Method for determining an autonomous driving trajectory and electronic device

Abstract

The application discloses a kind of automatic driving trajectory determination method and electronic equipment, it is related to automatic driving technical field, including after obtaining multiple predicted trajectories based on vehicle driving data and environmental semantic information trajectory prediction, multiple predicted trajectories and vehicle driving data, environmental semantic information input trajectory screening model, by trajectory screening model from multiple evaluation dimensions to multiple predicted trajectories Comprehensive evaluation is carried out, and candidate trajectory of evaluation score in front is output, it is realized to use trajectory screening model to combine actual automatic driving scene information to multiple predicted trajectories Comprehensive evaluation and screening out candidate trajectory of comprehensive performance in front, improve the environmental adaptability of candidate trajectory screened out, it can ensure that the target predicted trajectory finally determined matches actual driving environment, solve the technical problems that the accuracy and practical value of predicted trajectory determined in the related art are not high, reach the technical effects of improving trajectory prediction precision and practicality.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to a method for determining an autonomous driving trajectory and an electronic device. Background Technology

[0002] With the rapid development of autonomous driving technology, trajectory prediction, as the core perception and decision-making module of autonomous driving system, is directly related to driving safety and traffic efficiency in terms of prediction accuracy and reliability.

[0003] Currently, related technologies typically deploy multiple trajectory prediction models for parallel inference, and then select the optimal predicted trajectory by ranking or weighting the probability scores of the multiple predicted trajectories output by each model. However, probability scores only reflect the internal confidence of the trajectory prediction model in its own prediction results and cannot comprehensively assess the physical rationality, ride comfort, and safety redundancy of the trajectory in a real driving environment. Especially when the prediction results of multiple models differ significantly, simple score ranking often fails to select a high-quality trajectory that truly meets actual driving needs. Therefore, this method of determining the predicted trajectory cannot guarantee the accuracy and practical value of the trajectory. Summary of the Invention

[0004] This application provides a method and electronic device for determining the trajectory of an autonomous driving system, which at least solves the problem in related technologies that the optimal predicted trajectory is determined based on the probability scores of multiple predicted trajectories output by each trajectory prediction model, resulting in low accuracy and practical value of the determined predicted trajectory.

[0005] This application provides a method for determining the trajectory of an autonomous driving system, including: Acquire vehicle driving data and environmental semantic information of autonomous vehicles; Based on the vehicle driving data and the environmental semantic information, multiple different trajectory prediction models are used to predict the trajectory and obtain multiple predicted trajectories. The multiple predicted trajectories, the vehicle driving data, and the environmental semantic information are input into the trajectory filtering model, and at least one candidate trajectory is output by the trajectory filtering model. The trajectory filtering model is obtained by fine-tuning the visual language model using the training dataset of the autonomous driving scenario. The trajectory filtering model comprehensively evaluates the multiple predicted trajectories from multiple evaluation dimensions and outputs at least one candidate trajectory with the highest evaluation score. The evaluation score is used to characterize the comprehensive performance of the corresponding trajectory in the multiple evaluation dimensions. Based on the at least one candidate trajectory, a target predicted trajectory is determined so that the autonomous driving control system performs motion control based on the target predicted trajectory.

[0006] This application also provides an apparatus for determining the trajectory of an autonomous driving system, comprising: The information acquisition module is used to acquire vehicle driving data and environmental semantic information of autonomous vehicles; The trajectory prediction module is used to predict the trajectory based on the vehicle driving data and the environmental semantic information, using multiple different trajectory prediction models to obtain multiple predicted trajectories. The trajectory filtering module is used to input the multiple predicted trajectories, the vehicle driving data, and the environmental semantic information into the trajectory filtering model, and obtain at least one candidate trajectory output by the trajectory filtering model. The trajectory filtering model is obtained by fine-tuning the visual language model using the training dataset of the autonomous driving scenario. The trajectory filtering model comprehensively evaluates the multiple predicted trajectories from multiple evaluation dimensions and outputs at least one candidate trajectory with the highest evaluation score. The evaluation score is used to characterize the comprehensive performance of the corresponding trajectory in the multiple evaluation dimensions. The trajectory determination module is used to determine a target predicted trajectory based on the at least one candidate trajectory, so that the autonomous driving control system can perform motion control based on the target predicted trajectory.

[0007] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described methods for determining an autonomous driving trajectory.

[0008] This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of any of the above-described methods for determining an autonomous driving trajectory.

[0009] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described methods for determining an autonomous driving trajectory.

[0010] This application demonstrates that the trajectory selection model, obtained by fine-tuning the visual language model, can comprehensively evaluate multiple predicted trajectories from multiple evaluation dimensions and output at least one candidate trajectory with the highest evaluation score. After multiple different trajectory prediction models predict multiple trajectories based on vehicle driving data and environmental semantic information, these predicted trajectories, along with the vehicle driving data and environmental semantic information, are input into the trajectory selection model. The trajectory selection model then comprehensively evaluates these predicted trajectories from multiple evaluation dimensions based on the vehicle driving data and environmental semantic information, and outputs at least one candidate trajectory with the highest evaluation score. Thus, it achieves the goal of using the trajectory selection model in conjunction with actual autonomous driving scenario information to comprehensively evaluate multiple predicted trajectories from multiple evaluation dimensions and select the candidate trajectory with the highest overall performance. This method improves the environmental adaptability of the selected candidate trajectories and integrates a deep semantic understanding of traffic scenarios into the trajectory selection process. The decision-making basis is expanded from simple statistical probability to multiple considerations in the actual driving environment. Based on at least one candidate trajectory, the target predicted trajectory is determined for motion control of autonomous vehicles. This ensures that the final determined target predicted trajectory matches the actual driving environment, significantly improving the practical value of the trajectory. The target predicted trajectory has good environmental adaptability, accuracy, and practicality, thus improving the accuracy and practical value of trajectory prediction. Therefore, it can solve the technical problem in related technologies where determining the optimal predicted trajectory based on the probability scores of multiple predicted trajectories output by each trajectory prediction model results in low accuracy and practical value of the determined predicted trajectory. This method achieves the technical effect of improving the accuracy and practicality of trajectory prediction. Attached Figure Description

[0011] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic diagram of the application environment architecture of the autonomous driving trajectory determination method according to an exemplary embodiment of this application is shown; Figure 2 A flowchart illustrating an exemplary embodiment of this application for determining an autonomous driving trajectory; Figure 3 A flowchart illustrating a method for determining an autonomous driving trajectory, provided as another exemplary embodiment of this application; Figure 4 A flowchart illustrating a method for determining an autonomous driving trajectory, as provided in yet another exemplary embodiment of this application; Figure 5This is a schematic diagram of the structure of an autonomous driving trajectory determination device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0014] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0015] In critical fields such as autonomous driving and intelligent traffic management, the accuracy of trajectory prediction directly impacts system safety and decision-making efficiency. Among related technologies, autonomous driving trajectory prediction solutions primarily rely on deep learning models, analyzing information such as historical vehicle trajectories and environmental perception data to predict potential future paths. However, in practical applications, these solutions face the following key technical challenges: (1) Challenges in fusing prediction results from multiple models: Current mainstream solutions typically employ an ensemble learning approach, deploying multiple trajectory prediction models for parallel inference, and then selecting the optimal trajectory by ranking or weighting the probability scores output by each model. However, this probability score-based fusion method has significant limitations: the probability score only reflects the model's internal confidence in its own prediction results and cannot comprehensively assess the physical rationality, ride comfort, and safety redundancy of the trajectory in the actual driving environment. Especially when the prediction results from multiple models differ significantly, simple score ranking often fails to select a high-quality trajectory that truly meets the actual driving needs.

[0016] (2) Limitations of semantic understanding in end-to-end prediction: Traditional trajectory prediction models are mostly data-driven black-box models, lacking the ability to deeply understand the semantics of complex traffic scenarios. These models are unable to comprehensively consider multi-dimensional semantic factors such as traffic rule constraints, driver behavioral intentions, and multi-agent interaction behaviors, resulting in a significantly increased risk of prediction failure in edge scenarios (such as unprotected left turns, emergency avoidance, complex intersections, etc.), which seriously restricts the applicability of autonomous driving systems in real road environments.

[0017] (3) Insufficient adaptability to dynamic environments: Existing fusion strategies mostly adopt linear combinations of fixed weights or decision logic based on rule bases, which are difficult to adapt to complex and ever-changing driving environments. Different scenarios such as urban roads, highways, and rural roads have different requirements for trajectory preferences, and fixed-parameter fusion schemes cannot dynamically adjust the fusion strategy according to the characteristics of the actual scenario, resulting in poor performance of trajectory prediction results in specific scenarios and a lack of necessary environmental adaptability.

[0018] (4) Single evaluation dimension: The relevant technologies mainly focus on the position accuracy of the trajectory, ignoring important indicators such as trajectory smoothness, comfort, and energy efficiency, which cannot meet the comprehensive requirements of high-level autonomous driving systems for trajectory quality.

[0019] To address the aforementioned issues, this application provides an adaptive optimization algorithm for autonomous driving trajectory prediction based on the fusion of visual language model (VLM) selection and multi-model results from genetic algorithms. Specifically, addressing the technical bottlenecks of existing trajectory prediction schemes, such as low efficiency of multi-model fusion, insufficient semantic understanding capabilities, and poor environmental adaptability, this application proposes an innovative solution combining deep learning and evolutionary computation, which can significantly improve the prediction accuracy and driving safety of autonomous vehicles in complex traffic environments.

[0020] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] The specific application environment architecture or specific hardware architecture on which the method for determining the autonomous driving trajectory depends is described here.

[0022] Figure 1 The diagram illustrates the application environment architecture of an exemplary embodiment of the autonomous driving trajectory determination method of this application, as shown below. Figure 1As shown, the architecture includes three trajectory prediction models (Model 1, Model 2, and Model 3), a VLM model, a genetic algorithm fusion module, and an autonomous driving simulation platform. The three trajectory prediction models predict trajectories based on autonomous driving scenario data (including but not limited to vehicle driving data and environmental semantic information such as visual features extracted from environmental images) and output predicted trajectories. The VLM model comprehensively evaluates the predicted trajectories from multiple evaluation dimensions based on the autonomous driving scenario data, generating a comprehensive quality score of 0-100 for each predicted trajectory, completely replacing the traditional probability score ranking method. By setting dynamic thresholds or a Top-K selection strategy, the K highest-scoring high-quality trajectories are selected to enter the next stage of the fusion process, ensuring that the input to the genetic algorithm are all high-quality trajectories that have undergone preliminary screening. The genetic algorithm fusion module initializes the population, encoding the K high-quality trajectories as initial chromosomes. Through genetic operations such as selection, crossover, and mutation, it performs multi-generational evolution. In each generation, the quality of individuals is evaluated according to a fitness function, gradually improving the comprehensive performance of the fused trajectories. The autonomous driving simulation platform evaluates the quality of each trajectory and inputs it into the fitness function for individual quality evaluation. Based on the evaluation results, a Bayesian optimization algorithm is used to adjust the hyperparameter settings. This process can be iterated until the preset convergence condition is met. After multiple rounds of optimization, the trajectory with the best individual quality is finally found and sent to the trajectory tracking module of the autonomous driving control system. Combined with control algorithms such as Model Predictive Control (MPC) or Linaer Quadratic Regulator (LQR), the vehicle's precise trajectory tracking and motion control are completed.

[0023] It should be noted that, Figure 1 This is merely an example to illustrate the present application and should not be construed as limiting it. The example of inputting the K high-quality trajectories output by the VLM model into the genetic algorithm fusion module for processing is only an example. To ensure the speed and efficiency of trajectory determination and reduce computational overhead, the specific application environment architecture of this application may not include the genetic algorithm fusion module and the autonomous driving simulation platform. After the VLM model selects K high-quality trajectories, the target predicted trajectory can also be determined by random selection or by selecting the candidate trajectory with the highest comprehensive evaluation score.

[0024] This application provides a method for determining an autonomous driving trajectory. This method can be executed by an autonomous driving trajectory determination device provided in this application. The device can be implemented in software and / or hardware and can be integrated into an electronic device. The method is described in detail below with reference to its execution flow.

[0025] Figure 2A flowchart illustrating an exemplary embodiment of this application for determining an autonomous driving trajectory is shown below. Figure 2 As shown, the method for determining the autonomous driving trajectory may include the following steps: Step 101: Obtain vehicle driving data and environmental semantic information of the autonomous vehicle.

[0026] Vehicle driving data includes, but is not limited to, vehicle speed, acceleration, location information, and driving commands, which can be obtained through sensors and positioning systems on autonomous vehicles; environmental semantic information can include images captured by onboard cameras and visual features obtained by extracting features from the images.

[0027] Optionally, in order to ensure the accuracy of trajectory prediction as much as possible, multi-sensor data and environmental semantic information of autonomous vehicles can be acquired in real time, and preprocessing operations such as time synchronization, coordinate unification, and feature extraction can be performed to construct a complete environmental perception representation, and trajectory prediction and trajectory selection can be performed based on the environmental perception representation.

[0028] Step 102: Based on vehicle driving data and environmental semantic information, multiple different trajectory prediction models are used to predict trajectories and obtain multiple predicted trajectories.

[0029] The trajectory prediction model is pre-trained and can be a deep learning model. This model is trained using training samples from the autonomous driving field. Multiple different trajectory prediction models can be models with different network architectures and / or models trained using different training strategies.

[0030] In this embodiment, vehicle driving data and environmental semantic information are input into multiple pre-trained trajectory prediction models. These models then predict trajectories based on the vehicle driving data and environmental semantic information, outputting multiple predicted trajectories. Because different trajectory prediction models are used, the diversity of models can be leveraged to generate multiple differentiated predicted trajectories, covering a wider range of possible driving behavior spaces.

[0031] Step 103: Input multiple predicted trajectories, vehicle driving data and environmental semantic information into the trajectory filtering model, and obtain at least one candidate trajectory output by the trajectory filtering model. The trajectory filtering model is obtained by fine-tuning the visual language model using the training dataset of the autonomous driving scenario. The trajectory filtering model comprehensively evaluates the multiple predicted trajectories input from multiple evaluation dimensions and outputs at least one candidate trajectory with the highest evaluation score. The evaluation score is used to characterize the comprehensive performance of the corresponding trajectory in multiple evaluation dimensions.

[0032] The trajectory selection model is obtained by fine-tuning a visual language model using a training dataset of collected autonomous driving scenarios. For example, the visual language model can be the general visual language model Qwen2B. Based on large-scale autonomous driving scenario data, Qwen2B is fine-tuned to enable it to deeply understand traffic scenarios and accurately evaluate trajectory quality. This allows for comprehensive evaluation of input trajectories from multiple evaluation dimensions and the selection of high-quality candidate trajectories. The training dataset used for fine-tuning includes rich multimodal information, such as lane line geometry, traffic sign semantics, dynamic obstacle location distribution, historical trajectory sequences, and traffic light states, ensuring the model fully understands the scenario context. Fine-tuning refers to small-scale training on a pre-trained model for specific task objectives (downstream tasks) and task data (downstream data), achieving minor adjustments to the parameters of the pre-trained model, ultimately obtaining a model adapted to specific tasks and data.

[0033] In one alternative embodiment of this application, the visual language model can be fine-tuned directly using the collected training dataset to obtain a trained trajectory selection model.

[0034] In one optional embodiment of this application, when fine-tuning the visual language model to obtain a trajectory selection model, a trainable module can be added to the visual language model to obtain an initial model. For example, a lightweight fine-tuning method using Low-Rank Adaptation (LORA) can be used to add trainable low-rank matrix modules (i.e., trainable modules) to some layers of the visual language model to obtain the initial model. Next, the autonomous driving scenario data and corresponding multiple sample trajectories from the training dataset are input into the initial model for trajectory selection, and the first trajectory output by the initial model is obtained. The initial model comprehensively evaluates multiple sample trajectories from multiple evaluation dimensions and outputs the first trajectory. The first trajectory is at least one of the multiple sample trajectories. The autonomous driving scenario data includes sample vehicle driving data and sample environmental semantic information, and covers various weather conditions, light levels, and traffic densities. Next, based on the first trajectory and the standard trajectory corresponding to the autonomous driving scenario data from multiple sample trajectories, the loss value of the initial model is determined. The standard trajectory is the optimal trajectory corresponding to the autonomous driving scenario data. Other trajectories among the multiple sample trajectories besides the standard trajectory can be obtained by adjusting the standard trajectory. When calculating the loss value, if the first trajectory includes only one candidate trajectory, the loss value is calculated based on the difference between that candidate trajectory and the standard trajectory. If the first trajectory includes multiple candidate trajectories, a loss value can be calculated for the difference between each candidate trajectory and the standard trajectory, resulting in multiple loss values. The final loss value is then determined through weighted summation, averaging, etc., to determine whether the currently trained model has achieved the expected effect. Alternatively, the minimum or median of the multiple loss values ​​can be used as the final loss value; this application does not impose any restrictions on this. Next, the loss value is compared with a preset value to determine whether the currently trained model has achieved the expected effect. If the loss value is greater than the preset value, the network parameters of the trainable modules are adjusted and iterative training is performed until the obtained loss value is not less than the preset value or the number of training iterations reaches the preset number. Training is then complete, and the trajectory selection model is obtained. During training, only the network parameters of the trainable modules are adjusted, while the network parameters of the visual language model remain unchanged.

[0035] In this embodiment, an initial model is obtained by adding a trainable module to the visual language model. The initial model is then iteratively trained using the collected training dataset. During the training process, only the network parameters of the trainable model are adjusted, while the network parameters of the visual language model remain unchanged, until the loss value is not less than a preset value or the number of training iterations reaches a preset number. This yields a trajectory selection model. By adding a trainable module and adjusting only its network parameters, the visual language model can learn the trajectory selection task with very few additional parameters without changing its own parameters. This significantly reduces the number of parameters that need to be trained, substantially lowers the training overhead, and improves training speed and efficiency. Furthermore, the trajectory selection model obtained through training provides the conditions for selecting high-quality candidate trajectories from multiple predicted trajectories.

[0036] In one optional embodiment of this application, the sample trajectories in the training dataset can also be scored. The training model scores each input trajectory from multiple evaluation dimensions and outputs the trajectory's score in each evaluation dimension and the overall evaluation score, as well as the top k candidate trajectories with the highest overall evaluation scores.

[0037] In this embodiment, after obtaining multiple predicted trajectories, the vehicle driving data and environmental semantic information used to obtain the multiple predicted trajectories, along with the multiple predicted trajectories themselves, can be input into the trajectory selection model. The trajectory selection model then comprehensively evaluates the multiple predicted trajectories from multiple evaluation dimensions based on the vehicle driving data and environmental semantic information. During the internal analysis process, the trajectory selection model will comprehensively score each predicted trajectory from multiple evaluation dimensions and calculate an evaluation score. Finally, it will output at least one candidate trajectory with the highest evaluation score. The evaluation score is used to characterize the comprehensive performance of the corresponding trajectory in multiple evaluation dimensions. The higher the evaluation score, the better the comprehensive performance of the corresponding trajectory in multiple evaluation dimensions, and the higher the trajectory quality.

[0038] Step 104: Based on at least one candidate trajectory, determine the target predicted trajectory so that the autonomous driving control system can perform motion control based on the target predicted trajectory.

[0039] In this embodiment, after obtaining at least one candidate trajectory output by the trajectory screening model, the target predicted trajectory can be determined based on the at least one candidate trajectory. The autonomous driving system of the autonomous vehicle performs motion control based on the target predicted trajectory, so that the vehicle drives according to the template predicted trajectory and ensures driving safety.

[0040] As an example, if there is only one candidate trajectory, then that candidate trajectory is directly determined as the target predicted trajectory.

[0041] As an example, if there are two or more candidate trajectories for at least one candidate trajectory, one can be randomly selected as the target predicted trajectory. Alternatively, if the trajectory selection model outputs an evaluation score for each candidate trajectory along with the output of at least one candidate trajectory, the candidate trajectory with the highest evaluation score can be selected as the target predicted trajectory.

[0042] As an example, given at least one candidate trajectory, a genetic algorithm can be used to select the optimal trajectory as the target prediction trajectory through multiple rounds of optimization. Core parameters of the genetic algorithm include the number of generations (30-100), an adaptive crossover rate adjusted between 0.6 and 0.9, a mutation rate dynamically varying between 0.05 and 0.2 based on population diversity, and a fitness function to determine the fitness score of each generation of trajectories. During the genetic selection process, real-number encoding is used to encode the feature point sequence (including position coordinates, velocity, heading angle, and other state variables) of at least one candidate trajectory into chromosomes. Each chromosome represents a possible trajectory fusion scheme, and the encoding length is dynamically adjusted based on the prediction time domain and sampling frequency to ensure that the trajectory features are fully expressed without excessively increasing computational complexity. Next, an initial population is constructed based on at least one chromosome (i.e., the encoded candidate trajectory). The fitness score of each candidate trajectory in the initial population is determined based on the fitness function. Combining roulette wheel selection and elite retention strategies, individuals with high fitness have a greater chance of entering the next generation, while preventing the loss of superior genes. Genetic operations are performed on the determined next generation population. This can be achieved using a similarity-based multi-point crossover operator, which exchanges feature point sequences of different high-quality trajectories while maintaining trajectory continuity. The mutation probability is dynamically adjusted based on population diversity indicators, and minor perturbations are applied to the trajectory point coordinates to conform to vehicle dynamics constraints. This yields offspring trajectories, and their corresponding fitness scores are determined. The elite selection and genetic operations are repeated for multiple rounds of optimization until a preset convergence termination condition is met. For example, if the fitness improvement of the best individuals is less than 1-2% for 5-10 consecutive generations, or when the maximum number of generations is reached, the optimization process terminates. The trajectory with the highest fitness among the offspring trajectories at this point is then determined as the target predicted trajectory. By setting inter-generational improvement thresholds and the maximum number of generations, a balance between computational efficiency and optimization effectiveness is achieved. Using genetic algorithms to determine the target prediction trajectory can effectively explore potential high-quality regions in the trajectory solution space, overcome the limitations and biases of single models, and generate smoother, safer, and higher-quality trajectories that conform to human driving habits. Especially when multiple model predictions conflict, evolutionary search can find a compromise solution that balances the advantages of each. The fitness score of each trajectory can be determined based on its scores on multiple evaluation metrics. These metrics scores can be obtained by inputting the trajectory into an autonomous driving simulation platform for simulation evaluation; details can be found in the relevant descriptions in subsequent embodiments, and will not be elaborated here.

[0043] The autonomous driving trajectory determination method in this application embodiment, by fine-tuning the visual language model to obtain a trajectory selection model, can comprehensively evaluate multiple input predicted trajectories from multiple evaluation dimensions and output at least one candidate trajectory with the highest evaluation score. Thus, after multiple different trajectory prediction models predict multiple predicted trajectories based on vehicle driving data and environmental semantic information, these predicted trajectories, along with the vehicle driving data and environmental semantic information, are input into the trajectory selection model. The trajectory selection model then comprehensively evaluates these predicted trajectories from multiple evaluation dimensions based on the vehicle driving data and environmental semantic information, and outputs at least one candidate trajectory with the highest evaluation score. This achieves the goal of using the trajectory selection model in conjunction with actual autonomous driving scenario information to comprehensively evaluate multiple predicted trajectories from multiple evaluation dimensions and select the one with the highest overall performance. The candidate trajectories are improved, enhancing the environmental adaptability of the selected candidate trajectories. Furthermore, the trajectory selection process incorporates a deep semantic understanding of traffic scenarios, expanding the decision-making basis from simple statistical probability to multiple considerations in the actual driving environment. Based on at least one candidate trajectory, the target predicted trajectory is determined for motion control of the autonomous vehicle. This ensures that the final determined target predicted trajectory matches the actual driving environment, significantly improving the practical value of the trajectory. The target predicted trajectory possesses good environmental adaptability, accuracy, and practicality, thus improving the accuracy and practical value of trajectory prediction. Therefore, it can solve the technical problem in related technologies where determining the optimal predicted trajectory based on the probability scores of multiple predicted trajectories output by each trajectory prediction model results in low accuracy and practical value of the determined predicted trajectory, achieving the technical effect of improving trajectory prediction accuracy and practicality.

[0044] In one alternative embodiment of this application, such as Figure 3 As shown, based on the foregoing embodiments, step 104 may include the following sub-steps: Step 201: Determine the fusion trajectory based on at least one candidate trajectory.

[0045] The trajectory selection model may output one or more candidate trajectories. When multiple candidate trajectories are output, it would be computationally expensive to simulate each candidate trajectory to determine its score on each evaluation index, which would affect the efficiency of the final trajectory determination. Therefore, in this embodiment, a fused trajectory can be determined based on at least one candidate trajectory, and then the fused trajectory can be evaluated on the index scores of each evaluation index to determine whether it can be used as the target prediction trajectory.

[0046] As an example, if at least one candidate trajectory has only one trajectory, then that candidate trajectory is determined as the fused trajectory.

[0047] As an example, if at least one candidate trajectory has multiple trajectories, then the average of the multiple candidate trajectories is calculated to obtain the fused trajectory.

[0048] In this embodiment, when the trajectory screening model outputs multiple candidate trajectories, the fusion trajectory is obtained by averaging the multiple candidate trajectories. This ensures that the determined fusion trajectory contains the features of each candidate trajectory, which helps to find a target prediction trajectory that balances the advantages of each candidate trajectory. Furthermore, since only one fusion trajectory is determined, the subsequent evaluation of various index scores and trajectory adjustment by the autonomous driving simulation platform can be completed with less computation, reducing resource consumption and helping to quickly find a high-quality target prediction trajectory, thereby improving trajectory determination efficiency.

[0049] Step 202: Input the fused trajectory into the autonomous driving simulation platform for testing, and obtain the index scores of the fused trajectory on multiple evaluation indicators. The index score corresponding to an evaluation indicator is used to characterize the performance of the fused trajectory on the corresponding evaluation indicator.

[0050] Among them, the autonomous driving simulation platform can adopt a high-fidelity simulation platform such as CARLA 0.9.13, and construct a variety of driving scenarios, including urban main roads, residential roads, highway ramps, intersections without traffic lights, construction detour areas, school zones, sharp bends, and wet and slippery roads in rainy weather. Each scenario is subjected to 100-500 simulation tests to ensure the statistical reliability and significance of the evaluation results.

[0051] In this embodiment, the obtained fused trajectory is input into an autonomous driving simulation platform for testing. The platform evaluates the fused trajectory under multiple evaluation metrics, performs refined quantitative scoring based on the evaluation criteria of each metric, and outputs the fused trajectory's score for each metric. The score for each evaluation metric characterizes the performance of the fused trajectory on that metric; a higher score indicates better performance.

[0052] In one optional embodiment of this application, multiple evaluation metrics include lane keeping metrics, drivable area metrics, safe distance metrics, obstacle avoidance metrics, comfort metrics, and driving efficiency metrics. Among them, the lane keeping index score is determined based on the lateral deviation of each coordinate point on the trajectory from the lane centerline, with a score range of 0-100. This score is used to evaluate the road compliance of the trajectory; the smaller the lateral deviation, the higher the score. The drivable area index score is determined by detecting whether the trajectory exceeds the drivable area, such as road boundaries and shoulders. The score for drivable areas is 0 or 100. The trajectory scores 0 if it exceeds the drivable area and 100 if it does not. The safe distance index score is determined based on the distance between the trajectory and obstacles, with a score range of 0-100. The autonomous driving simulation model dynamically calculates the minimum distance between the trajectory and obstacles to evaluate whether it meets the safety threshold requirements and gives the corresponding index score. The trajectory scores 0 if it collides with an obstacle, and the larger the minimum distance between the trajectory and the obstacle, the higher the score. The obstacle avoidance index score is based on... The score for the avoidance performance index is determined by the probability of a collision while traveling along the trajectory, with a value of 0 or 100. The autonomous driving simulation model evaluates the avoidance performance based on indicators such as Time-to-Collision (TTC) and collision probability, and gives corresponding index scores. The higher the collision probability, the lower the score. The score for the comfort index is determined based on the acceleration of the trajectory, with a score range of 0-100. The autonomous driving simulation model evaluates comfort by analyzing dynamic indicators such as lateral acceleration and jerk (i.e., jerkiness, representing the rate of change of acceleration over time) of the trajectory, and gives corresponding index scores. Generally, the greater the lateral acceleration and jerk, the lower the index score. The score for the driving efficiency index is determined based on the estimated travel time of the trajectory, with a score range of 0-100. The autonomous driving simulation model gives the index score by evaluating efficiency indicators such as the estimated travel time of the trajectory and the number of stops. Generally, the longer the estimated travel time and the more stops, the lower the index score.

[0053] In this embodiment of the application, by setting multiple evaluation indicators and evaluation criteria for the scores of each indicator, the autonomous driving simulation platform can perform refined quantitative scoring of the input trajectory from multiple evaluation indicators, ensuring the accuracy of trajectory evaluation and improving the conditions for obtaining high-quality target prediction trajectories.

[0054] Step 203: Based on the index scores of the fusion trajectory on multiple evaluation indicators, determine the fitness score of the fusion trajectory. The fitness score is used to characterize the comprehensive performance of the fusion trajectory on multiple evaluation indicators.

[0055] In this embodiment, after determining the scores of the fusion trajectory on multiple evaluation metrics, the fitness score of the fusion trajectory can be determined based on the scores of each metric. The fitness score is used to characterize the overall performance of the fusion trajectory on multiple evaluation metrics. The higher the fitness score, the better the overall performance of the fusion trajectory on multiple evaluation metrics.

[0056] As an example, the mean, sum, etc. of the scores of each indicator can be used to determine the fitness score of the fusion trajectory.

[0057] As an example, the fitness score of the fused trajectory can be obtained by weighted summation of the scores of each indicator. This involves first obtaining the weight coefficients corresponding to multiple evaluation indicators, and then weighted summation of the weight coefficients and scores of the same evaluation indicator among these indicators to obtain the fitness score of the fused trajectory. For example, a fitness function that comprehensively considers multiple indicators such as safety, comfort, and efficiency can be designed. The fitness function is expressed as Fitness = w1 × LaneKeeping + w2 × DrivableArea + w3 × SafetyDistance + w4 × ObstacleAvoidance + w5 × ComfortScore + w6 × Efficiency, where w1 to w6 are adaptively adjustable weight coefficients, corresponding to the index scores of LaneKeeping, DrivableArea, SafetyDistance, ObstacleAvoidance, ComfortScore, and Efficiency, respectively. The initial values ​​of the weight coefficients are w1 = 0.25, w2 = 0.15, w3 = 0.25, w4 = 0.15, w5 = 0.1, and w6 = 0.1, and they are automatically adjusted according to the characteristics of the scenario during the optimization process.

[0058] In this embodiment, the fitness score of the fusion trajectory is obtained by weighting and summing multiple evaluation indicators according to their respective weight coefficients. The weight coefficients of each evaluation indicator can be flexibly set according to the importance of each evaluation indicator, which improves the flexibility of the scheme.

[0059] Step 204: If the fitness score is greater than or equal to the score threshold, the fused trajectory is determined as the target predicted trajectory.

[0060] The scoring threshold can be set according to actual needs, such as setting the scoring threshold to 85, 90, etc.

[0061] In this embodiment, after determining the fitness score of the fused trajectory, the fitness score can be compared with the score threshold. If the fitness score is greater than or equal to the score threshold, the fused trajectory is considered to perform well in terms of safety, comfort, efficiency, etc., and thus the fused trajectory is determined as the final target prediction trajectory for vehicle driving control.

[0062] Step 205: If the fitness score is less than the score threshold, adjust the fused trajectory based on at least one candidate trajectory to obtain the second trajectory.

[0063] In this embodiment, if the fitness score of the fused trajectory is less than the score threshold, it is considered that the fused trajectory does not perform well in terms of safety, comfort, efficiency, etc. In this case, the fused trajectory can be adjusted based on at least one candidate trajectory to obtain a new trajectory (referred to as the second trajectory for ease of description and distinction).

[0064] For example, when adjusting the fused trajectory based on at least one candidate trajectory, a candidate trajectory can be randomly selected, and some trajectory coordinate points on the candidate trajectory can be exchanged with the corresponding trajectory coordinate points on the fused trajectory to obtain a second trajectory.

[0065] For example, when adjusting the fused trajectory based on at least one candidate trajectory, the idea of ​​a genetic algorithm can be used. First, the fitness score of each candidate trajectory is determined. Combining roulette wheel selection and elite retention strategies, this ensures that individuals with high fitness have a greater chance of entering the next generation, while preventing the loss of superior genes. For instance, a candidate trajectory with a high fitness score can be selected for adjusting the fused trajectory. During adjustment, a multi-point crossover operator based on similarity can be used to exchange the feature point sequences of the fused trajectory and the candidate trajectory with a high fitness score, while maintaining trajectory continuity, to obtain the second trajectory. Optionally, after exchanging the feature point sequences of the fused trajectory and the candidate trajectory, the mutation probability can be dynamically adjusted according to the population diversity index. Small perturbations that conform to vehicle dynamics constraints can be applied to the coordinates of the trajectory points on the exchanged trajectory to obtain the second trajectory.

[0066] Step 206: Input the second trajectory into the autonomous driving simulation platform for testing, and obtain the performance scores of the autonomous driving simulation platform for the second trajectory on multiple evaluation indicators.

[0067] In this embodiment, after obtaining the second trajectory, the second trajectory is input into the autonomous driving simulation platform for testing. The autonomous driving simulation platform evaluates the second trajectory under multiple evaluation indicators, performs refined quantitative scoring based on the evaluation standards of each evaluation indicator, and outputs the indicator scores of the second trajectory under each evaluation indicator.

[0068] Step 207: Determine the fitness score of the second trajectory based on the index scores of the second trajectory on multiple evaluation metrics.

[0069] In this embodiment, after obtaining the index scores of the second trajectory on multiple evaluation metrics, the fitness score of the second trajectory can be further determined based on these index scores.

[0070] It should be noted that the specific implementation method for determining the fitness score of the second trajectory is the same as the method for determining the fitness score of the fused trajectory in the aforementioned steps, and will not be repeated here. Additionally, it should be noted that for the case where the fitness score is determined by weighted summation of the scores of each indicator, the weight coefficients corresponding to each evaluation indicator can be fixed or adaptively adjusted; this application does not impose any restrictions on this.

[0071] For example, after determining the fitness score of a new trajectory, if the fitness score is less than a score threshold, the weight coefficients corresponding to each evaluation index can be adjusted for use in the next fitness score calculation. Therefore, in one optional embodiment of this application, after determining the fitness score of a trajectory, if the fitness score is less than a score threshold, the weight coefficients corresponding to multiple evaluation indices are adjusted; the adjusted weight coefficients are used to calculate the fitness score of the new trajectory obtained from the next trajectory adjustment. For example, after determining the fitness score of the fused trajectory, if the fitness score is less than a score threshold, the weight coefficients of each evaluation index are adjusted. When determining the fitness score of the second trajectory, the adjusted weight coefficients are used to calculate the fitness score; if the fitness score of the second trajectory is also less than a score threshold, the weight coefficients of each evaluation index are further adjusted, and the adjusted weight coefficients are used to calculate the fitness score of the new trajectory obtained from the trajectory adjustment of the second trajectory. Optionally, the weight coefficients can be automatically adjusted using a Bayesian optimization algorithm. Furthermore, for cases where a genetic algorithm is used to determine the target prediction trajectory based on at least one candidate trajectory, when the fitness score is less than the score threshold, the Bayesian optimization algorithm can not only adjust the weight coefficients but also determine key hyperparameters such as crossover rate, mutation rate, and population size in the genetic algorithm. This forms a complete closed-loop system of perception-decision-evaluation-optimization, enabling the system to continuously adapt to new driving environments and scenario characteristics through online learning and incremental optimization.

[0072] In this embodiment, when the obtained fitness score is less than the score threshold, the weight coefficients corresponding to multiple evaluation indicators are adjusted, and the fitness score of the new trajectory obtained in the next trajectory adjustment is calculated using the adjusted weight coefficients. Thus, through the close integration of the autonomous driving simulation platform and the automatic optimization mechanism of weight coefficients, the dynamic adjustment of the weight coefficients of the evaluation indicators is realized, which significantly improves the adaptability and robustness of this solution in complex and ever-changing environments.

[0073] Step 208: If the fitness score of the second trajectory is greater than or equal to the score threshold, the second trajectory is determined as the target prediction trajectory.

[0074] In this embodiment, after determining the fitness score of the second trajectory, it is compared with the score threshold. If the fitness score is greater than or equal to the score threshold, the second trajectory is considered to perform well in terms of safety, comfort, efficiency, etc., and thus the second trajectory is determined as the final target prediction trajectory for vehicle driving control.

[0075] Step 209: If the fitness score of the second trajectory is less than the score threshold, the second trajectory is adjusted based on at least one candidate trajectory to obtain the third trajectory, and the fitness score of the third trajectory is determined. Based on the fitness score of the third trajectory, it is determined whether the third trajectory is determined as the target prediction trajectory.

[0076] In this embodiment, if the fitness score of the second trajectory is less than the score threshold, it is considered that the second trajectory does not perform well in terms of safety, comfort, efficiency, etc. In this case, the second trajectory can be adjusted based on at least one candidate trajectory to obtain a new trajectory (referred to as the third trajectory for ease of description and distinction). Next, the obtained third trajectory is input into the autonomous driving simulation platform for testing, and the index scores of each evaluation index output by the autonomous driving simulation platform for the third trajectory are obtained. Then, the fitness score of the third trajectory is determined based on the index scores. If its fitness score is greater than or equal to the score threshold, the third trajectory is determined as the target prediction trajectory; otherwise, the third trajectory is adjusted based on at least one candidate trajectory to obtain a new trajectory, and the above process is repeated until the target prediction trajectory is determined.

[0077] To avoid situations where the fitness score of the trajectory obtained after multiple trajectory adjustments is still less than the score threshold, thus delaying the determination of the target predicted trajectory, an adjustment count threshold can be set to represent the maximum number of trajectory adjustments allowed. When the trajectory adjustment count reaches this threshold, even if the fitness score of the latest adjusted trajectory is still less than the score threshold, no further trajectory adjustments are made, and the trajectory at this point is determined as the target predicted trajectory, ensuring timely planning of driving paths for autonomous vehicles. Therefore, in one optional embodiment of this application, before adjusting the second trajectory to obtain the third trajectory when the fitness score of the second trajectory is less than the score threshold, the current trajectory adjustment count can be obtained. The initial value of the trajectory adjustment count is 0, and it is incremented by one each time a new trajectory is obtained. That is, when it is determined that the fitness score of the second trajectory is less than the score threshold, the current trajectory adjustment count is 1. This count is compared with the adjustment count threshold. If the current trajectory adjustment count is less than the adjustment count threshold, the second trajectory is adjusted based on at least one candidate trajectory to obtain the third trajectory. Simultaneously, the trajectory adjustment count is incremented by 1, making the current trajectory adjustment count 2. If the fitness score of the second trajectory is determined to be less than the score threshold, and the current number of trajectory adjustments equals the adjustment threshold, then the second trajectory is identified as the target predicted trajectory. Similarly, after obtaining the third trajectory, if the fitness score of the third trajectory is less than the score threshold, the current number of trajectory adjustments is compared with the adjustment threshold. If it is not less than the adjustment threshold, the third trajectory is not adjusted further and is directly identified as the target predicted trajectory; if it is less than the adjustment threshold, the third trajectory is adjusted based on at least one candidate trajectory to obtain the fourth trajectory. At the same time, the trajectory adjustment count is incremented by 1, making the trajectory adjustment count 3. This process is repeated until the number of trajectory adjustments is not less than the adjustment threshold, or the obtained fitness score is greater than or equal to the score threshold, at which point the target predicted trajectory is determined. Therefore, by setting an adjustment number threshold, before each trajectory adjustment to obtain a new trajectory, it is first determined whether the current trajectory adjustment number is less than the adjustment number threshold. Only when the current trajectory adjustment number is less than the adjustment number threshold is trajectory adjustment performed to obtain a new trajectory. If the current trajectory adjustment number is equal to the adjustment number threshold, trajectory adjustment is no longer performed, and the current trajectory is directly determined as the target prediction trajectory. This can avoid the situation where the fitness score of the trajectory obtained after multiple trajectory adjustments is still less than the score threshold, and the target prediction trajectory cannot be determined for a long time. It avoids the situation of infinite loop and failure to converge, ensuring the real-time performance and response efficiency of trajectory prediction, as well as the interpretability and determinism of the solution.

[0078] The autonomous driving trajectory determination method of this application embodiment first determines a fused trajectory based on at least one candidate trajectory, inputs the fused trajectory into an autonomous driving simulation platform for testing, and the autonomous driving simulation platform scores the fused trajectory based on multiple evaluation indicators and outputs the indicator scores for each evaluation indicator. Based on the indicator scores of the fused trajectory on multiple evaluation indicators, a fitness score is determined for the fused trajectory. If the fitness score is greater than or equal to a score threshold, the fused trajectory is determined as the target prediction trajectory. This ensures the accuracy of the determined target trajectory. By determining a single fused trajectory, the target prediction trajectory can be quickly selected using fewer computational resources, improving trajectory determination efficiency. If the fitness score of the fused trajectory is less than the score threshold, the fused trajectory is adjusted based on at least one candidate trajectory to obtain a second trajectory, and then... The fitness score of the second trajectory is determined. If the fitness score of the second trajectory is greater than or equal to a score threshold, the second trajectory is identified as the target prediction trajectory. If the fitness score of the second trajectory is less than the score threshold, the second trajectory is adjusted based on at least one candidate trajectory to obtain a third trajectory, and the fitness score of the third trajectory is determined. The fitness score of the third trajectory is used to determine whether to identify the third trajectory as the target prediction trajectory. Thus, for trajectories with low fitness scores, trajectory adjustment is performed based on at least one candidate trajectory to integrate the features of the candidate trajectory into the new trajectory. The quality of the new trajectory is verified through an autonomous driving simulation platform. Through multiple optimizations, an optimal trajectory that balances the advantages of various dimensions can be found, generating a smoother, safer trajectory that conforms to human driving habits, ensuring the accuracy and practicality of the final determined trajectory. This solution, through a multi-dimensional comprehensive evaluation system and closed-loop optimization mechanism, ensures special handling capabilities for extreme or rare traffic scenarios (such as emergency braking, sudden obstacles, complex interactions, etc.), significantly reducing the probability of prediction failure in these high-risk scenarios.

[0079] In one alternative embodiment of this application, such as Figure 4 As shown, based on the foregoing embodiments, step 103 may include the following sub-steps: Step 301: Obtain the prompt text corresponding to the trajectory selection model. The prompt text includes descriptions of multiple evaluation dimensions, which are used to indicate the evaluation rules of the trajectory selection model in multiple evaluation dimensions. The prompt text is used to instruct the trajectory selection model to comprehensively evaluate the input candidate trajectories from multiple evaluation dimensions and output at least one candidate trajectory with the highest evaluation score.

[0080] In this embodiment, the prompt text of the trajectory selection model is a high-quality prompt text optimized during the fine-tuning of the visual language model. The prompt text includes descriptive information for multiple evaluation dimensions, indicating the evaluation rules of the trajectory selection model across these dimensions. The prompt text instructs the trajectory selection model to comprehensively evaluate the input predicted trajectory from multiple evaluation dimensions and output at least one candidate trajectory with the highest evaluation score. For example, in addition to descriptive information for multiple evaluation dimensions, the prompt text may also include task description information. This task description information indicates the task the trajectory selection model needs to complete. For instance, the task description information could be, "Please comprehensively evaluate each input trajectory from the following evaluation dimensions, determine the comprehensive evaluation score of each trajectory, and output the top 3 trajectories with the highest evaluation scores," thus instructing the trajectory selection model to comprehensively evaluate the input trajectory from multiple evaluation dimensions and output the top 3 candidate trajectories with the highest evaluation scores.

[0081] In one optional embodiment of this application, the descriptive information of multiple evaluation dimensions includes: the conformity between the trajectory and the lane structure, used to assess the consistency between the predicted trajectory and road geometric constraints, including lane centerline deviation and curvature continuity; the reasonableness of the safe distance between the trajectory and surrounding dynamic obstacles, assessed by calculating the distance between each point of the predicted trajectory and dynamic obstacles, such as adjacent vehicles and pedestrians; the smoothness of the trajectory and ride comfort, assessed by analyzing the rate of curvature change and acceleration continuity of the predicted trajectory to ensure a comfortable ride; the degree of compliance with traffic rules, assessed by checking whether the predicted trajectory complies with traffic rule constraints, such as traffic signals, signs, and markings; and the expected driving efficiency, assessed by analyzing the trajectory's travel time and energy efficiency. By setting multiple evaluation dimensions, the trajectory screening model performs preliminary screening of predicted trajectories from multiple evaluation dimensions, expanding the decision-making basis from simple statistical probability to multiple considerations such as safety, comfort, and compliance in the actual driving environment, significantly improving the practical value of the trajectory.

[0082] Step 302: Input the prompt text, multiple predicted trajectories, vehicle driving data, and environmental semantic information into the trajectory filtering model, and obtain at least one candidate trajectory selected by the trajectory filtering model from multiple predicted trajectories based on vehicle driving data and environmental semantic information under the guidance of the prompt text.

[0083] In this embodiment, multiple predicted trajectories, vehicle driving data, and environmental semantic information are input into the trajectory filtering model. Under the guidance of the prompt text, the trajectory filtering model combines vehicle driving data and environmental semantic information to comprehensively evaluate the multiple predicted trajectories from multiple evaluation dimensions and outputs at least one candidate trajectory with the highest evaluation score.

[0084] The autonomous driving trajectory determination method of this application embodiment obtains prompt text, which includes descriptive information of multiple evaluation dimensions to instruct the trajectory selection model on the evaluation rules of multiple evaluation dimensions. The prompt text, multiple predicted trajectories, vehicle driving data, and environmental semantic information are input into the trajectory selection model. Under the instruction of the prompt text, the trajectory selection model selects at least one candidate trajectory from multiple predicted trajectories based on vehicle driving data and environmental semantic information and outputs it. Thus, the trajectory selection criteria are expanded from simple statistical probability to multiple considerations such as safety, comfort, and compliance in the actual driving environment, which greatly improves the accuracy and practical value of the trajectory.

[0085] In summary, the proposed solution, by cleverly combining the semantic understanding capabilities of visual language models with the global search advantages of genetic algorithms, effectively solves key technical challenges in autonomous driving trajectory prediction, such as low efficiency of multi-model fusion and poor scene adaptability. This provides strong technical support for the safe and reliable deployment of high-level autonomous driving systems. Experimental results show that, under the same test conditions, compared to the traditional method of determining the final trajectory based on probability scores, this solution improves trajectory safety by approximately 23%, comfort by approximately 18%, and prediction success rate by approximately 31% in complex scenarios, demonstrating significant technical advantages and practical value.

[0086] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0087] Embodiments of this application also provide an autonomous driving trajectory determination device, which can be implemented in software and / or hardware and can be integrated into an electronic device.

[0088] Figure 5 This is a schematic diagram of the structure of an autonomous driving trajectory determination device provided in an embodiment of this application, as shown below. Figure 5 As shown, the autonomous driving trajectory determination device 50 includes: an information acquisition module 510, a trajectory prediction module 520, a trajectory filtering module 530, and a trajectory determination module 540.

[0089] Among them, the information acquisition module 510 is used to acquire vehicle driving data and environmental semantic information of autonomous vehicles; The trajectory prediction module 520 is used to predict trajectories based on vehicle driving data and environmental semantic information, using multiple different trajectory prediction models to obtain multiple predicted trajectories. The trajectory filtering module 530 is used to input multiple predicted trajectories, vehicle driving data and environmental semantic information into the trajectory filtering model, and obtain at least one candidate trajectory output by the trajectory filtering model. The trajectory filtering model is obtained by fine-tuning the visual language model using the training dataset of the autonomous driving scenario. The trajectory filtering model comprehensively evaluates the multiple predicted trajectories input from multiple evaluation dimensions and outputs at least one candidate trajectory with the highest evaluation score. The evaluation score is used to characterize the comprehensive performance of the corresponding trajectory in multiple evaluation dimensions. The trajectory determination module 540 is used to determine the target predicted trajectory based on at least one candidate trajectory, so that the autonomous driving control system can perform motion control based on the target predicted trajectory.

[0090] Optionally, the trajectory determination module 540 is also used for: The fusion trajectory is determined based on at least one candidate trajectory; The fused trajectory is input into the autonomous driving simulation platform for testing, and the scores of the fused trajectory on multiple evaluation indicators are obtained. The score of an evaluation indicator is used to characterize the performance of the fused trajectory on the corresponding evaluation indicator. Based on the scores of the fusion trajectory on multiple evaluation metrics, the fitness score of the fusion trajectory is determined. The fitness score is used to characterize the comprehensive performance of the fusion trajectory on multiple evaluation metrics. If the fitness score is greater than or equal to the score threshold, the fused trajectory is determined as the target predicted trajectory.

[0091] Further optionally, the trajectory determination module 540 is also used for: If the fitness score is less than the score threshold, the fused trajectory is adjusted based on at least one candidate trajectory to obtain a second trajectory; The second trajectory is input into the autonomous driving simulation platform for testing, and the scores of the autonomous driving simulation platform for the second trajectory on multiple evaluation indicators are obtained. The fitness score of the second trajectory is determined based on the scores of the second trajectory on multiple evaluation metrics. If the fitness score of the second trajectory is greater than or equal to the score threshold, the second trajectory is determined as the target prediction trajectory. If the fitness score of the second trajectory is less than the score threshold, the second trajectory is adjusted based on at least one candidate trajectory to obtain a third trajectory, and the fitness score of the third trajectory is determined. Based on the fitness score of the third trajectory, it is determined whether the third trajectory is identified as the target prediction trajectory.

[0092] Further optionally, the trajectory determination module 540 is also used for: Get the current number of trajectory adjustments. The trajectory adjustment count is incremented by one each time a new trajectory is obtained through adjustment. If the number of trajectory adjustments is less than the adjustment threshold, the second trajectory is adjusted based on at least one candidate trajectory to obtain the third trajectory; If the number of trajectory adjustments equals the adjustment threshold, the second trajectory is determined as the target predicted trajectory.

[0093] For a description of the features in the embodiment corresponding to the autonomous driving trajectory determination device, please refer to the relevant description of the embodiment corresponding to the autonomous driving trajectory determination method, which will not be repeated here.

[0094] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above embodiments of the method for determining an autonomous driving trajectory.

[0095] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above embodiments of the method for determining an autonomous driving trajectory when it is run.

[0096] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0097] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the method for determining an autonomous driving trajectory.

[0098] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above embodiments of the method for determining an autonomous driving trajectory.

[0099] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be executed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), application-specific standard parts (ASSP), a system-on-chip (SoC), a complex programmable logic device (CPLD), a microcontroller unit (MCU), etc. The terms "system," "computing device," or "apparatus" as used herein encompass various means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0100] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0101] The foregoing has provided a detailed description of the method for determining an autonomous driving trajectory, the electronic device, the storage medium, and the program product provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only intended to help understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A method for determining the trajectory of an autonomous driving system, characterized in that, include: Acquire vehicle driving data and environmental semantic information of autonomous vehicles; Based on the vehicle driving data and the environmental semantic information, multiple different trajectory prediction models are used to predict the trajectory and obtain multiple predicted trajectories. The multiple predicted trajectories, the vehicle driving data, and the environmental semantic information are input into the trajectory filtering model, and at least one candidate trajectory is output by the trajectory filtering model. The trajectory filtering model is obtained by fine-tuning the visual language model using the training dataset of the autonomous driving scenario. The trajectory filtering model comprehensively evaluates the multiple predicted trajectories from multiple evaluation dimensions and outputs at least one candidate trajectory with the highest evaluation score. The evaluation score is used to characterize the comprehensive performance of the corresponding trajectory in the multiple evaluation dimensions. Based on the at least one candidate trajectory, the target predicted trajectory is determined.

2. The method for determining the trajectory of an autonomous driving system according to claim 1, characterized in that, The process of fine-tuning the visual language model to obtain the trajectory selection model includes: A trainable module is added to the visual language model to obtain an initial model; The autonomous driving scenario data and corresponding multiple sample trajectories in the training dataset are input into the initial model for trajectory filtering, and the first trajectory output by the initial model is obtained. The initial model comprehensively evaluates the multiple sample trajectories from multiple evaluation dimensions and outputs the first trajectory. The first trajectory is at least one of the multiple sample trajectories. The autonomous driving scenario data includes sample vehicle driving data and sample environmental semantic information. Based on the first trajectory and the standard trajectory corresponding to the autonomous driving scenario data among the multiple sample trajectories, the loss value of the initial model is determined; If the loss value is greater than the preset value, the network parameters of the trainable module are adjusted and iterative training is performed until the loss value is not less than the preset value or the number of training iterations reaches the preset number. The training is then completed, and the trajectory selection model is obtained.

3. The method for determining the trajectory of an autonomous driving system according to claim 1, characterized in that, Determining the target predicted trajectory based on the at least one candidate trajectory includes: Based on the at least one candidate trajectory, a fusion trajectory is determined; The fused trajectory is input into an autonomous driving simulation platform for testing, and the scores of the fused trajectory on multiple evaluation indicators are obtained from the autonomous driving simulation platform. The score of an evaluation indicator is used to characterize the performance of the fused trajectory on the corresponding evaluation indicator. Based on the index scores of the fusion trajectory on multiple evaluation indicators, the fitness score of the fusion trajectory is determined, and the fitness score is used to characterize the comprehensive performance of the fusion trajectory on the multiple evaluation indicators. If the fitness score is greater than or equal to the score threshold, the fused trajectory is determined as the target predicted trajectory.

4. The method for determining the trajectory of an autonomous driving system according to claim 3, characterized in that, The process of determining the fitness score of the fusion trajectory based on its scores across multiple evaluation metrics includes: Obtain the weight coefficients corresponding to the multiple evaluation indicators; The fitness score of the fusion trajectory is obtained by weighting and summing the weight coefficients and scores of the same evaluation index among the multiple evaluation indicators.

5. The method for determining the trajectory of an autonomous driving system according to claim 3, characterized in that, After determining the fitness score of the fusion trajectory, the method further includes: If the fitness score is less than the score threshold, the fused trajectory is adjusted based on the at least one candidate trajectory to obtain a second trajectory; The second trajectory is input into the autonomous driving simulation platform for testing, and the autonomous driving simulation platform obtains the index scores of the second trajectory on the multiple evaluation indicators. The fitness score of the second trajectory is determined based on the scores of the second trajectory on multiple evaluation metrics. If the fitness score of the second trajectory is greater than or equal to the score threshold, the second trajectory is determined as the target predicted trajectory; If the fitness score of the second trajectory is less than the score threshold, the second trajectory is adjusted based on the at least one candidate trajectory to obtain a third trajectory, and the fitness score of the third trajectory is determined, so as to determine whether the third trajectory is determined as the target prediction trajectory based on the fitness score of the third trajectory.

6. The method for determining the trajectory of an autonomous driving system according to claim 5, characterized in that, The step of adjusting the second trajectory based on the at least one candidate trajectory to obtain the third trajectory includes: Get the current number of trajectory adjustments, which is incremented by one each time a new trajectory is obtained through adjustment; If the number of trajectory adjustments is less than the adjustment number threshold, the second trajectory is adjusted based on the at least one candidate trajectory to obtain the third trajectory.

7. The method for determining the trajectory of an autonomous driving system according to claim 6, characterized in that, The method further includes: If the number of trajectory adjustments equals the adjustment number threshold, the second trajectory is determined as the target predicted trajectory.

8. The method for determining the trajectory of an autonomous driving system according to claim 5, characterized in that, The method further includes: If the fitness score is less than the score threshold, adjust the weight coefficients corresponding to the multiple evaluation indicators. The adjusted weighting coefficients are used to calculate the fitness score of the new trajectory obtained after the next trajectory adjustment.

9. The method for determining the trajectory of an autonomous driving system according to claim 3, characterized in that, Determining the fused trajectory based on the at least one candidate trajectory includes: If the number of trajectories in the at least one candidate trajectory is one, the candidate trajectory is determined as the fused trajectory; When there are multiple candidate trajectories, the average value of the multiple candidate trajectories is calculated to obtain the fused trajectory.

10. The method for determining the trajectory of an autonomous driving system according to claim 3, characterized in that, The multiple evaluation indicators include lane keeping indicators, drivable area indicators, safe distance indicators, obstacle avoidance indicators, comfort indicators, and driving efficiency indicators; The lane keeping index score is determined based on the lateral deviation between each coordinate point on the trajectory and the lane centerline. The score of the drivable area indicator is determined by detecting whether the trajectory exceeds the non-drivable area; The score of the safety distance indicator is determined based on the distance between the trajectory and the obstacle; The obstacle avoidance index score is determined based on the probability of a collision occurring while traveling along the trajectory; The comfort index score is determined based on the acceleration of the trajectory; The score of the driving efficiency index is determined based on the estimated travel time of the trajectory.

11. The method for determining the trajectory of an automated driving system according to any one of claims 1-10, characterized in that, The step of inputting the multiple predicted trajectories, the vehicle driving data, and the environmental semantic information into the trajectory filtering model, and obtaining at least one candidate trajectory output by the trajectory filtering model, includes: Obtain the prompt text corresponding to the trajectory filtering model. The prompt text includes descriptive information of the multiple evaluation dimensions and is used to indicate the evaluation rules of the trajectory filtering model in the multiple evaluation dimensions. The prompt text is used to instruct the trajectory filtering model to comprehensively evaluate the input predicted trajectory from the multiple evaluation dimensions and output at least one candidate trajectory with the highest evaluation score. The prompt text, the multiple predicted trajectories, the vehicle driving data, and the environmental semantic information are input into the trajectory filtering model, and the at least one candidate trajectory is selected by the trajectory filtering model from the multiple predicted trajectories based on the vehicle driving data and the environmental semantic information under the instruction of the prompt text.

12. The method for determining the trajectory of an autonomous driving system according to claim 11, characterized in that, The descriptive information for the multiple evaluation dimensions includes: The conformity between the trajectory and the lane structure is used to evaluate the consistency between the predicted trajectory and the road geometric constraints; The reasonableness of the safe distance between the trajectory and surrounding dynamic obstacles is assessed by calculating the distance between each point of the predicted trajectory and the dynamic obstacles to evaluate the safety redundancy; Track smoothness and ride comfort are evaluated by analyzing the rate of change of curvature and acceleration continuity of the predicted trajectory. The degree of compliance with traffic rules is assessed by checking whether the predicted trajectory conforms to traffic rule constraints; Expected driving efficiency is assessed by analyzing the travel time and energy efficiency of the trajectory.

13. The method according to claim 1, characterized in that, Determining the target predicted trajectory based on the at least one candidate trajectory includes: If the at least one candidate trajectory is one, then the corresponding candidate trajectory is taken as the target predicted trajectory; If there are multiple candidate trajectories, one of them is randomly selected as the target predicted trajectory.

14. The method according to claim 1, characterized in that, Determining the target predicted trajectory based on the at least one candidate trajectory includes: The feature point sequence of at least one candidate trajectory is encoded with real numbers to obtain multiple chromosomes; An initial population is constructed based on at least one of the multiple chromosomes; The genetic operation is repeated to perform multiple rounds of iterative optimization on the initial population until the preset convergence termination condition is met. The candidate trajectory corresponding to the chromosome with the best fitness score in the last iteration is determined as the target predicted trajectory. The fitness score is used to characterize the comprehensive performance of the corresponding candidate trajectory on multiple evaluation indicators.

15. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the method for determining an autonomous driving trajectory as described in any one of claims 1 to 14 when executing the computer program.

Images