves according to the autonomous vehicle group model to calculate whether the size of its vehicle group reaches the upper limit N max , and judges whether v j can join its vehicle group; if approved, v j establishes a perception contribution relationship with v i , and other free nodes can also join the vehicle group through v j (lines 5-18 of Algorithm 3);

[0179] 3) If any node in G i does not receive a data packet from v j within a period of time, it is considered that v j leaves the vehicle group G i (lines 19-23 of Algorithm 3);

[0180]

[0181] Step 5 Simulation experiment verification

[0182] In order to verify the effectiveness of the autonomous vehicle group formation method, the simulation software is used for simulation experiment in the present application.

[0183] (1) Simulation experiment data and method

[0184] This application uses SUMO (Simulation of Urban Mobility) to construct two open scenarios:

[0185] Scenario 1: such as Figure 4 The image shows a square-layout open-world scenario with 12 two-way and three-lane roads, each 400 meters long. There are 4 crossroads, 8 T-junctions, 4 L-junctions, and 16 traffic lights. 10% of the vehicles are designated as passenger vehicles. 10 to 20 roadside obstacles of varying sizes are randomly placed on each road, along with 100 pedestrians.

[0186] Scenario 2: such as Figure 5 As shown, the area is located near the Yan'an Elevated Road in Shanghai, China, covering a 4x4 km area. It extends from 121.45845°E to 121.49426°E and from 31.21975°N to 31.24810°N. The area includes highways, elevated roads, intersections, and T-junctions.

[0187] (2) Evaluation indicators

[0188] To evaluate the performance of CPF, state-of-the-art technologies for autonomous vehicle swarming were selected as benchmark methods to evaluate the CPF of the proposed method. These include: Vehicle Formation Method on Highways (AVGF), Vehicle Formation Method in Urban Scenarios (AVGM), and Distributed Vehicle Formation Method Based on Sidechain Consensus (SCCAVGF). The evaluation metrics are as follows: Average vehicle swarm contribution, persistence, reliability, and timeliness measure the average individual's overall perception ability of the surrounding environment; persistence of CP relationships among vehicle swarm members under disturbance studies the ability of autonomous vehicle swarm members to form reliable CP relationships under open disturbance scenarios; and the time overhead of transmitting perception data among vehicle swarm members. These are the average of the contribution, persistence, reliability, and timeliness of all vehicle groups throughout the simulation. Node survival time is the cumulative survival time of a node in the vehicle swarm throughout the entire simulation. Node stability time represents the cumulative duration of a node within the same vehicle swarm throughout the entire simulation period. Vehicle swarm survival time represents the average survival time of the vehicle swarm throughout the entire simulation.

[0189] (3) Analysis of simulation experiment results

[0190] Figure 6The figure shows the relationship between average group contribution and maximum number of vehicles. Each subfigure shows the boxplot of CPF, AVGM, SCCAVGF and AVGF. In each boxplot, the upper bound represents the maximum value of the method, the lower bound represents the minimum value of the method, the dotted line represents the average value of the method, and the black dot represents the distribution of the results. The same representation is used in the remaining boxplots in this application. Among the four methods, CPF performs best, indicating that the CPF formed by the unmanned vehicle group has a stronger understanding of the surrounding environment. SCCAVGF, AVGM and AVGF form a connected and stable vehicle group without considering the perception information as the evaluation standard, resulting in a weaker understanding of the environment than CPF. Due to the reduced mobility of unmanned vehicles, the average group contribution increases with the increase of the number of vehicles. By observing each subfigure, it can be observed that CPF shows a lower average group contribution growth rate at 20 km / h compared to 60 km / h. This is because in the case of 20 km / h, the movement speed of the vehicle is more limited than the number of vehicles. When the maximum speed of the vehicle is 60 km / h, the increase in the maximum number of vehicles leads to the limitation of vehicle mobility, resulting in the increase of the size of the vehicle group and the amount of perception information transmitted between members.

[0191] Figure 7 The figure shows the average group persistence of CPF, AVGM, SCCAVGF, AVGF under different maximum number of vehicles. The experimental results show that the vehicle group formed by CPF has a longer perception contribution interaction time. In AVGF, when the leading node fails or leaves the vehicle group, it will cause the vehicle group to disintegrate, resulting in AVGF being at a significant disadvantage in terms of average group persistence. Compared with the AVGF algorithm, the AVGM algorithm not only has a secondary leading node set, but also selects a new leading node from the candidate set when the leading node leaves, resulting in suboptimal performance. CPF considers whether the vehicle can maintain a long-term perception contribution state after joining the vehicle group, so it has a higher average group persistence than the other three nodes. When the maximum speed is 20, 30, 40 km / h, the average group persistence increases with the increase of the maximum number of vehicles. This is because the limited mobility of unmanned vehicles leads to the increase of the size of the unmanned vehicle group. As shown in Figure 7 The figure shows the average group persistence of CPF, AVGM, SCCAVGF, AVGF under different maximum number of vehicles. The experimental results show that the vehicle group formed by CPF has a longer perception contribution interaction time. In AVGF, when the leading node fails or leaves the vehicle group, it will cause the vehicle group to disintegrate, resulting in AVGF being at a significant disadvantage in terms of average group persistence. Compared with the AVGF algorithm, the AVGM algorithm not only has a secondary leading node set, but also selects a new leading node from the candidate set when the leading node leaves, resulting in suboptimal performance. CPF considers whether the vehicle can maintain a long-term perception contribution state after joining the vehicle group, so it has a higher average group persistence than the other three nodes. When the maximum speed is 20, 30, 40 km / h, the average group persistence increases with the increase of the maximum number of vehicles. This is because the limited mobility of unmanned vehicles leads to the increase of the size of the unmanned vehicle group. As shown in

[0192] Figure 8The relationship between the average group reliability and the maximum number of vehicles is illustrated. Among the four methods, CPF shows superior performance, indicating that the vehicle group formed using CPF exhibits a more reliable perception contribution relationship than the vehicle groups formed using the other three methods. From Figure 8 It can be seen that CPF has the lowest average group reliability when the maximum number of vehicles is 200, and reaches a peak at 600. The average group reliability increases with the number of vehicles. This can be attributed to the limited movement of vehicles in scenario 1, resulting in a high number of perception contribution interaction pairs within the vehicle group, and a low interruption probability, resulting in a high contribution reliability. The subplots present the relationship between the average group contribution and the maximum vehicle speed. It is found that as the maximum speed increases, the average group contribution rate first increases and then decreases. This can be attributed to the fact that a higher speed allows vehicles to travel a greater distance within a given time unit, thereby increasing the contribution of perception information. However, as the speed difference between vehicles increases, the residence time of autonomous vehicles in the pre-perception area decreases, resulting in a decrease in perception contribution interaction between vehicles. The box plot shows that the average group contribution of CPF is 0.202, 0.493, 1.13, 1.35, and 1.74 for 200, 300, 400, 500, and 600 vehicles, respectively. This value increases with the number of vehicles, indicating that the perception demand and contribution increase with the number of vehicles, resulting in more frequent perception contribution interactions.

[0193] Figure 9 The relationship between the average group real-time performance and the maximum number of vehicles is illustrated. The box plot analysis shows that when the maximum speed is 20 km / h, the average group punctuality rate values of CPF, AVGM, SCCAVGF, and AVGF are 1241, 952, 510, and 253, respectively. Compared with AVGM, SCCAVGF, and AVGF, the average group punctuality rate of CPF increases by 30.3%, 143.7%, and 3.91 times, respectively. These findings indicate that CPF is superior to the other three comparison methods in terms of the time cost of vehicle group members' perception contribution-related. As the number of vehicles increases, the average group punctuality rate shows an upward trend. This can be attributed to the increase in the number of vehicles in the scenario, resulting in an increase in the number of vehicle pairs with lower contribution delays and higher communication efficiency.

[0194] Figure 10CPF, AVGM, SCCAVGF and AVGF are shown in terms of node survival time, node stable time and vehicle group survival time. As shown in the subgraph, the vehicle survival time of CPF decreases with the increase of maximum vehicle speed. This is because the mobility of vehicles increases with the increase of speed, so that the time of individual autonomous vehicles in a single group is shorter. As shown in the box plot, the average node survival time of CPF is 19% higher than SCCAVGF, 28% higher than AVGF, and 4% higher than AVGM, indicating that CPF can make the survival time of vehicle nodes in the unmanned vehicle group longer than the same kind. As can be seen from the subgraph, CPF has the longest average stable time of nodes compared with the other three comparison methods. This shows that among the four methods, CPF considers the perception needs and contribution degree of vehicles when building the vehicle group, so CPF can make the vehicle nodes maintain a longer perception contribution time in the same vehicle group.

[0195] Figure 11 The relationship between average group contribution and maximum vehicle speed in scenario 2 is shown. CPF performs best among the four methods, indicating that CPF has stronger environmental perception ability than the three similar methods. Because scenario 2 is large and the distribution of unmanned vehicles is sparse, the average group contribution of all methods is low, which leads to a decrease in the amount of effective perception information transmitted through perception contribution in the vehicle group. The average group contribution increases with the increase of speed from 20 to 40 km / h, and further decreases. With the increase of maximum vehicle speed, the average group contribution first increases and then decreases. This is because the distance they move in a unit of time increases with the increase of speed, which leads to more perception information being contributed. In addition, the residence time of unmanned vehicles in the pre-perception area decreases with the increase of speed difference between vehicles, leading to a decrease in the interaction of perception contribution between vehicles. The box plot shows that the average group contribution average of CPF is 0.202, 0.493, 1.13, 1.35, 1.74 in 200, 300, 400, 500, 600 vehicles, respectively. Perception needs and perception contribution increase with the increase of vehicle number, leading to more frequent CP interaction.

[0196] Figure 12is the relationship between the average group persistence and the maximum vehicle speed in scenario 2. No matter the maximum speed of the vehicles, the average group persistence of CPF is significantly higher than the other three comparative methods. This is because CPF considers the impact of inter-vehicle perception on group formation, resulting in longer perception contribution interaction time than the other three nodes. As shown in the box plot, the average group persistence of CPF is 3.09, 3.54, 3.7, 3.84, and 4.22 at 200, 300, 400, 500, and 600, respectively. It increases with the increase of the number of vehicles, which is due to the increase of the number of vehicles leading to more vehicles having the same destination, similar location, and similar motion behavior. In addition, they can also maintain long-term perception contribution relationships in the group. Since the perception contribution duration decreases with the increase of the speed difference between the unmanned vehicles, the average group persistence decreases. The average group persistence of CPF fluctuates less than the other three comparative methods. In other words, CPF has a higher tolerance to speed changes, indicating that the unmanned vehicle group formed by CPF is more suitable for survival in an open scenario with a larger speed difference.

[0197] Figure 13 is the relationship between the average group reliability and the maximum vehicle speed in scenario 2. CPF has the highest average group reliability at each maximum vehicle number compared to the same vehicles, indicating that CPF has a more reliable ability to share perception information among group members. As can be seen from the box plot in the figure, the average value of the average group reliability of CPF increases with the increase of the number of vehicles. This is because the mobility restriction of the vehicles increases with the increase of the number of vehicles, making the perception contribution link between group members more reliable.

[0198] Figure 14 shows the relationship between the average group real-time performance and the maximum vehicle speed in scenario 2. It is worth noting that no matter the maximum speed of the vehicles, CPF is always superior to the other three comparative methods in terms of average group real-time performance. This indicates that CPF effectively reduces the time overhead of perception contribution among group members. The average group real-time performance shows an upward trend with the increase of the number of vehicles. This can be attributed to the decrease in mobility of unmanned vehicles with the increase of vehicle density. Therefore, group members can choose vehicles with lower latency and higher efficiency for perception contribution, thereby improving the average group real-time performance. Figure 15 illustrates the performance of CPF, AVGM, SCCAVGF, and AVGF in terms of node survival time, node stable time, and group survival time. It is worth noting that CPF outperforms the other three comparative methods. This indicates that the group formed by CPF can maintain a longer survival time, the duration of the vehicle node as a group member is longer, and the duration of maintaining within the same group is longer.

[0199] The above description is only a description of the preferred embodiments of the present application, and is not any limitation on the scope of the present application. Any modification or change made by any person skilled in the art according to the disclosed technical content above should be regarded as an equivalent effective embodiment, and belongs to the protection scope of the technical scheme of the present application.

Claims

1.A method for constructing an open autonomous vehicle fleet based on perception contribution, characterized in that, It comprises the following steps: Step 1. Define the perception contribution, including the measurement of contribution, the measurement of demand, and the measurement of contribution degree; Step 2. Predict the influence factors of perception contribution in an open scenario; Step 3. Construct a model of autonomous vehicle group based on perception contribution; Step 4. Autonomous vehicle group formation method; In step 1: The measurement of contribution: The sensor sensing range includes the area grid where the object cannot be sensed, the area grid where there is no obstruction in the sensed area, and the area grid where there is an obstacle in the sensed area. The position in the sensing range x The degree of occlusion is represented as follows: in, , and These represent unobstructed passage, obstructed passage, and invalid perception, respectively; based on autonomous vehicles. In position x degree of obstruction Further measurement Location within the sensing range x The contribution, namely: Based on , The perceptual contribution of a grid is the sum of the contributions of all the grids within its perceptual range, denoted as: wherein, is a vehicle within a perception range a set of squares; The measurement of demand: driverless vehicles The internal requirement of autonomous vehicles is to provide perception information from other vehicles for areas within their perception range that cannot be perceived due to obstacles or other interference; Internal demand It is a grid set Each position The accumulation of demand, that is: wherein is at a location demand; Unmanned vehicle The external demand is the unmanned vehicle The perception data outside the perception range, if The perception result of the pre-perception area, which is similar to the driving direction and closer to the unmanned vehicle, can be obtained in advance, so that the accuracy of the decision can be better guaranteed ; The perception demand for the area outside the pre-perception area is ; on the contrary, the perception demand of the unmanned vehicle for the point in the pre-perception area depends on its speed, relative distance and direction; then The external perception demand of the unmanned vehicle is: wherein, is the speed, denotes the braking time of the vehicle , is the moving direction and position of the connecting line, and is the distance between the vehicle and the connecting line, i.e.: wherein and are coordinates of the vehicle and respectively. Obtained Position down The perceived demand, i.e. the product of the inner and outer perceived demands, is expressed as: The contribution degree is measured by the effectiveness of the information transfer between vehicles through the perception contribution; and The contribution degree is calculated by the contribution of and the demand of and is expressed as: wherein, and respectively represent and a set of grids within the perception range; In step 2, the prediction of influence factors of perception contribution in an open scenario includes input features, construction of contribution influence factor prediction framework, and design of contribution influence factor prediction algorithm, wherein The edge features of autonomous vehicle are the input features of the contribution influence factor prediction framework, which are the features within the perception range of autonomous vehicle at a specific time; The proposed contribution influencing factor prediction framework uses the edge features of autonomous vehicles as input and predicts contribution influencing factors based on Transformer to provide a basis for vehicle swarm formation. The framework consists of five parts: a vehicle edge feature input layer, an autonomous vehicle edge feature extraction layer, an attention layer, a contribution influencing factor output layer, and a regression layer for contribution impact. The prediction of contribution influencing factors yields the following six outputs: contribution interruption probability. ,Right now and The probability of interruption in the perceived contribution relationship; contribution transmission throughput. That is, the contribution made by perception per unit time. and The amount of sensed information successfully transmitted between them; contribution to reliability That is, in the near future and The reliable transmission of information between them is probabilistic; contribution delay Received Time delay when the contribution is perceived; contribution efficiency ,Right now Send unit test data Time spent; duration and The longest duration of the perceived contribution relationship between them; The contribution influence factor prediction algorithm: initialize the weight parameters of the contribution influence factor prediction framework, and divide the training set into several sub-training sets of the same size according to the step size; For each iteration of each sub-training set, input the training set into the contribution influence factor prediction model to obtain the predicted value of the contribution influence factor; through a loss function compute loss, backpropagate to update gradients, update prediction model parameters ​ When the maximum number of iterations is reached, stop training and output the trained contribution influence factor prediction model; Input the input data into the trained contribution influence factor prediction model to obtain the predicted value of the contribution influence factor and output it; In step 3: In an open scenario, autonomous vehicle group members collaborate by contributing perception information to each other, effectively filtering external environmental interference through perception contribution; the autonomous vehicle group in an open scenario is represented as: Wherein, 1) is a set of platoon members, and is an unmanned vehicle; 2) is a set of perceived contribution relationships among the members of the vehicle group; 3) For the vehicle type set, 0 and 1 represent group members and free vehicles, respectively; 4) represents a vehicle type mapping function; In the autonomous vehicle group, the members enhance their own perception ability by using the perception information obtained by the edge nodes; the contribution degree of the group measures the overall perception ability of the members to the surrounding environment, which is calculated by the sum of all contributions, i.e. wherein and are members of a vehicle fleet ​ Persistent representation of vehicle group the persistent ability of members to perceive contribution relationships between one another, i.e.: wherein, is to the perceived contribution duration between, is the number of unmanned vehicles in Reachability is the car platoon In an open scenario, the inter-vehicle reachability of the perception contribution among the group members can be formed; the greater the reachability of the car platoon, the higher the reliability of the perception contribution relationship among the members of the car platoon; the contribution interruption probability, the perception contribution transmission throughput and the contribution reliability are used to measure, that is: wherein, , , are the contribution interruption probability, the perceived contribution transmission throughput and the contribution reliability, respectively, between and ; There is a time overhead when autonomous vehicles transmit perception information through perception contribution; the contribution delay and efficiency are used to measure the real-time performance of the group, i.e. wherein, and respectively and contribution latency and efficiency; Based on the contribution degree, sustainability, reliability, and real-time performance of the autonomous vehicle group, the autonomous vehicle group model is constructed as: wherein, is a group of unmanned vehicles in an open scenario, represents a set of free nodes joining the group of vehicles , is the relative distance between and represents is within a pre-perception area belonging to , indicates and have the same direction of motion, and is a set of free vehicles in an open scenario; Step 4 comprises: Step 4.1 Autonomous vehicle group model based on multi-objective optimization and its optimization algorithm based on multi-objective optimization; First, convert the autonomous vehicle group model into an optimization problem with four objectives, and combine the constraints of the group; use particle swarm optimization to obtain the Pareto optimal solution to determine the optimal vehicle set to join the autonomous vehicle group; Unmanned vehicle fleet model Pareto optimal solution is: in, It is the Pareto optimal solution to a multi-objective optimization problem, i.e. The best collection of vehicles, It is an n-dimensional The solution set of a vector-based autonomous vehicle swarm model; where, yes Vie Vector, when The Middle i When the dimension element is 1, it indicates agreement with the autonomous vehicle node. Join the car group ,when The Middle i When the dimension element is 0, it indicates rejection; ultimately, it agrees to join the car group. set of autonomous vehicle nodes for: The optimization algorithm based on multi-objective optimization of the autonomous vehicle group model (shown as algorithm 2): 1) In an open scenario, autonomous vehicles communicate with each other. The free vehicle with the largest contribution within a certain range is given the ability to build an autonomous vehicle swarm. When these vehicles receive requests to join from other free vehicles, they perform initial screening based on the vehicle's direction, position, and hop count. Then, the number of requesting vehicles is counted to determine the solution set. Size Next, initialize the number of particles. Inertial weight Learning factor and and the maximum number of iterations ; 2) generating a random binary vector and a random vector as the position and particle velocity, respectively, evaluating each particle with the unmanned vehicle group model, assigning the best position of each particle to its current position, dividing all particles into non-dominated solutions and dominated solutions according to the Pareto dominance relationship, and assigning the set of all non-dominated solutions to the current Pareto front as the current Pareto front; 3) In the iteration phase, do sub-iterations; in each iteration, for each particle, randomly select one non-dominated solution from and assign it to the local best position of the particle; then, update the velocity and position of the particle, evaluate it with the AV fleet model, and update its global best position; if a particle is a non-dominated solution and is not dominated by any solution in , add it to ; when get the Pareto optimal solution , the vector in each dimension of represents whether a free node can join the group or not; 4) Finally, the vehicle node with median value 1 joins the set of nodes that agree to join the platoon and outputs and outputs; Step 4.2 Autonomous vehicle group formation algorithm Autonomous vehicle group formation algorithm based on perception contribution (CPF) (shown as algorithm 3): 1) free vehicle sending a join request carrying its own awareness 2) node received when a request is sent, According to the solution of unmanned vehicle group model, calculate whether the size of the vehicle group reaches the upper limit , and determine whether it can join the vehicle group; if approved, with establish a perception contribution relationship, and other free nodes can also join the vehicle group by ; 3) If any node in does not receive a data packet from for a period of time, then the node is considered to have left the platoon . 2.The method of claim 1, wherein, In step 2, the contribution influence factor prediction algorithm is algorithm 1: 。 3.The method of claim 1, wherein, In step 4, the optimization algorithm based on multi-objective optimization of the autonomous vehicle group model is algorithm 2: Algorithm 2: Algorithm for solving the model of the unmanned vehicle group based on multi-objective optimization input: number of particles dimensionality inertia weight learning factor and maximum number of iterations Output: Set of nodes that agreed to join the platoon 。 4.The method of claim 1, wherein, In step 4, the algorithm 3: 。

Images