Automatic driving decision-making method, device, computer equipment, readable storage medium and program product

Abstract

The present application relates to an autonomous driving decision-making method, apparatus, computer device, computer-readable storage medium, and computer program product. The method comprises: obtaining environmental information of the target object's environment and multiple first candidate paths for the target object; determining multiple first predicted trajectories corresponding to each traffic participant in the current environment and first importance weights corresponding to each first predicted trajectory based on a trajectory prediction model and environmental information; the trajectory prediction model is trained using a decision cost function; based on the first candidate paths, the first predicted trajectories, the first importance weights, and the decision cost function, the decision costs corresponding to each first candidate path are determined, and the target path is determined from the first candidate paths based on the decision costs. The use of this method can improve the safety of autonomous driving.

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to an autonomous driving decision-making method, apparatus, computer equipment, computer-readable storage medium, and computer program product. Background Art

[0002] In complex traffic environments, autonomous vehicles will face more dangerous environmental conditions and more complex decision-making behaviors. The safe driving of autonomous vehicles requires accurate and reliable decision-making methods.

[0003] In traditional technology, the vehicle-mounted system of an autonomous vehicle includes a prediction module and a planning module. The prediction module predicts the future trajectories of traffic participants by sensing environmental information and information about traffic participants. The planning module combines environmental information and the future trajectories of traffic participants to determine a target path that does not overlap with or is a certain distance away from the future trajectory.

[0004] However, in traditional methods, the planning module has a poor utilization rate of the output results of the prediction module, and the effects of the prediction module and the planning module on information decision-making are independent of each other, which leads to poor safety of autonomous driving. Summary of the Invention

[0005] Based on this, it is necessary to provide an autonomous driving decision-making method, device, computer equipment, computer-readable storage medium and computer program product to address the above-mentioned technical problems.

[0006] In a first aspect, the present application provides an autonomous driving decision-making method, comprising:

[0007] Acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object;

[0008] Determining, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model is trained using a decision cost function;

[0009] Based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, a decision cost corresponding to each of the first candidate paths is determined, and a target path is determined among the first candidate paths according to the decision cost.

[0010] In one embodiment, before obtaining the environmental information of the environment in which the target object is located and the plurality of first candidate paths of the target object, the method further includes:

[0011] Obtaining a sample data set and a trajectory label and a path selection label corresponding to each sample data in the sample data set;

[0012] Performing prediction processing on each of the sample data according to the trajectory prediction model to be trained to obtain a second predicted trajectory and a second importance weight corresponding to each of the second predicted trajectories;

[0013] determining an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and determining a path planning loss based on the second importance weight, the decision cost function, and the path selection label;

[0014] The model parameters of the trajectory prediction model to be trained are updated by back propagation until the accuracy prediction loss and the path planning loss meet a preset loss threshold, thereby obtaining a trained trajectory prediction model.

[0015] In one embodiment, determining the path planning loss based on the second importance weight, the decision cost function, and the path selection label includes:

[0016] Constructing a historical sampling space according to the sample data set; the historical sampling space includes a plurality of historical candidate paths determined based on the sample data;

[0017] Among the historical candidate paths, determining the historical candidate path with the smallest distance to the path selection label as the expert trajectory;

[0018] A path planning loss is determined based on the expert trajectory and the decision cost function.

[0019] In one embodiment, before determining the decision cost corresponding to each first candidate path based on the first candidate path, the first predicted trajectory, the first importance weight, and the decision cost function, the method further includes:

[0020] Obtaining initial candidate paths, and screening the initial candidate paths according to a preset dynamic constraint sub-model to obtain pre-screened candidate paths;

[0021] The initially screened candidate paths are screened a second time according to a preset collision safety constraint sub-model to obtain a first candidate path.

[0022] In one embodiment, the decision cost function includes a decision safety cost function, a decision efficiency cost function, and a decision comfort cost function; the decision cost includes a decision safety cost, a decision efficiency cost, and a decision comfort cost;

[0023] The determining, based on the first candidate paths, the first predicted trajectories, the first importance weights, and the decision cost function, of the decision costs corresponding to the first candidate paths includes:

[0024] determining the decision safety cost based on the initial candidate path, the first predicted trajectory, the first importance weight, and the decision safety cost function;

[0025] determining the decision efficiency cost based on the initial candidate path, the destination coordinates, and a decision efficiency cost function;

[0026] The decision comfort cost is determined based on the initial candidate path and a decision comfort cost function.

[0027] In one embodiment, obtaining the initial candidate path includes:

[0028] Get the preset path starting point and path end point;

[0029] generating a global path based on the path starting point and the path end point;

[0030] A plurality of initial candidate paths for the target object are generated according to preset transverse and longitudinal offset distances and the global path.

[0031] In a second aspect, the present application also provides an autonomous driving decision-making device, comprising:

[0032] A first acquisition module is used to acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object;

[0033] a first prediction module, configured to determine, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model being trained using a decision cost function;

[0034] a planning module, configured to determine a decision cost corresponding to each of the first candidate paths based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, and to determine a target path among the first candidate paths according to the decision cost.

[0035] In one embodiment, the apparatus further comprises:

[0036] A second acquisition module is used to acquire a sample data set and a trajectory label and a path selection label corresponding to each sample data in the sample data set;

[0037] a second prediction module, configured to perform prediction processing on each of the sample data according to the trajectory prediction model to be trained, to obtain a second predicted trajectory and a second importance weight corresponding to each of the second predicted trajectories;

[0038] a loss calculation module, configured to determine an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and to determine a path planning loss based on the second importance weight, the decision cost function, and the path selection label;

[0039] The training module is used to update the model parameters of the trajectory prediction model to be trained through back propagation until the accuracy prediction loss and the path planning loss meet a preset loss threshold, thereby obtaining a trained trajectory prediction model.

[0040] In one embodiment, the loss calculation module is specifically configured to construct a historical sampling space based on the sample data set; the historical sampling space includes a plurality of historical candidate paths determined based on the sample data;

[0041] Among the historical candidate paths, determining the historical candidate path with the smallest distance to the path selection label as the expert trajectory;

[0042] A path planning loss is determined based on the expert trajectory and the decision cost function.

[0043] In one embodiment, the apparatus further comprises:

[0044] A first screening module is used to obtain initial candidate paths and screen the initial candidate paths according to a preset dynamic constraint sub-model to obtain pre-screened candidate paths;

[0045] The second screening module is used to perform a secondary screening on the initially screened candidate paths according to a preset collision safety constraint sub-model to obtain a first candidate path.

[0046] In one embodiment, the decision cost function includes a decision safety cost function, a decision efficiency cost function, and a decision comfort cost function; the decision cost includes a decision safety cost, a decision efficiency cost, and a decision comfort cost;

[0047] The planning module is specifically configured to determine the decision safety cost based on the initial candidate path, the first predicted trajectory, the first importance weight, and the decision safety cost function;

[0048] determining the decision efficiency cost based on the initial candidate path, the destination coordinates, and a decision efficiency cost function;

[0049] The decision comfort cost is determined based on the initial candidate path and a decision comfort cost function.

[0050] In one embodiment, the planning module is specifically used to obtain a preset path starting point and path end point;

[0051] generating a global path based on the path starting point and the path end point;

[0052] A plurality of initial candidate paths for the target object are generated according to preset transverse and longitudinal offset distances and the global path.

[0053] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0054] Acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object;

[0055] Determining, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model is trained using a decision cost function;

[0056] Based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, a decision cost corresponding to each of the first candidate paths is determined, and a target path is determined among the first candidate paths according to the decision cost.

[0057] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0058] Acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object;

[0059] Determining, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model is trained using a decision cost function;

[0060] Based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, a decision cost corresponding to each of the first candidate paths is determined, and a target path is determined among the first candidate paths according to the decision cost.

[0061] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the following steps:

[0062] Acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object;

[0063] Determining, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model is trained using a decision cost function;

[0064] Based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, a decision cost corresponding to each of the first candidate paths is determined, and a target path is determined among the first candidate paths according to the decision cost.

[0065] The above-mentioned autonomous driving decision-making method, device, computer equipment, computer-readable storage medium and computer program product analyze environmental information through a trajectory prediction model trained by a quantitative assessment of safety reflected by a decision cost function to obtain multiple first predicted trajectories of each traffic participant and a first importance weight of each first predicted trajectory. At the same time, the planning of the target path is also obtained by screening each first candidate path based on the constraints of the decision cost function and combining the first predicted trajectory and the first importance weight, so that the decision cost function is simultaneously introduced into the prediction and path planning of autonomous driving. The first importance weight is transmitted as safety information in the two stages of trajectory prediction and path planning, strengthening the connection between trajectory prediction and path planning, and improving the utilization rate of the output results of the trajectory prediction model in the path planning stage, thereby improving the accuracy of path planning guided by safety, and thereby improving the safety of autonomous driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying any creative work.

[0067] Figure 1 This is a diagram of an application environment of an autonomous driving decision-making method in one embodiment;

[0068] Figure 2 is a schematic diagram of a decision cost function in one embodiment;

[0069] Figure 3 Schematic diagram of a process for training a trajectory prediction model in one embodiment;

[0070] Figure 4 A schematic diagram of a flow chart of an example of training a trajectory prediction model in one embodiment;

[0071] Figure 5 A schematic diagram of a process for determining path planning loss in one embodiment;

[0072] Figure 6 FIG1 is a schematic diagram of a process for determining a first candidate path through path screening in one embodiment;

[0073] Figure 7 FIG1 is a flow chart of calculating decision cost according to a decision cost function in one embodiment;

[0074] Figure 8 A schematic diagram of a process for generating an initial candidate path in one embodiment;

[0075] Figure 9 This is a structural block diagram of an autonomous driving decision-making device in one embodiment;

[0076] Figure 10 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0077] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0078] In one embodiment, Figure 1 As shown, a method for autonomous driving decision-making is provided. This embodiment uses the method applied to a terminal as an example. The terminal can be an onboard terminal of an autonomous driving vehicle. It is understood that the method can also be applied to a server, or to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0079] Step 102: Acquire environmental information of the target object's environment and a plurality of first candidate paths of the target object.

[0080] In the embodiments of this application, an autonomous vehicle is used as the target object for illustration. The safe driving of an autonomous vehicle requires accurate and reliable decision-making to determine the control of the autonomous vehicle. The terminal selects the target path of the autonomous vehicle in the future time step based on the environmental information transmitted from the upstream, including the information of traffic participants, to achieve autonomous driving. Therefore, the structure used for autonomous driving decision-making in the terminal can be divided into a trajectory prediction model and a planning model. The trajectory prediction model is used to predict the possible future trajectories of other traffic participants in the current environment; the planning model is used to select the target path based on the prediction results output by the trajectory prediction model.

[0081] First, the terminal uses sensors (e.g., cameras, lidar, millimeter-wave radar, etc.) to collect environmental information around the vehicle. This information can include information about traffic participants, road structure, and obstacles. Traffic participant information includes the current position and speed of vehicles, pedestrians, bicycles, and other traffic participants; road structure information includes road structures such as lane markings, traffic signs, and traffic lights; and obstacle information includes buildings and roadblocks. The terminal then processes and fuses the sensor data to generate a comprehensive understanding of the current environment. Simultaneously, based on this environmental information, the terminal pregenerates multiple first candidate paths to cover different driving strategies. For example, different first candidate paths represent paths for lane maintenance, lane change, and deceleration.

[0082] Step 104 : determining a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each first predicted trajectory based on the trajectory prediction model and the environmental information.

[0083] Among them, the trajectory prediction model is trained through the decision cost function.

[0084] The first predicted trajectory and the first importance weight are output results of the trajectory prediction model in the application phase.

[0085] In the embodiment of the present application, the terminal constructs and trains the trajectory prediction model .in, For environmental data, is the decision cost function. For high-density traffic and high-risk scenarios such as intersections and highways, the terminal builds a trajectory prediction model based on a deep neural network. The trajectory prediction model uses environmental data As input, environmental data The features include: historical trajectories of traffic participants, interaction status of traffic participants, historical trajectories of the self-vehicle and future planning.

[0086] The terminal inputs environmental information, including road environment data, into the trajectory prediction model. By mapping the pre-trained environmental information with the first predicted trajectories and first importance weights in the trajectory prediction model, the model obtains multiple possible first predicted trajectories for each traffic participant in the current environment within a certain future time step, along with the first importance weights corresponding to each first predicted trajectory within that time step. This represents the risk and importance of the first predicted trajectory in the target path decision. The trajectory prediction model consists of a historical trajectory encoder, an ego vehicle state information encoder, an interaction feature encoder, a predicted trajectory decoder, and a planning importance perception module decoder.

[0087] Among them, the interactive feature encoder adopts a graph neural network structure, the historical trajectory encoder and the self-vehicle state information encoder adopt a network structure based on long short-term memory (LSTM), and the predicted trajectory decoder and the planning perception decoder uniformly adopt a network structure based on the gated recurrent unit (GRU).

[0088] The interaction feature encoder is used to encode the spatial interaction features between the ego vehicle and traffic participants, or between multiple traffic participants. In autonomous driving scenarios, the interaction relationships between traffic participants (e.g., vehicles, pedestrians, bicycles, etc.) are very complex. Using a graph neural network as an interaction feature encoder can effectively model the detailed spatial and topological relationships between the ego vehicle and traffic participants, as well as between each traffic participant. At the same time, at each time step, the position and state of traffic participants are changing dynamically. The use of a graph neural network can adapt to scenarios where the state of traffic participants is constantly changing. In addition, the topological invariance of graph neural networks makes them insensitive to the connection method of nodes, and can better handle the changing relationships between nodes (traffic participants) in complex traffic scenarios. Therefore, graph neural networks are more suitable for processing the interaction features between traffic participants and the ego vehicle in dynamic environments.

[0089] In the interaction feature encoder, traffic participants are represented as nodes in a graph, and interactions between them (e.g., distance, speed, etc.) are represented as edges. A graph neural network aggregates information from adjacent nodes through a message passing mechanism. Each node updates its own state based on the states (features) of its neighbors. This message passing process enables the interaction feature encoder to capture the mutual influence between traffic participants and learn complex interaction patterns. It also dynamically adjusts the graph edges based on the current environment, enabling flexible responses to dynamic interactions and ultimately generating interaction features for each traffic participant.

[0090] For the history trajectory encoder and the ego vehicle state information encoder, a long short-term memory (LSTM) structure can capture the temporal dependencies between the history trajectory and the ego vehicle state. The LSTM structure effectively addresses the vanishing and exploding gradient problems in long sequence training by introducing a gating mechanism (input gate, forget gate, and output gate). This allows the LSTM structure to effectively capture information over long timeframes, retaining and updating state over several time steps, making it suitable for processing complex history trajectory sequences. Therefore, the history trajectory encoder and the ego vehicle state information encoder adopt a long short-term memory (LSTM) structure. The LSTM structure uses a gating structure to control the flow of information, determining the extent to which each input affects the current hidden state and output. It uses memory cells to store long-term dependent information, while using a forget gate to selectively delete unnecessary information. At each time step, the LSTM structure updates its internal state and generates an output based on the current input, thereby encoding the time series of information and obtaining features between the history trajectory and the ego vehicle state.

[0091] The trajectory prediction and importance weighting in the prediction trajectory decoder and planning importance perception module decoder both involve processing time series data. In autonomous driving scenarios, the future location of each traffic participant is closely related to their current and past positions. The gating mechanism of the GRU structure enables it to adapt to this dynamic temporal dependency and effectively learn trends and patterns in time series data. The prediction trajectory decoder and planning perception decoder need to gradually generate prediction targets based on the current state. Therefore, in autonomous driving, trajectory and state information change rapidly and variably. The GRU structure makes the trajectory prediction model highly adaptable to rapidly changing environments.

[0092] Step 106 : Based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, a decision cost corresponding to each first candidate path is determined, and a target path is determined from the first candidate paths according to the decision cost.

[0093] In the embodiment of the present application, a decision cost function is pre-set in the planning model of the terminal, such as Figure 2 As shown in the figure, for driving risk scenarios and common planning task problems in autonomous driving, such as turning, going straight, etc., a decision cost function based on driving needs is constructed. The decision cost function mainly includes the decision safety cost function, the decision efficiency cost function and the decision comfort cost function. The unified decision cost function is based on the idea of ​​minimizing the optimal planning. The form of is the linear weighting of the above decision safety cost function, decision efficiency cost function and decision comfort cost function:

[0094]

[0095] in, represents the weight coefficient, ,Right now .

[0096] The terminal calculates the multiple first predicted trajectories corresponding to each traffic participant output by the trajectory prediction model and the first importance weight corresponding to each first predicted trajectory in combination with the decision cost function and the first candidate path to obtain the decision cost of each first candidate path. The decision cost evaluates the safety, efficiency and comfort of each first candidate path, and selects the first candidate path with the lowest decision cost as the optimal path, that is, as the target path.

[0097] The target path is considered part of the global path. That is, after obtaining a unique optimal trajectory as the target path, the ego vehicle will advance along the target path for one time step at the current moment. At the next time step, steps 102 through 106 will be repeated until the terminal determines that the ego vehicle has reached the end of the global path, at which point all processes will be terminated. During use, the decision frequency is set to 10 Hz, the historical and future prediction time domains are set to 6 and 20 frames, respectively, and the future decision time domain is set to 20 frames.

[0098] In the above-mentioned autonomous driving decision-making method, the trajectory prediction model trained by the quantitative evaluation of safety reflected by the decision cost function analyzes the environmental information to obtain multiple first predicted trajectories of each traffic participant and the first importance weight of each first predicted trajectory. At the same time, the planning of the target path is also obtained by screening each first candidate path based on the constraints of the decision cost function and the first predicted trajectory and the first importance weight, so that the decision cost function is introduced into the prediction and path planning of autonomous driving at the same time. The first importance weight is transmitted as safety information in the two stages of trajectory prediction and path planning, strengthening the connection between trajectory prediction and path planning, and improving the utilization rate of the output results of the trajectory prediction model in the path planning stage, thereby improving the accuracy of path planning guided by safety, and thus improving the safety of autonomous driving.

[0099] In an exemplary embodiment, Figure 3 As shown, before step 102, the trajectory prediction model needs to be trained first, and the method also includes steps 302 to 308. Among them:

[0100] Step 302: Obtain a sample data set and a trajectory label and a path selection label corresponding to each sample data in the sample data set.

[0101] In an embodiment of the present application, the terminal uses historical data in high-density vehicle scenarios such as intersections and highways as sample data to construct a sample data set. The sample data includes the historical trajectory of each traffic participant in the preset scenario, the interaction characteristics between each traffic participant and the self-vehicle and between each traffic participant, the historical trajectory of the self-vehicle and the future path. Specifically, the terminal can obtain a random number based on a random number seed in advance, and randomly initialize the model parameters of the trajectory prediction model to be trained based on the random number, and divide it into a training set and a validation set according to the date and time corresponding to the sample data to obtain a sample data set. Among them, the trajectory label reflects the actual path that other traffic participants should take in a specific environment, and is used to guide the training to generate predicted trajectories; the path selection label represents the path selected by the planning model under the state of the sample data, and is used to guide the generation of importance weights.

[0102] Step 304 : performing prediction processing on each sample data according to the trajectory prediction model to be trained to obtain a second predicted trajectory and a second importance weight corresponding to each second predicted trajectory.

[0103] The second predicted trajectory and the second importance weight are output results of the trajectory prediction model in the training phase.

[0104] In an embodiment of the present application, the terminal inputs the training set from the sample data set into the trajectory prediction model to be trained. The trajectory prediction model to be trained generates and outputs prediction results for each sample data, including a second predicted trajectory corresponding to each traffic participant contained in the sample data and a second importance weight corresponding to each second predicted trajectory. The second predicted trajectory currently output by the trajectory prediction model to be trained may differ from the trajectory label. At the same time, the second importance weight based on the current prediction output may cause the subsequent planning model to select a path other than the path selection label as the target path. Therefore, the model parameters need to be adjusted to optimize the accuracy of the output results of the trajectory prediction model to be trained.

[0105] Step 306 : determining an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and determining a path planning loss based on the second importance weight, the decision cost function, and the path selection label.

[0106] In the embodiment of the present application, for the trajectory prediction model to be trained The training process, in which is the historical trajectory of environmental objects in the sample data, is the policy cost function, is the second predicted trajectory output by the trajectory prediction model to be trained, It is the second importance weight output by the trajectory prediction model to be trained.

[0107] The terminal calculates the accuracy loss of the trajectory prediction model to be trained by comparing the second predicted trajectory with the actual trajectory label. Specifically, the predicted trajectory accuracy Calculate the accuracy prediction loss through the Evidence Lower Bound (ELBO) principle To constrain:

[0108]

[0109] Among them, X is the historical trajectory of the environment object, E is the target path of the ego vehicle, Y is the second predicted trajectory, and Z is the intermediate variable of the auxiliary ELBO loss. and The two distributions of the intermediate variable Z are obtained, where For the training process, For inference, these two distributions are fitted during training. For the decoder, Characterize the calculation of KL divergence, characterize so that and The distribution is as close as possible. is the mathematical expectation, which represents the Z-pair decoder with distributed sampling The impact of output.

[0110] For path planning loss, the terminal uses the second importance weight, the decision cost function, and the path selection label to comprehensively calculate the loss. The decision cost function is used to consider the actual cost of path selection and is combined with the path selection label to evaluate the performance of the trajectory prediction model to be trained in path selection.

[0111] In step 308 , the model parameters of the trajectory prediction model to be trained are updated through back propagation until the accuracy prediction loss and the path planning loss meet the preset loss conditions, thereby obtaining a trained trajectory prediction model.

[0112] In an embodiment of the present application, the training process of the trajectory prediction model to be trained updates the model parameters through back propagation, and verifies the path selection under the second predicted trajectory and the second importance weight through the verification set. If the accuracy prediction loss and the path planning loss still do not meet the preset loss conditions, the model parameters of the trajectory prediction model to be trained will continue to be adjusted until the accuracy prediction loss and the path planning loss meet the preset loss conditions. The trajectory prediction model to be trained is considered to be within an acceptable accuracy range, and a trained trajectory prediction model is obtained.

[0113] In this embodiment, the loss of the trajectory prediction model to be trained is calculated based on the prediction results using sample data and corresponding trajectory labels and path selection labels, and the model parameters of the trajectory prediction model are continuously iteratively updated to obtain a high-performance trajectory prediction model. In an autonomous vehicle, this training mechanism enables the trajectory prediction model to be trained to better understand the environment and the degree to which different predicted trajectories affect decision-making.

[0114] In a specific embodiment, Figure 4 As shown, an example of training a trajectory prediction model is provided, including:

[0115] Step 401: The trajectory prediction model to be trained is randomly initialized;

[0116] Step 402: Divide the sample data set into a training set and a validation set, and randomly shuffle the data arrangement;

[0117] Step 403 : The prediction capability of the trajectory prediction model to be trained is evaluated based on the scene trajectory true value in the validation set, and the decision capability is evaluated based on the scene trajectory true value and the decision cost function, thereby completing the training of the trajectory prediction model to be trained.

[0118] In an exemplary embodiment, Figure 5 As shown, step 306 includes steps 502 to 506. Among them:

[0119] Step 502: construct a historical sampling space based on the sample data set.

[0120] The historical sampling space includes multiple historical candidate paths determined based on sample data.

[0121] In the embodiment of the present application, the sample data includes the starting point and end point of the self-vehicle. When the terminal trains the trajectory prediction model to be trained, multiple historical candidate paths are constructed according to the same principle as step 102 to constitute the historical sampling space of the training process. That is, during the training process of the trajectory prediction model to be trained, it is necessary to guide the generation of importance weights in the selection of multiple historical candidate paths.

[0122] Step 504 : Among the historical candidate paths, the historical candidate path with the smallest distance to the path selection label is determined as the expert trajectory.

[0123] In this embodiment of the present application, the importance of the predicted trajectory to the future of the ego vehicle (the second importance weight) Constraints are implemented through planning perception. The specific training process is as follows: First, a historical sampling space is constructed based on the sample data set. , including candidate planning paths for the vehicle in different scenarios , The terminal uses the path selection label in the historical sampling space as the real future planning behavior of humans, and selects the path with the closest L2 distance to the path selection label in the historical sampling space as the security expert trajectory. :

[0124]

[0125] in, Characterizing candidate planning paths exist The location at the moment, Indicates the path corresponding to the path selection label.

[0126] Step 506 : Determine the path planning loss based on the expert trajectory and the decision cost function.

[0127] In the embodiment of the present application, the terminal uses the expert trajectory as the label trajectory to guide the second predicted trajectory generated by the trajectory prediction model to be trained and the second importance weight Generate a decision cost function Select the label trajectory, that is, supervise the planning model to maximize the probability of selecting the label trajectory, thereby obtaining the planning-related path planning loss :

[0128]

[0129]

[0130] Where, Refers to calculating the decision cost function Extensive information about the scene is required.

[0131] This completes the training of the trajectory prediction model. On the one hand, it supervises the task of predicting trajectory accuracy. On the other hand, from a planning perspective, it uses a data-driven approach to automatically select important modes of the trajectory and generate importance weights.

[0132] In this embodiment, by referring to the historical candidate paths corresponding to the sample data set and quantitatively evaluating the decision costs, the trajectory prediction model can be improved to generate importance weights that are more suitable for each predicted trajectory through more accurate decision selection, so that the predicted trajectories and importance weights generated by it can better guide the decision-making process of the planning model.

[0133] In an exemplary embodiment, Figure 6 As shown, before step 106, the method further includes steps 602 to 604. Among them:

[0134] Step 602: Acquire initial candidate paths, and screen the initial candidate paths according to a preset dynamic constraint sub-model to obtain pre-screened candidate paths.

[0135] In this embodiment of the present application, the planning model includes a collision safety constraint submodel, a dynamic constraint submodel, and a global path submodel. The global path submodel is used to generate an initial candidate path. The terminal is pre-configured with a preset collision safety constraint submodel. This preset collision safety constraint submodel is constructed based on road environment information, such as the shape, position, and road edge lines of traffic participants, taking into account driving safety issues, setting a minimum safe distance for object interaction, and constructing a collision prevention circular curve.

[0136] First, the terminal obtains a series of initial candidate paths from the global path sub-model. These paths may be generated based on factors such as map information, target location, and traffic regulations. These initial candidate paths are then screened using a preset dynamic constraint sub-model. The dynamic constraint sub-model considers the vehicle's dynamic characteristics, such as maximum control speed, acceleration, and maximum control angle, to ensure the feasibility of the initially screened candidate paths.

[0137] Step 604 : Perform a secondary screening on the initially screened candidate paths according to the preset collision safety constraint sub-model to obtain a first candidate path.

[0138] In this embodiment of the present application, the terminal also pre-configures a dynamic constraint sub-model constructed based on vehicle dynamics information and current state information, combined with vehicle limit variable constraint information (maximum control speed and acceleration, and maximum control angle). For each initially screened candidate path, the terminal uses a pre-set collision safety constraint sub-model to evaluate whether its distance from other objects in the road environment meets the minimum safety distance requirement and whether a collision is likely. This completes a secondary screening of the initially screened candidate paths, resulting in a first candidate path that meets both dynamic constraints and collision safety requirements, serving as the basis for subsequent path planning and decision-making.

[0139] In this embodiment, by presetting the collision safety constraint sub-model and the preset dynamic constraint sub-model, it can be ensured that the first candidate path finally obtained not only meets the requirements of global planning, but also takes into account the dynamic characteristics and driving safety of the vehicle. In addition, the preliminary screening according to the preset collision safety constraint sub-model and the preset dynamic constraint sub-model can reduce the amount of calculation compared to calculating the decision cost of each initial candidate path, thereby improving the computational efficiency of path planning.

[0140] In an exemplary embodiment, Figure 7As shown, the decision cost function includes a decision safety cost function, a decision efficiency cost function, and a decision comfort cost function; the decision cost includes a decision safety cost, a decision efficiency cost, and a decision comfort cost; and step 106 includes steps 702 to 706. Wherein:

[0141] Step 702 : determining a decision safety cost based on the initial candidate path, the first predicted trajectory, the first importance weight, and the decision safety cost function.

[0142] In the embodiment of the present application, the terminal first constructs a decision safety cost function to evaluate the collision risk between the first candidate path and the first predicted trajectory. , and the first predicted trajectory of a traffic participant , based on the basic idea of ​​keeping a safe distance from surrounding traffic participants, the decision safety cost function The calculation formula is:

[0143]

[0144] in, Indicates the number of multimodalities, representing the number of different traffic participants and exhibiting different behavior patterns (e.g., going straight, turning, accelerating, etc.), represents the planning time domain, Indicates the first importance weight generated by the trajectory prediction model. Characterized by time A traffic participant at the moment The first predicted trajectory of Characterizing the first predicted trajectory The corresponding first importance weight.

[0145] Step 704 : determining a decision efficiency cost based on the initial candidate path, the destination coordinates, and the decision efficiency cost function.

[0146] In the embodiment of the present application, for the decision efficiency cost function, the terminal accelerates the autonomous driving car to reach the destination The basic idea of ​​efficiency, decision efficiency cost function The calculation formula is:

[0147]

[0148] in, is the first candidate path in The location at the moment, For the destination.

[0149] Step 706 : Determine the decision comfort cost based on the initial candidate paths and the decision comfort cost function.

[0150] In the embodiment of the present application, for constructing a decision comfort cost function, the terminal is based on the basic idea of ​​ensuring the safety of the autonomous driving car and controlling the cost, and the decision comfort cost function is constructed. The calculation formula is:

[0151]

[0152] in, The distance traveled in the forward direction after the route is converted to the Frenet coordinate system.

[0153] This embodiment achieves a comprehensive evaluation of autonomous vehicle path planning by constructing decision-making safety, efficiency, and comfort cost functions. Overall, these three cost functions help the autonomous driving system balance safety, efficiency, and comfort during path planning, resulting in more optimal decisions. Furthermore, the decision-making cost functions provide unified constraints and design for the upstream trajectory prediction model and the downstream planning model, achieving consistency between upstream and downstream tasks, thereby improving planning effectiveness and ensuring a unified safety and utility across autonomous driving decision-making methods, thereby enhancing the safety and reliability of the autonomous driving system.

[0154] In an exemplary embodiment, Figure 8 As shown, step 602 includes steps 802 to 806. Among them:

[0155] Step 802: Obtain the preset path starting point and path end point.

[0156] In an embodiment of the present application, the terminal may use the current position of the vehicle as the starting point of the path and obtain the destination input by the user as the end point of the path.

[0157] Step 804: Generate a global path based on the path start point and the path end point.

[0158] In the embodiment of the present application, during the first planning, the terminal first generates a global planning path, and obtains a global path G connecting the path starting point and the path end point through a deep search method.

[0159] Step 806 : Generate multiple initial candidate paths for the target object according to the preset horizontal and vertical offset distances and the global path.

[0160] In the embodiment of the present application, each subsequent planning sampling local planning path wherein the specific sampling method is combined with the global path G, such as Figure 9 As shown in Figure 1, different horizontal and vertical offset distances are used to obtain candidate trajectory points using a quartic spline curve to form a trajectory sampling set. The trajectory sampling set contains multiple sampled candidate trajectories pointing to the target location point, forming the initial candidate path.

[0161] In this embodiment, multiple initial candidate paths are generated based on the global path according to preset lateral and longitudinal offset distances. The candidate paths provide abundant options for subsequent path screening and optimization, thereby improving the diversity and accuracy of path screening.

[0162] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0163] Based on the same inventive concept, the present application also provides an autonomous driving decision-making device for implementing the aforementioned autonomous driving decision-making method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations in one or more of the following autonomous driving decision-making device embodiments can be found in the aforementioned limitations on the autonomous driving decision-making method and will not be further elaborated here.

[0164] In an exemplary embodiment, Figure 9 As shown, an automatic driving decision-making device 900 is provided, comprising: a first acquisition module 901, a first prediction module 902 and a planning module 903, wherein:

[0165] A first acquisition module 901 is configured to acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object;

[0166] A first prediction module 902 is configured to determine, based on a trajectory prediction model and environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each first predicted trajectory; the trajectory prediction model is trained using a decision cost function;

[0167] The planning module 903 is configured to determine the decision cost corresponding to each first candidate path based on the first candidate path, the first predicted trajectory, the first importance weight, and the decision cost function, and determine the target path from the first candidate paths according to the decision cost.

[0168] In one embodiment, the apparatus further comprises:

[0169] The second acquisition module is used to obtain the sample data set and the trajectory label and path selection label corresponding to each sample data in the sample data set;

[0170] A second prediction module is used to perform prediction processing on each sample data according to the trajectory prediction model to be trained, and obtain a second predicted trajectory and a second importance weight corresponding to each second predicted trajectory;

[0171] a loss calculation module, configured to determine an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and to determine a path planning loss based on the second importance weight, the decision cost function, and the path selection label;

[0172] The training module is used to update the model parameters of the trajectory prediction model to be trained through back propagation until the accuracy prediction loss and path planning loss meet the preset loss threshold, thereby obtaining a trained trajectory prediction model.

[0173] In one embodiment, the loss calculation module is specifically configured to construct a historical sampling space based on the sample data set; the historical sampling space includes a plurality of historical candidate paths determined based on the sample data;

[0174] Among all the historical candidate paths, the historical candidate path with the smallest distance to the path selection label is determined as the expert trajectory;

[0175] Based on the expert trajectory and the decision cost function, the path planning loss is determined.

[0176] In one embodiment, the apparatus further comprises:

[0177] The first screening module is used to obtain initial candidate paths and screen the initial candidate paths according to a preset dynamic constraint sub-model to obtain preliminary screening candidate paths;

[0178] The second screening module is used to perform a secondary screening on the initially screened candidate paths according to a preset collision safety constraint sub-model to obtain a first candidate path.

[0179] In one embodiment, the decision cost function includes a decision safety cost function, a decision efficiency cost function, and a decision comfort cost function; the decision cost includes a decision safety cost, a decision efficiency cost, and a decision comfort cost;

[0180] The planning module 903 is specifically configured to determine a decision safety cost based on the initial candidate path, the first predicted trajectory, the first importance weight, and the decision safety cost function;

[0181] determining a decision efficiency cost based on the initial candidate paths, the destination coordinates, and the decision efficiency cost function;

[0182] The decision comfort cost is determined based on the initial candidate paths and the decision comfort cost function.

[0183] In one embodiment, the planning module 903 is specifically used to obtain a preset path starting point and path end point;

[0184] Generate a global path based on the path start point and path end point;

[0185] Generate multiple initial candidate paths for the target object based on preset horizontal and vertical offset distances and the global path.

[0186] Each module in the aforementioned autonomous driving decision-making device may be implemented in whole or in part through software, hardware, or a combination thereof. Each module may be embedded in or independent of a processor in a computer device in hardware form, or may be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0187] In an exemplary embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 10 As shown. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals via wired or wireless means, and the wireless means can be implemented via Wi-Fi, a mobile cellular network, near-field communication (NFC), or other technologies. When executed by the processor, the computer program implements an autonomous driving decision-making method. The display unit of the computer device is used to form a visually visible image, and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse.

[0188] Those skilled in the art will understand that Figure 10The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0189] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0190] Acquire environmental information of the environment in which the target object is located and a plurality of first candidate paths of the target object;

[0191] Determining, based on a trajectory prediction model and environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each first predicted trajectory; the trajectory prediction model is trained using a decision cost function;

[0192] Based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, a decision cost corresponding to each first candidate path is determined, and a target path is determined from the first candidate paths according to the decision cost.

[0193] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0194] Obtain the sample data set and the trajectory label and path selection label corresponding to each sample data in the sample data set;

[0195] Perform prediction processing on each sample data according to the trajectory prediction model to be trained to obtain a second predicted trajectory and a second importance weight corresponding to each second predicted trajectory;

[0196] determining an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and determining a path planning loss based on the second importance weight, the decision cost function, and the path selection label;

[0197] The model parameters of the trajectory prediction model to be trained are updated through back propagation until the accuracy prediction loss and path planning loss meet the preset loss threshold, and the trained trajectory prediction model is obtained.

[0198] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0199] Constructing a historical sampling space based on the sample data set; the historical sampling space includes multiple historical candidate paths determined based on the sample data;

[0200] Among all the historical candidate paths, the historical candidate path with the smallest distance to the path selection label is determined as the expert trajectory;

[0201] Based on the expert trajectory and the decision cost function, the path planning loss is determined.

[0202] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0203] Obtaining initial candidate paths, screening the initial candidate paths according to a preset dynamic constraint sub-model to obtain preliminary screening candidate paths;

[0204] The primary candidate paths are screened again according to the preset collision safety constraint sub-model to obtain the first candidate path.

[0205] In one embodiment, the decision cost function includes a decision safety cost function, a decision efficiency cost function, and a decision comfort cost function; the decision cost includes a decision safety cost, a decision efficiency cost, and a decision comfort cost; and when the processor executes the computer program, the processor further implements the following steps:

[0206] determining a decision safety cost based on the initial candidate path, the first predicted trajectory, the first importance weight, and a decision safety cost function;

[0207] determining a decision efficiency cost based on the initial candidate paths, the destination coordinates, and the decision efficiency cost function;

[0208] The decision comfort cost is determined based on the initial candidate paths and the decision comfort cost function.

[0209] In one embodiment, when the processor executes the computer program, the processor further implements the following steps:

[0210] Get the preset path starting point and path end point;

[0211] Generate a global path based on the path start point and path end point;

[0212] Generate multiple initial candidate paths for the target object based on preset horizontal and vertical offset distances and the global path.

[0213] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0214] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0215] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0216] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), quantum computing-based data processing logic devices, artificial intelligence (AI) processors, and the like.

[0217] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0218] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. An automatic driving decision-making method, characterized in that: The method comprises: Acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object; Determining, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model is trained using a decision cost function; determining a decision cost corresponding to each of the first candidate paths based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, and determining a target path among the first candidate paths according to the decision cost; Before acquiring the environmental information of the environment in which the target object is located and the plurality of first candidate paths of the target object, the method further includes: Obtaining a sample data set and a trajectory label and a path selection label corresponding to each sample data in the sample data set; Performing prediction processing on each of the sample data according to the trajectory prediction model to be trained to obtain a second predicted trajectory and a second importance weight corresponding to each of the second predicted trajectories; determining an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and determining a path planning loss based on the second importance weight, the decision cost function, and the path selection label; The model parameters of the trajectory prediction model to be trained are updated by back propagation until the accuracy prediction loss and the path planning loss meet a preset loss threshold, thereby obtaining a trained trajectory prediction model.

2. The method according to claim 1, characterized in that The determining of the path planning loss based on the second importance weight, the decision cost function, and the path selection label includes: Constructing a historical sampling space according to the sample data set; the historical sampling space includes a plurality of historical candidate paths determined based on the sample data; Among the historical candidate paths, determining the historical candidate path with the smallest distance to the path selection label as the expert trajectory; A path planning loss is determined based on the expert trajectory and the decision cost function.

3. The method according to claim 1, characterized in that Before determining the decision cost corresponding to each of the first candidate paths based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, the method further includes: Obtaining initial candidate paths, and screening the initial candidate paths according to a preset dynamic constraint sub-model to obtain pre-screened candidate paths; The initially screened candidate paths are screened a second time according to a preset collision safety constraint sub-model to obtain a first candidate path.

4. The method according to claim 1 or 3, characterized in that The decision cost function includes a decision safety cost function, a decision efficiency cost function and a decision comfort cost function; the decision cost includes a decision safety cost, a decision efficiency cost and a decision comfort cost; The determining, based on the first candidate paths, the first predicted trajectories, the first importance weights, and the decision cost function, of the decision costs corresponding to the first candidate paths includes: determining the decision safety cost based on the initial candidate path, the first predicted trajectory, the first importance weight, and the decision safety cost function; determining the decision efficiency cost based on the initial candidate path, the destination coordinates, and a decision efficiency cost function; The decision comfort cost is determined based on the initial candidate path and a decision comfort cost function.

5. The method according to claim 3, characterized in that The obtaining of the initial candidate path includes: Get the preset path starting point and path end point; generating a global path based on the path starting point and the path end point; A plurality of initial candidate paths for the target object are generated according to preset transverse and longitudinal offset distances and the global path.

6. An automatic driving decision-making device, characterized in that: The device comprises: A first acquisition module is used to acquire environmental information of an environment in which a target object is located and a plurality of first candidate paths of the target object; a first prediction module, configured to determine, based on a trajectory prediction model and the environmental information, a plurality of first predicted trajectories corresponding to each traffic participant in the current environment and a first importance weight corresponding to each of the first predicted trajectories; the trajectory prediction model being trained using a decision cost function; a planning module, configured to determine a decision cost corresponding to each of the first candidate paths based on the first candidate paths, the first predicted trajectory, the first importance weight, and the decision cost function, and determine a target path among the first candidate paths according to the decision cost; The device further comprises: A second acquisition module is used to acquire a sample data set and a trajectory label and a path selection label corresponding to each sample data in the sample data set; a second prediction module, configured to perform prediction processing on each of the sample data according to the trajectory prediction model to be trained, to obtain a second predicted trajectory and a second importance weight corresponding to each of the second predicted trajectories; a loss calculation module, configured to determine an accuracy prediction loss based on the trajectory label and the second predicted trajectory, and to determine a path planning loss based on the second importance weight, the decision cost function, and the path selection label; The training module is used to update the model parameters of the trajectory prediction model to be trained through back propagation until the accuracy prediction loss and the path planning loss meet a preset loss threshold, thereby obtaining a trained trajectory prediction model.

7. The device according to claim 6, characterized in that The loss calculation module is specifically used to construct a historical sampling space according to the sample data set; the historical sampling space includes a plurality of historical candidate paths determined based on the sample data; Among the historical candidate paths, determining the historical candidate path with the smallest distance to the path selection label as the expert trajectory; A path planning loss is determined based on the expert trajectory and the decision cost function.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images