A non-conservative intelligent vehicle decision planning integrated method
By employing a multi-layer directed graph convolutional neural network encoding and decoding architecture and graph convolutional neural networks in intelligent vehicles, the future trajectories of traffic participants are predicted, solving the problem of conservative decision-making in complex environments for intelligent vehicles. This achieves efficient trajectory planning and decision-making, improving traffic efficiency and comfort.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2023-04-04
- Publication Date
- 2026-07-21
AI Technical Summary
Existing intelligent vehicles make overly conservative decisions in complex and heterogeneous environments, resulting in low traffic efficiency. Furthermore, existing models struggle to effectively predict the impact of future social interactions, leading to conservative trajectory planning and reduced comfort and traffic efficiency.
A social interaction knowledge learning model based on a multi-layer directed graph convolutional neural network encoding and decoding architecture is adopted. It combines graph convolutional neural networks and temporal dependency capture networks, and plans the trajectory of intelligent vehicles online after offline training. By predicting the future trajectories of all traffic participants, the environmental state is updated and decisions are made iteratively to avoid erroneous decisions.
It improves the traffic efficiency of intelligent vehicles in complex and heterogeneous environments, avoids erroneous decisions, realizes non-conservative decision-making and planning, and enhances the comfort and traffic efficiency of intelligent vehicles.
Smart Images
Figure CN116523156B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent vehicle trajectory planning, and in particular to a non-conservative integrated method for intelligent vehicle decision-making and planning in complex and heterogeneous environments. Background Technology
[0002] With the development of intelligent vehicles and people's desire for fast and safe travel, current intelligent driving vehicles, due to their overly conservative decision-making in complex and heterogeneous environments, suffer from low traffic efficiency due to frequent interference from pedestrians and cyclists. Therefore, developing a non-conservative intelligent vehicle with the driving characteristics of a professional driver is one way to improve the utilization and comfort of intelligent vehicles.
[0003] To enable intelligent vehicles to reach their destinations quickly and safely, the current mainstream architecture prioritizes action decision-making followed by trajectory planning. Action decision-making is the primary factor determining the efficiency of intelligent vehicle traffic. Dynamic Bayesian networks, sharing the same architecture as human driver decision-making processes, offer good scalability and interpretability, and are therefore widely used. They have achieved good results in lane-changing decisions for intelligent vehicles on highways. Deep learning algorithms are similar, but both face the challenge of dealing with the vast and variable amount of environmental information in complex and heterogeneous environments. This makes it difficult for decision-making models to converge easily or achieve complete coverage, hindering their generalization in complex and heterogeneous environments. Furthermore, the architecture of prioritizing action decision-making before trajectory planning relies on current traffic participant information, neglecting the impact of predicted social interactions on the future trajectory. This makes the planned trajectory conservative, leading to unnecessary deceleration or even sudden deceleration in the intelligent vehicle. This significantly reduces the comfort and traffic efficiency of the intelligent vehicle.
[0004] Researchers have also proposed an integrated architecture for model-based predictive decision-making and planning, but its purpose is to avoid flawed decisions that lead to infeasible planning trajectories. This does not address the issue of predicting future social interactions to improve traffic efficiency. Furthermore, it still faces the challenge of dealing with the vast and variable amount of environmental information in complex and heterogeneous environments, a challenge inherent in the aforementioned decision-making models. Summary of the Invention
[0005] The purpose of this invention is to provide a non-conservative integrated decision-making and planning method for intelligent vehicles in complex and heterogeneous environments. This method can adapt to complex and heterogeneous environments, deeply analyze and predict social interactions, improve vehicle traffic efficiency, and avoid erroneous decisions that lead to infeasible planned trajectories.
[0006] The objective of this invention can be achieved through the following technical solutions:
[0007] A non-conservative integrated decision-making and planning method for intelligent vehicles in complex and heterogeneous environments includes the following steps:
[0008] A social interaction knowledge learning model is constructed and trained offline. The input of the social interaction knowledge learning model is the position, speed, acceleration and category information of traffic participants, and the output is the predicted trajectory of different traffic participants.
[0009] The system acquires real-time online status data of surrounding traffic participants and intelligent vehicle status data, and then stitches them together to obtain the environmental status, which includes the position, speed, acceleration, and category information of all traffic participants.
[0010] Using the environmental state as input to the trained social interaction knowledge learning model, the predicted trajectories of all traffic participants, including intelligent vehicles, are obtained.
[0011] Update the environmental state based on the predicted trajectory;
[0012] The updated environmental state is re-input into the social interaction knowledge learning model to iteratively complete the intelligent vehicle trajectory decision planning. The planned trajectory of the intelligent vehicle is the result of splicing the first point of the predicted trajectory obtained in each iteration.
[0013] The social interaction knowledge learning model includes an encoding part and a decoding part. The encoding part constructs location maps, velocity maps, acceleration maps, and category maps based on the location, speed, acceleration, and category information of traffic participants, respectively, incorporating expert experience. Based on four parallel spatial relationship capture networks and temporal dependency capture networks, it captures the topological structure relationships and temporal dependencies of the location maps, velocity maps, acceleration maps, and category maps, respectively, to obtain the hidden states of location, speed, acceleration, and category. The decoding part decodes the hidden states of location, speed, acceleration, and category, respectively, and updates the model parameters based on the difference between the predicted trajectory and the actual trajectory to obtain the future trajectory prediction results for different traffic participants.
[0014] The method for constructing the location map is as follows:
[0015] G P =(U P E P )
[0016]
[0017]
[0018]
[0019] In the formula: G P Represents location diagram; UP For nodes in the location map, the location is represented by the position of the traffic participant; E P The edges of the location graph represent the interaction relationships influenced by their relative positions. This represents the relative position vector between traffic participant i and traffic participant j; It is the expert-experienced correction coefficient for the relative positional relationship introduced; It is the velocity vector of traffic participant i, representing the direction of the center line of his field of view.
[0020] The method for constructing the velocity map is as follows:
[0021] G V =(U V E V )
[0022]
[0023]
[0024]
[0025] In the formula: G V Represents velocity graph; U V For nodes on the speed map, the speed is represented by the speed of traffic participants; E V The edges of the velocity graph represent the interaction relationships influenced by relative velocity. It is an expert-empirical correction coefficient for the relative velocity relationship. This represents the relative position vector between traffic participant i and traffic participant j; These are the velocity vectors of traffic participants i and j, respectively.
[0026] The method for constructing the acceleration map is as follows:
[0027] G A =(U A E A )
[0028]
[0029]
[0030]
[0031] In the formula: G A Represents acceleration graph; U A For nodes in the acceleration graph, the acceleration of traffic participants is used; E A The edges of the acceleration graph represent the interaction relationships under the influence of relative acceleration. It is the expert-empirical correction coefficient for the relative acceleration relationship introduced; It is the acceleration vector of traffic participant i; It is the acceleration vector of traffic participant j; This represents the relative position vector between traffic participant i and traffic participant j.
[0032] The method for constructing the category diagram is as follows:
[0033] G C =(U C E C )
[0034]
[0035]
[0036]
[0037] In the formula: G C Representative category diagram; U C For each node in the category graph, represents the risk coefficient of a traffic participant, where c∈[1,2,3] corresponds to pedestrians, bicycles, and cars, respectively; E C The edges of the category graph represent the degree of danger. This represents the expert experience correction factor related to the category of transportation agency; It is the velocity vector of traffic participant i. It is the velocity vector of traffic participant j; This represents the relative position vector between traffic participant i and traffic participant j.
[0038] The spatial relationship capture network employs a graph convolutional neural network.
[0039] The time dependency capture network is a temporal convolutional network, a long short-term neural network, a recurrent neural network, or a gated neural network.
[0040] The specific steps for updating the environmental state based on the predicted trajectory are as follows: remove the first point from the historical trajectories of all traffic participants, and then concatenate the removed historical trajectories with the first point of the predicted trajectory to update the environmental state.
[0041] The status data of surrounding traffic participants comes from onboard sensors installed on the smart car.
[0042] Compared with the prior art, the present invention has the following beneficial effects:
[0043] (1) For the social interaction knowledge learning part, this invention adopts an encoding and decoding architecture based on multi-layer directed graph convolutional neural networks, wherein the encoding part includes a position map, velocity map, acceleration map and category map. Figure 4 The invention comprises several components. The social interaction knowledge learning model is capable of learning to select interfering features from multiple pieces of information, identifying interfering neighbors from multiple interacting traffic participants, and considering the differentiated social attributes of different traffic participants. The decoding component uses the error between the actual future trajectory and the predicted trajectory of traffic participants to evaluate the effectiveness of the model's social interaction learning and to correct the undetermined coefficients in the social interaction knowledge learning model. This is crucial for predicting social interactions and forms the basis for considering the impact of social interaction predictions on future trajectories. Based on social interaction knowledge learning, this invention can deeply analyze social interactions within an integrated decision-making and planning approach, addressing the challenges of abundant and variable environmental information in complex and heterogeneous environments.
[0044] (2) Regarding the online rolling planning part, this invention adopts a rolling planning architecture to predict the future trajectories of all participants and use the future trajectory of the intelligent vehicle as the planned trajectory. Each trajectory planning step is achieved through multi-step iterative trajectory planning, meaning each step only obtains the trajectory points for the next step. After obtaining new trajectory points, the first point of the current model input is discarded and appended to the current model input to form a new model input for the next planning step. This iterative approach considers the impact of predictions on future trajectories based on social interaction, thereby improving the traffic efficiency of intelligent driving vehicles and avoiding erroneous decisions. This is a key aspect of the non-conservative integrated decision-making and planning method for intelligent vehicles.
[0045] (3) The social interaction knowledge learning model of the present invention is constructed and trained offline and called when it is used for iterative trajectory planning online, which reduces the amount of online computation and improves the response speed. Attached Figure Description
[0046] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0047] Figure 2 This is a schematic diagram of the decoder principle in one embodiment;
[0048] Figure 3 Flowchart for initial trajectory prediction;
[0049] Figure 4 Flowchart for updating environment status. Detailed Implementation
[0050] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0051] This embodiment provides a non-conservative integrated decision-making and planning method for intelligent vehicles in complex and heterogeneous environments, such as... Figure 1 As shown, it includes the following steps:
[0052] 1) Construct and train a social interaction knowledge learning model offline.
[0053] like Figure 1 As shown in the upper part, the offline social interaction knowledge learning model adopts an encoding and decoding architecture based on multi-layer directed graph convolutional neural networks. The encoding part constructs location, velocity, acceleration, and category maps incorporating expert experience based on the location, velocity, acceleration, and category information of traffic participants. It then uses four parallel spatial relationship capturing networks and temporal dependency capturing networks to capture the topological and temporal dependencies of these maps, obtaining the hidden states of location, velocity, acceleration, and category. The decoding part decodes these hidden states and evaluates their effectiveness based on the difference between the predicted and actual trajectories, updating the model parameters to obtain future trajectory predictions for different traffic participants.
[0054] Specifically, the method for constructing the location map is as follows:
[0055] G P =(U P E P )
[0056]
[0057]
[0058]
[0059] In the formula: G P Represents location diagram; U P For nodes in the location map, the location is represented by the position of the traffic participant; E P The edges of the location graph represent the interaction relationships influenced by their relative positions. This represents the relative position vector between traffic participant i and traffic participant j; It is the expert-experienced correction coefficient for the relative positional relationship introduced; It is the velocity vector of traffic participant i, representing the direction of the center line of his field of view.
[0060] Obstacles outside the agent's field of view will not affect the future trajectory of traffic participants. Based on this, assuming the traffic participants' field of view is within 180°, and considering experience with traffic participants' avoidance, the closer their relative positions, the more pronounced their interaction and antagonism. Therefore, using the traffic participant's speed as the center line of the field of view, when the angle between the traffic participant's speed and their relative position is less than or equal to ±90°, it will affect the prediction of the traffic participant's trajectory, and the magnitude of this impact depends on both the speed and the relative position. Based on this, an expert-experienced correction coefficient for the aforementioned relative position relationship is determined.
[0061] The method for constructing the velocity map is as follows:
[0062] G v =(U V E V )
[0063]
[0064]
[0065]
[0066] In the formula: G V Represents velocity graph; U V For nodes on the speed map, the speed is represented by the speed of traffic participants; E V The edges of the velocity graph represent the interaction relationships influenced by relative velocity. It is an expert-empirical correction coefficient for the relative velocity relationship.
[0067] Within the visible field of view, the interaction between traffic participants depends not only on their relative positions but also on their relative speeds and accelerations. For example, even if two participants are close together, the conflict between them is not intense because their relative speeds and accelerations are parallel. Conversely, if they are far apart but their relative speeds and accelerations are perpendicular, their interaction exhibits stronger conflict. Therefore, the impact of interaction between traffic participants gradually increases as their relative speeds and accelerations increase and their angle increases to perpendicularity. This allows us to determine the expert-empirical correction coefficients for the aforementioned relative speed relationship and the following relative acceleration relationship.
[0068] The method for constructing the acceleration diagram is as follows:
[0069] G A =(U A E A )
[0070]
[0071]
[0072]
[0073] In the formula: G A Represents acceleration graph; U A For nodes in the acceleration graph, the acceleration of traffic participants is used; E A The edges of the acceleration graph represent the interaction relationships under the influence of relative acceleration. It is the expert-empirical correction coefficient for the relative acceleration relationship introduced; It is the acceleration vector of traffic participant i; It is the acceleration vector of traffic participant j.
[0074] In complex and heterogeneous environments, pedestrians, bicycles, and cars share roads. However, due to the inconsistencies in the size, mobility, and materials of these traffic participants, the risk levels differ among the various agents. This is the fundamental reason for the different social attributes among different traffic participants.
[0075] This embodiment introduces an expert-experienced category graph, using traffic participant hazard coefficients as nodes and hazard levels as edges. Expert-experienced correction coefficients related to traffic agent categories are introduced into the edges. The traffic participant hazard coefficient depends on the traffic participant category, with nodes 1, 2, and 3 defined for pedestrians, bicycles, and cars, respectively. When there are identical traffic participants ahead, the smaller the relative distance, the higher the hazard level. Similarly, a perpendicular velocity vector angle indicates a higher hazard level than a parallel velocity vector angle. Therefore, the hazard level is related to the relative distance between traffic participants and the angle between their velocity vectors.
[0076] Therefore, the method for constructing the category graph is as follows:
[0077] G c =(U c E c )
[0078]
[0079]
[0080]
[0081] In the formula: G C Representative category diagram; U CFor each node in the category graph, represents the risk coefficient of a traffic participant, where c∈[1,2,3] corresponds to pedestrians, bicycles, and cars, respectively; E C The edges of the category graph represent the degree of danger. This represents the expert experience correction coefficient related to the category of transportation agency.
[0082] Considering the excellent topology learning ability of graph convolutional neural networks, the spatial relationship capture network adopts graph convolutional neural networks, and it is not recommended to use other networks.
[0083] For networks that capture time dependencies, temporal convolutional networks or other sequential neural networks can be used, such as long short-term neural networks, recurrent neural networks, or gated neural networks.
[0084] In this embodiment, the decoding part uses a convolutional neural network. It's worth noting that the error between the actual future trajectory and the predicted trajectory of traffic participants is used to evaluate the effectiveness of the model's social interaction learning and to correct the undetermined coefficients in the social interaction knowledge learning model. For example... Figure 2 As shown, the input to the convolutional neural network is the hidden state output by the encoding part, including position, velocity, acceleration, and category hidden states. i The hidden layer, whose number of units can be increased depending on the computing platform's capabilities, contains learned weighted matrix coefficients and activation function coefficients. The output of the convolutional neural network is the future trajectory of traffic participants, including horizontal and vertical coordinates.
[0085] 2) Real-time online acquisition of status data of surrounding traffic participants and intelligent vehicles, and splicing processing to obtain the environmental status.
[0086] This embodiment obtains the state data of surrounding traffic participants through onboard sensors of the intelligent driving vehicle, such as radar, lidar, and cameras. Unlike the input of offline social interaction knowledge learning models, the state of the intelligent vehicle needs to be added as one of the model inputs here. The environmental state includes the position, speed, acceleration, and category information of all traffic participants.
[0087] 3) Using the environmental state as input to the trained social interaction knowledge learning model, the predicted trajectories of all traffic participants, including intelligent vehicles, are obtained.
[0088] The initial trajectory prediction process is as follows: Figure 3 As shown, the state data of the intelligent vehicle and the state data of the surrounding traffic participants are stitched together to obtain the overall environmental information, which is then used as the input of the social interaction knowledge learning model to obtain a prediction of the trajectories of all traffic participants.
[0089] 4) Update the environmental state based on the future trajectories of all traffic participants predicted in the previous step.
[0090] It is worth noting that, such as Figure 4 As shown, considering the impact of prediction error iteration, only the first point of the predicted trajectory is used as the new state of the traffic participants, and this first point is concatenated with the input used in this prediction (i.e., the historical trajectory relative to this prediction). To ensure that the input dimension of the model remains unchanged, the first point of the input used in this prediction is discarded, forming the input of a new social interaction knowledge learning model. This is the basis for predicting the impact of future social interactions on the planned trajectory. That is, the method for updating the environmental state is: for the current step, the first point of all traffic participants' historical trajectories is removed, and the removed historical trajectories are concatenated with the first point of the predicted trajectory to update the environmental state.
[0091] 5) Re-input the updated environmental state into the social interaction knowledge learning model and iteratively complete the intelligent vehicle trajectory decision planning (until the iteration prediction time equals the trajectory time to be planned). The planned trajectory of the intelligent vehicle is the splicing result of the first point of the predicted trajectory obtained in each iteration.
[0092] For the online rolling planning part, it uses a social interaction knowledge learning model trained offline to plan the trajectory of the intelligent vehicle. It's important to note that trajectory planning is achieved through multi-step iterative planning; each step only obtains the trajectory points for the next step and uses these points as input to the model for the next planning iteration. Then, the trajectory points obtained from the first point of the predicted trajectory in each iteration are stitched together to form the planned trajectory for intelligent driving, thus completing the planning process.
[0093] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A non-conservative integrated decision-making and planning method for intelligent vehicles in complex and heterogeneous environments, characterized in that, Includes the following steps: A social interaction knowledge learning model is constructed and trained offline. The input of the social interaction knowledge learning model is the position, speed, acceleration and category information of traffic participants, and the output is the predicted trajectory of different traffic participants. The system acquires real-time online status data of surrounding traffic participants and intelligent vehicle status data, and then stitches them together to obtain the environmental status, which includes the position, speed, acceleration, and category information of all traffic participants. Using the environmental state as input to the trained social interaction knowledge learning model, the predicted trajectories of all traffic participants, including intelligent vehicles, are obtained. Update the environmental state based on the predicted trajectory; The updated environmental state is re-input into the social interaction knowledge learning model to iteratively complete the intelligent vehicle trajectory decision planning. The planned trajectory of the intelligent vehicle is the result of splicing the first point of the predicted trajectory obtained in each iteration. The social interaction knowledge learning model includes an encoding part and a decoding part. The encoding part constructs location maps, velocity maps, acceleration maps, and category maps based on the location, speed, acceleration, and category information of traffic participants, respectively, incorporating expert experience. Based on four parallel spatial relationship capture networks and temporal dependency capture networks, it captures the topological structure relationships and temporal dependencies of the location maps, velocity maps, acceleration maps, and category maps, respectively, to obtain the hidden states of location, speed, acceleration, and category. The decoding part decodes the hidden states of location, speed, acceleration, and category, respectively, and updates the model parameters based on the difference between the predicted trajectory and the actual trajectory to obtain the future trajectory prediction results for different traffic participants.
2. The non-conservative intelligent vehicle decision-making and planning integrated method under complex heterogeneous environments according to claim 1, characterized in that, The method for constructing the location map is as follows: In the formula: Representative location map; The nodes in the location map are represented by the locations of traffic participants; The edges of the location graph represent the interaction relationships influenced by their relative positions. Indicates traffic participants With traffic participants The relative position vectors between them; It is the expert-experienced correction coefficient for the relative positional relationship introduced; Traffic participants The velocity vector represents the direction of the center line of his field of view.
3. The non-conservative intelligent vehicle decision-making and planning integrated method under complex heterogeneous environments according to claim 1, characterized in that, The method for constructing the velocity map is as follows: In the formula: Represents the velocity graph; The nodes on the speed map are represented by the speed of traffic participants; The edges of the velocity graph represent the interaction relationships influenced by relative velocity. It is an expert-empirical correction coefficient for the relative velocity relationship. Indicates traffic participants With traffic participants The relative position vectors between them; , They are traffic participants , The velocity vector.
4. The non-conservative intelligent vehicle decision-making and planning integrated method in complex heterogeneous environments according to claim 1, characterized in that, The method for constructing the acceleration map is as follows: In the formula: Represents acceleration graph; For nodes in the acceleration graph, the acceleration of traffic participants is used as the representation. The edges of the acceleration graph represent the interaction relationships under the influence of relative acceleration. It is the expert-empirical correction coefficient for the relative acceleration relationship introduced; Traffic participants The acceleration vector; Traffic participants The acceleration vector; Indicates traffic participants With traffic participants The relative position vectors between them.
5. The non-conservative intelligent vehicle decision-making and planning integrated method according to claim 1, characterized in that, The method for constructing the category diagram is as follows: In the formula: Representative category diagram; The nodes in the category graph represent the risk coefficient of traffic participants. These correspond to pedestrians, bicycles, and cars, respectively. The edges of the category graph represent the degree of danger. This represents the expert experience correction factor related to the category of transportation agency; Traffic participants The velocity vector, Traffic participants The velocity vector; Indicates traffic participants With traffic participants The relative position vectors between them.
6. The non-conservative intelligent vehicle decision-making and planning integrated method according to claim 1, characterized in that, The spatial relationship capture network employs a graph convolutional neural network.
7. The non-conservative intelligent vehicle decision-making and planning integrated method according to claim 1, characterized in that, The time dependency capture network is a temporal convolutional network, a long short-term neural network, a recurrent neural network, or a gated neural network.
8. The non-conservative intelligent vehicle decision-making and planning integrated method in complex heterogeneous environments according to claim 1, characterized in that, The specific steps for updating the environmental state based on the predicted trajectory are as follows: remove the first point from the historical trajectories of all traffic participants, and then concatenate the removed historical trajectories with the first point of the predicted trajectory to update the environmental state.
9. The non-conservative intelligent vehicle decision-making and planning integrated method in complex heterogeneous environments according to claim 1, characterized in that, The status data of surrounding traffic participants comes from onboard sensors installed on the smart car.