Driving trajectory prediction control method and system for unmanned vehicle

Abstract

The application discloses a driving trajectory prediction control method and system for unmanned vehicles, and relates to the technical field of vehicle trajectory control. The method comprises the following steps: collecting real-time driving scene data through a multi-source heterogeneous sensor network to obtain multi-modal driving data; inputting the multi-modal driving data into a hybrid prediction network for trajectory prediction to obtain a multi-modal prediction trajectory set; performing prediction control optimization on the multi-modal prediction trajectory set to construct a target prediction control optimization data set; solving a target optimization control sequence based on the target prediction control optimization data set, performing trajectory tracking control on the unmanned vehicle according to the target optimization control sequence, and generating a driving trajectory control result. The technical problem of low trajectory control accuracy and driving stability of the unmanned vehicle in a complex road environment in the prior art is solved, and the technical effect of improving the trajectory control precision and driving stability of the unmanned vehicle is achieved.

Description

Technical Field

[0001] This invention relates to the field of vehicle trajectory control technology, and more specifically to a driving trajectory prediction and control method and system for unmanned vehicles. Background Technology

[0002] With the continuous development of intelligent transportation and autonomous driving technologies, driverless vehicles have been gradually applied to various scenarios such as urban roads, highways, and industrial park logistics. During the operation of driverless vehicles, the vehicle needs to generate a corresponding driving trajectory based on road structure, surrounding vehicles, pedestrian status, and real-time traffic environment, and complete vehicle control according to the planned trajectory. Therefore, trajectory prediction and trajectory control are key aspects of driverless systems. Most existing trajectory control methods for driverless vehicles are based on fixed rules or single-scenario models for trajectory planning and vehicle control. In complex road environments and when traffic participants change rapidly, problems such as large trajectory prediction deviations, delayed vehicle control response, and insufficient trajectory tracking stability easily arise, thus affecting the smoothness and safety of vehicle operation. Furthermore, some existing control methods struggle to balance the coordination between trajectory prediction results and real-time vehicle control when facing complex operating conditions, leading to trajectory deviations or control oscillations during steering, obstacle avoidance, or dynamic following. Summary of the Invention

[0003] This application provides a driving trajectory prediction and control method and system for autonomous vehicles, which solves the technical problem of low trajectory control accuracy and driving stability of autonomous vehicles in complex road environments in the prior art.

[0004] The first aspect of this application provides a driving trajectory prediction and control method for an autonomous vehicle, the method comprising:

[0005] Real-time driving scenario data is collected through a multi-source heterogeneous sensor network to obtain multimodal driving data; the multimodal driving data is input into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set; predictive control optimization is performed on the multimodal predicted trajectory set to construct a target predictive control optimization dataset; a target optimization control sequence is solved based on the target predictive control optimization dataset, and trajectory tracking control is performed on the autonomous vehicle according to the target optimization control sequence to generate driving trajectory control results.

[0006] A second aspect of this application provides a driving trajectory prediction and control system for an autonomous vehicle, the system comprising:

[0007] Data acquisition component: Collects real-time driving scene data through a multi-source heterogeneous sensor network to obtain multimodal driving data; Trajectory prediction component: Inputs the multimodal driving data into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set; Optimization component: Performs predictive control optimization on the multimodal predicted trajectory set to construct a target predictive control optimization dataset; Trajectory control component: Solves for a target optimization control sequence based on the target predictive control optimization dataset, performs trajectory tracking control on the autonomous vehicle according to the target optimization control sequence, and generates driving trajectory control results.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, real-time driving scenario data is collected through a multi-source heterogeneous sensor network to obtain multimodal driving data. Next, the multimodal driving data is input into a hybrid prediction network for trajectory prediction, resulting in a multimodal predicted trajectory set. Then, predictive control optimization is performed on the multimodal predicted trajectory set to construct a target predictive control optimization dataset. Finally, a target optimization control sequence is solved based on the target predictive control optimization dataset, and trajectory tracking control is executed on the autonomous vehicle according to the target optimization control sequence to generate driving trajectory control results. This solves the technical problem of low trajectory control accuracy and driving stability of autonomous vehicles in complex road environments in existing technologies, achieving the technical effect of improving the trajectory control accuracy and driving stability of autonomous vehicles. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 A schematic flowchart of a driving trajectory prediction and control method for autonomous vehicles provided in an embodiment of this application;

[0012] Figure 2 This is a schematic diagram of the driving trajectory prediction and control system for unmanned vehicles provided in an embodiment of this application.

[0013] Figure labeling: Data acquisition component 11, trajectory prediction component 12, optimization component 13, trajectory control component 14. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] Example 1, as Figure 1 As shown, this application provides a driving trajectory prediction and control method for autonomous vehicles, wherein the method includes:

[0016] Multimodal driving data is obtained by collecting real-time driving scenario data through a multi-source heterogeneous sensor network.

[0017] Furthermore, multimodal driving data is obtained by collecting real-time driving scenario data through a multi-source heterogeneous sensor network. Methods include:

[0018] The system retrieves real-time motion state parameters and scene complexity parameters of the autonomous vehicle; dynamically adjusts the multi-source heterogeneous sensor network based on the real-time motion state parameters and scene complexity parameters to collect raw sensor datasets; performs spatiotemporal synchronization calibration on the raw sensor datasets to generate synchronized sensor data frames; performs adaptive noise suppression on the synchronized sensor data frames to obtain multiple sensor data signals and dynamically gain signal quality to obtain the multimodal driving data.

[0019] During the operation of the autonomous vehicle, the onboard central control unit first retrieves the real-time motion state parameters and scene complexity parameters of the autonomous vehicle. The real-time motion state parameters include at least vehicle speed, vehicle acceleration, yaw rate, steering wheel angle, braking status, and the vehicle's current position coordinates. The scene complexity parameters include at least the number of road targets, number of lanes, road curvature, obstacle density, weather conditions, and light intensity. Subsequently, a multi-source heterogeneous sensor network is dynamically adjusted based on the real-time motion state parameters and scene complexity parameters. This multi-source heterogeneous sensor network includes at least cameras, LiDAR, millimeter-wave radar, ultrasonic sensors, inertial measurement units, and a high-precision positioning module. When the vehicle is identified as being on a high-speed straight road, the sampling frequency of the millimeter-wave radar and LiDAR is increased; when the vehicle is identified as being in a low-speed complex traffic scene, the data acquisition weight of the cameras and ultrasonic sensors is increased to collect the corresponding raw sensor dataset. Afterward, the raw sensor dataset undergoes spatiotemporal synchronization calibration. The system aligns the timestamps of data from each sensor using a unified synchronization clock, and performs coordinate mapping and transformation on the spatial poses of different sensors based on a unified vehicle coordinate system. This eliminates temporal and spatial deviations between multi-source data, generating synchronized sensing data frames. Subsequently, adaptive noise suppression processing is applied to these synchronized sensing data frames. This includes image denoising and brightness equalization for camera image data, outlier removal and point cloud smoothing for LiDAR point cloud data, and false target filtering and distance fluctuation smoothing for millimeter-wave radar data, resulting in multiple sensing data signals. Next, dynamic signal quality gain is applied based on the signal quality indices corresponding to each sensing data signal. These indices include at least signal-to-noise ratio, target detection continuity, data integrity, and data stability. Sensing data signals with higher quality have higher fusion gain weights, while those with lower quality have lower fusion gain weights. Finally, the multiple sensing data signals processed by dynamic gain are uniformly fused and encoded to obtain the multimodal driving data.

[0020] The multimodal driving data is input into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set.

[0021] Furthermore, the method involves inputting the multimodal driving data into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set, including:

[0022] A hybrid prediction network is constructed, comprising a hierarchical interaction perception module, a multimodal prior learning module, and a decoder module. The hierarchical interaction perception module performs heterogeneous encoding on the multimodal driving data to extract deep interaction features. The multimodal prior learning module learns driving trajectories to construct multiple latent variables of driving intention. The decoder module jointly decodes the deep interaction features and the multiple latent variables of driving intention to generate multiple hypothetical driving trajectory data. Feasibility analysis is performed on the multiple hypothetical driving trajectory data to obtain the multimodal predicted trajectory set.

[0023] First, a hybrid prediction network is constructed and deployed on the onboard computing platform of the autonomous vehicle. This network includes a hierarchical interaction perception module, a multimodal prior learning module, and a decoder module. The hierarchical interaction perception module establishes spatial interaction relationships between different traffic participants, the multimodal prior learning module learns historical driving behavior and scene evolution patterns, and the decoder module generates future predicted trajectories. Then, the hierarchical interaction perception module performs heterogeneous encoding on the multimodal driving data. This includes visual feature convolutional encoding on camera image data, spatial structure encoding on laser point cloud data, target motion state encoding on millimeter-wave radar data, and temporal trajectory encoding on historical vehicle trajectory data. Based on the positional, speed, and directional relationships between traffic participants, hierarchical interaction correlation analysis is performed to extract deep interaction features. Finally, the multimodal prior learning module learns the driving trajectory, utilizing historical driving sample data to study vehicle following behavior. Pattern learning is performed on lane-changing behavior, obstacle avoidance behavior, and steering behavior, and multiple latent variables of driving intention are constructed in conjunction with the current traffic scene state. Different latent variables of driving intention correspond to different future driving trends. Subsequently, the deep interaction features and the multiple latent variables of driving intention are jointly decoded by the decoder module. The multiple latent variables of driving intention are input into the corresponding decoder branches, and cross-attention calculation is performed in conjunction with the deep interaction features to generate multiple time step position coordinate sequences in the future prediction time domain to construct multiple hypothetical driving trajectory data. Then, the multiple hypothetical driving trajectory data are subjected to feasibility analysis. The hypothetical driving trajectory data are subjected to trajectory continuity detection, vehicle kinematic constraint detection, road boundary constraint detection, and obstacle collision risk detection, respectively, and abnormal trajectories with trajectory abrupt changes, vehicle inaccessibility, or collision risk exceeding a preset threshold are eliminated. Finally, the remaining hypothetical driving trajectory data that meet the driving conditions are sorted according to trajectory confidence to obtain the multimodal predicted trajectory set.

[0024] Furthermore, the method involves jointly decoding the deep interaction features and the multiple latent variables of driving intent using the decoder module to generate multiple hypothetical driving trajectory data, including:

[0025] The process involves traversing multiple latent variables of driving intent for decoding analysis, constructing multiple decoder branches, each corresponding to one of the latent variables of driving intent; retrieving target time location information, concatenating the target time location information with the multiple latent variables of driving intent to generate an initial query vector; performing cross-attention calculation based on the initial query vector and the deep interaction features to retrieve intent interaction information; activating the multiple decoder branches to perform time step location analysis according to the predicted time domain data, determining multiple time step location coordinates for autoregression, and constructing the multiple hypothetical driving trajectory data.

[0026] First, the multiple latent variables of driving intention are decoded and analyzed. Multiple decoder branches are constructed based on the driving behavior trend corresponding to each latent variable of driving intention. These decoder branches correspond to following trajectory branches, lane-changing trajectory branches, obstacle avoidance trajectory branches, and steering trajectory branches, respectively, and each decoder branch is configured with an independent trajectory parameter update channel. Next, the target time location information is retrieved. This target time location information includes at least the current position coordinates of the autonomous vehicle, the current heading angle, the current speed value, and the current acceleration value. The target time location information is then concatenated with the corresponding latent variables of driving intention to generate an initial query vector. Afterward, cross-attention calculation is performed based on the initial query vector and the deep interaction features, where the initial query vector is used as the query vector, and the deep interaction features are used as the... Using key-value features, the interaction relationships between traffic participants, road boundary constraints, and dynamic obstacle movement trends are correlated and retrieved to obtain corresponding intentional interaction information. Subsequently, the multiple decoder branches are activated to perform time-step position analysis based on the predicted time-domain data. A preset prediction time-domain length is used as the trajectory prediction window, and the position coordinates, velocity values, and heading angle data corresponding to each future moment are output step by step in the order of time steps. After the prediction result of the current time step is generated, the prediction result of the current time step is fed back to the next time step as an input condition for autoregressive update. Then, the corresponding trajectory coordinate sequence is constructed based on the position coordinates of multiple time steps generated in continuous time steps, and the trajectory change amplitude is smoothed and limited by the vehicle motion continuity constraint. Finally, the multiple trajectory coordinate sequences output by each decoder branch are output as multiple hypothetical driving trajectory data.

[0027] Predictive control optimization is performed on the multimodal predicted trajectory set to construct a target predictive control optimization dataset.

[0028] Furthermore, the method for performing predictive control optimization on the multimodal predicted trajectory set to construct a target predictive control optimization dataset includes:

[0029] The multimodal predicted trajectory set is traversed and spatiotemporally aligned to generate a dynamic obstacle probability occupancy grid map; vehicle kinematic constraints are set and statistical analysis is performed in conjunction with the multimodal predicted trajectory set to construct the reachability state space of the autonomous vehicle; the dynamic obstacle probability occupancy grid map and the reachability state space are structured and encoded to construct the target prediction control optimization dataset.

[0030] First, the multimodal predicted trajectory set is traversed for spatiotemporal alignment. This involves synchronizing the future time step position coordinates of each predicted trajectory according to a unified prediction time axis and performing spatial mapping based on a unified road coordinate system, establishing spatiotemporal relationships between different predicted trajectories within the same prediction time domain. Next, based on the future position distribution, speed changes, and direction changes of traffic targets in each predicted trajectory, the occupancy probability of the road area where the target is located is statistically analyzed. The road area is then discretized into a grid according to a preset grid division rule, where each grid cell corresponds to the probability value of being occupied by a dynamic obstacle in the future prediction time domain, generating a dynamic obstacle probability occupancy grid map. Finally, vehicle kinematic constraints are set and statistical analysis is performed using the multimodal predicted trajectory set. The kinematic constraints include at least the vehicle's maximum steering angle constraint, maximum lateral acceleration constraint, minimum turning radius constraint, and vehicle stability constraint. Based on the changes in the future motion state of the autonomous vehicle under different control input conditions, multiple reachable motion state nodes of the vehicle in the prediction time domain are generated to construct the reachable state space of the autonomous vehicle. Subsequently, the dynamic obstacle probability occupancy grid map and the reachable state space are subjected to structured encoding processing. In this process, the dynamic obstacle risk areas, vehicle passable areas, and vehicle reachable trajectory nodes are characterized, and corresponding spatiotemporal correlations and state transition relationships are established. Finally, the encoded obstacle probability distribution information, vehicle reachable state information, and trajectory risk constraint information are fused and organized to construct the target prediction control optimization dataset.

[0031] Based on the target predictive control optimization dataset, the target optimization control sequence is solved, and the trajectory tracking control of the unmanned vehicle is performed according to the target optimization control sequence to generate the driving trajectory control result.

[0032] Furthermore, the method for solving the target optimization control sequence based on the target predictive control optimization dataset includes:

[0033] A hierarchical parallel optimization solver is constructed, comprising an upper discrete search layer and a lower continuous optimization layer. The target prediction control optimization dataset is synchronized to the hierarchical parallel optimization solver. S1: The upper discrete search layer performs a graph search on the reachable state space to generate a candidate state sequence. S2: The lower continuous optimization layer combines the candidate state sequence as an initial guess solution for control optimization to obtain the target optimization control sequence.

[0034] First, a hierarchical parallel optimization solver is constructed, comprising an upper discrete search layer and a lower continuous optimization layer. The upper discrete search layer is used to quickly search for candidate motion paths that meet the passage conditions in complex dynamic scenarios, and the lower continuous optimization layer is used to continuously optimize the control variables of the candidate motion paths. Then, the target prediction control optimization dataset is synchronized to the hierarchical parallel optimization solver. The target prediction control optimization dataset includes at least a dynamic obstacle probability occupancy grid, a vehicle reachable state space, trajectory risk constraint information, and road passage constraint information. Next, step S1 is executed, performing a graph search on the reachable state space through the upper discrete search layer. The state node corresponding to the vehicle's current position is used as the starting search node, and the state node corresponding to the target trajectory endpoint is used as the target search node. The search is based on the dynamic obstacle risk cost and trajectory... A search evaluation function is established based on the smoothing cost and road traffic cost. Path search is performed on the connection relationships between multiple state nodes to generate multiple candidate state sequences that satisfy vehicle motion constraints. Subsequently, step S2 is executed, where the candidate state sequences are used as initial guess solutions in conjunction with the lower continuous optimization layer for control optimization. The trajectory nodes in the candidate state sequences are used as the initial trajectory for continuous optimization, and the steering angle control, longitudinal acceleration control, and braking control are continuously optimized using the vehicle dynamics model. Simultaneously, the trajectory tracking error, control input change rate, and vehicle lateral stability are jointly constrained and solved. Then, the optimization results are iteratively converged in each control cycle. Iterative optimization stops when the control error between two consecutive optimization results is less than a preset convergence threshold. Finally, the target optimized control sequence that satisfies the vehicle motion stability and trajectory safety constraints is output.

[0035] Furthermore, the method of performing trajectory tracking control on the autonomous vehicle according to the target optimized control sequence to generate driving trajectory control results includes:

[0036] The target optimized control sequence is parsed according to the control cycle to determine the desired control quantity; the autonomous vehicle is periodically tracked and controlled based on the target optimized control sequence to obtain the periodic control quantity; the periodic control quantity and the desired control quantity are compared in multiple dimensions to generate a control comparison result; when the control comparison result is greater than a preset control deviation threshold, a control compensation command is generated; the target optimized control sequence is trajectory compensated using the control compensation command, and the target optimized control sequence is backtracked and updated to continuously track and control the autonomous vehicle, generating the driving trajectory control result.

[0037] First, the target optimized control sequence is parsed according to the control cycle, where the control cycle is a preset time interval control window. Within each control cycle, the target steering angle, target vehicle speed, target longitudinal acceleration, and target heading angle information for the corresponding time step are extracted to determine the desired control quantity. Then, based on the target optimized control sequence, the autonomous vehicle is subjected to periodic tracking control. The onboard controller sends corresponding control commands to the steering actuator, drive actuator, and braking actuator according to the desired control quantity for the current control cycle, and collects the actual operating status data of the autonomous vehicle within the current control cycle in real time through the vehicle status feedback interface to obtain the periodic control quantity. Afterward, the periodic control quantity and the desired control quantity are compared in multiple dimensions, where the multi-dimensional comparison includes at least a comparison of lateral trajectory offset and vehicle... Speed ​​deviation, heading angle deviation, and acceleration deviation are compared, and a comprehensive error calculation is performed based on the deviation values ​​of each dimension to generate a control comparison result. Subsequently, when the control comparison result exceeds a preset control deviation threshold, a control compensation command is generated, wherein the corresponding compensation steering amount, compensation acceleration amount, and compensation braking force adjustment amount are determined based on the deviation direction, deviation amplitude, and deviation change trend. Then, the target optimized control sequence is trajectory compensated using the control compensation command, wherein the actual vehicle operating state corresponding to the current control cycle is used as the new state starting point, the subsequent unexecuted target optimized control sequence is backtracked and updated, and the trajectory control parameters corresponding to the subsequent time steps are readjusted. Finally, based on the updated target optimized control sequence, continuous closed-loop trajectory tracking control is performed on the autonomous vehicle to generate a driving trajectory control result.

[0038] In summary, the embodiments of this application have at least the following technical effects:

[0039] First, real-time driving scenario data is collected through a multi-source heterogeneous sensor network to obtain multimodal driving data. Next, the multimodal driving data is input into a hybrid prediction network for trajectory prediction, resulting in a multimodal predicted trajectory set. Then, predictive control optimization is performed on the multimodal predicted trajectory set to construct a target predictive control optimization dataset. Finally, a target optimization control sequence is solved based on the target predictive control optimization dataset, and trajectory tracking control is executed on the autonomous vehicle according to the target optimization control sequence to generate driving trajectory control results. This solves the technical problem of low trajectory control accuracy and driving stability of autonomous vehicles in complex road environments in existing technologies, achieving the technical effect of improving the trajectory control accuracy and driving stability of autonomous vehicles.

[0040] Example 2, based on the same inventive concept as the driving trajectory prediction and control method for autonomous vehicles in the foregoing examples, such as... Figure 2 As shown, this application provides a driving trajectory prediction and control system for autonomous vehicles, wherein the system includes:

[0041] Data acquisition component 11: Collects real-time driving scene data through a multi-source heterogeneous sensor network to obtain multimodal driving data; Trajectory prediction component 12: Inputs the multimodal driving data into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set; Optimization component 13: Performs predictive control optimization on the multimodal predicted trajectory set to construct a target predictive control optimization dataset; Trajectory control component 14: Solves for a target optimization control sequence based on the target predictive control optimization dataset, performs trajectory tracking control on the autonomous vehicle according to the target optimization control sequence, and generates driving trajectory control results.

[0042] Furthermore, the data acquisition component 11 is used to perform the following methods:

[0043] The system retrieves real-time motion state parameters and scene complexity parameters of the autonomous vehicle; dynamically adjusts the multi-source heterogeneous sensor network based on the real-time motion state parameters and scene complexity parameters to collect raw sensor datasets; performs spatiotemporal synchronization calibration on the raw sensor datasets to generate synchronized sensor data frames; performs adaptive noise suppression on the synchronized sensor data frames to obtain multiple sensor data signals and dynamically gain signal quality to obtain the multimodal driving data.

[0044] Furthermore, the trajectory prediction component 12 is used to perform the following method:

[0045] A hybrid prediction network is constructed, comprising a hierarchical interaction perception module, a multimodal prior learning module, and a decoder module. The hierarchical interaction perception module performs heterogeneous encoding on the multimodal driving data to extract deep interaction features. The multimodal prior learning module learns driving trajectories to construct multiple latent variables of driving intention. The decoder module jointly decodes the deep interaction features and the multiple latent variables of driving intention to generate multiple hypothetical driving trajectory data. Feasibility analysis is performed on the multiple hypothetical driving trajectory data to obtain the multimodal predicted trajectory set.

[0046] Furthermore, the trajectory prediction component 12 is used to perform the following method:

[0047] The process involves traversing multiple latent variables of driving intent for decoding analysis, constructing multiple decoder branches, each corresponding to one of the latent variables of driving intent; retrieving target time location information, concatenating the target time location information with the multiple latent variables of driving intent to generate an initial query vector; performing cross-attention calculation based on the initial query vector and the deep interaction features to retrieve intent interaction information; activating the multiple decoder branches to perform time step location analysis according to the predicted time domain data, determining multiple time step location coordinates for autoregression, and constructing the multiple hypothetical driving trajectory data.

[0048] Furthermore, the optimization component 13 is used to perform the following method:

[0049] The multimodal predicted trajectory set is traversed and spatiotemporally aligned to generate a dynamic obstacle probability occupancy grid map; vehicle kinematic constraints are set and statistical analysis is performed in conjunction with the multimodal predicted trajectory set to construct the reachability state space of the autonomous vehicle; the dynamic obstacle probability occupancy grid map and the reachability state space are structured and encoded to construct the target prediction control optimization dataset.

[0050] Furthermore, the trajectory control component 14 is used to perform the following methods:

[0051] A hierarchical parallel optimization solver is constructed, comprising an upper discrete search layer and a lower continuous optimization layer. The target prediction control optimization dataset is synchronized to the hierarchical parallel optimization solver. S1: The upper discrete search layer performs a graph search on the reachable state space to generate a candidate state sequence. S2: The lower continuous optimization layer combines the candidate state sequence as an initial guess solution for control optimization to obtain the target optimization control sequence.

[0052] Furthermore, the trajectory control component 14 is used to perform the following methods:

[0053] The target optimized control sequence is parsed according to the control cycle to determine the desired control quantity; the autonomous vehicle is periodically tracked and controlled based on the target optimized control sequence to obtain the periodic control quantity; the periodic control quantity and the desired control quantity are compared in multiple dimensions to generate a control comparison result; when the control comparison result is greater than a preset control deviation threshold, a control compensation command is generated; the target optimized control sequence is trajectory compensated using the control compensation command, and the target optimized control sequence is backtracked and updated to continuously track and control the autonomous vehicle, generating the driving trajectory control result.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A driving trajectory prediction and control method for autonomous vehicles, characterized in that, The method includes: Multimodal driving data is obtained by collecting real-time driving scenario data through a multi-source heterogeneous sensor network; The multimodal driving data is input into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set; Predictive control optimization is performed on the multimodal predicted trajectory set to construct a target predictive control optimization dataset; Based on the target predictive control optimization dataset, the target optimization control sequence is solved, and the trajectory tracking control of the unmanned vehicle is performed according to the target optimization control sequence to generate the driving trajectory control result.

2. The driving trajectory prediction and control method for unmanned vehicles as described in claim 1, characterized in that, Multimodal driving data is obtained by collecting real-time driving scenario data through a multi-source heterogeneous sensor network. The methods include: Retrieve real-time motion state parameters and scene complexity parameters of the autonomous vehicle; The multi-source heterogeneous sensor network is dynamically adjusted based on the real-time motion state parameters and the scene complexity parameters, and the original sensor dataset is collected. The original sensor dataset is spatiotemporally synchronized and calibrated to generate a synchronized sensor data frame. Adaptive noise suppression is performed on the synchronous sensing data frame to obtain multiple sensing data signals, and dynamic signal quality gain is applied to obtain the multimodal driving data.

3. The driving trajectory prediction and control method for unmanned vehicles as described in claim 1, characterized in that, The multimodal driving data is input into a hybrid prediction network for trajectory prediction to obtain a multimodal predicted trajectory set. The method includes: A hybrid prediction network is constructed, which includes a hierarchical interaction perception module, a multimodal prior learning module, and a decoder module; The hierarchical interaction perception module performs heterogeneous encoding on the multimodal driving data to extract deep interaction features; The driving trajectory is learned through the multimodal prior learning module, and multiple latent variables of driving intention are constructed. The decoder module jointly decodes the deep interaction features and the multiple latent variables of driving intention to generate multiple hypothetical driving trajectory data. A feasibility analysis is performed on the multiple hypothetical driving trajectory data to obtain the multimodal predicted trajectory set.

4. The driving trajectory prediction and control method for unmanned vehicles as described in claim 3, characterized in that, The decoder module jointly decodes the deep interaction features and the multiple latent variables of driving intent to generate multiple hypothetical driving trajectory data. The method includes: The multiple latent variables of driving intent are traversed for decoding analysis, and multiple decoder branches are constructed. The multiple decoder branches have a corresponding relationship with the multiple latent variables of driving intent. Retrieve the target time location information, and concatenate the target time location information with the multiple driving intention latent variables to generate an initial query vector; Cross-attention calculation is performed based on the initial query vector and the deep interaction features to retrieve intent interaction information; The multiple decoder branches are activated to perform time step position analysis based on the predicted time domain data, determine the time step position coordinates of multiple time steps for autoregression, and construct the multiple hypothetical driving trajectory data.

5. The driving trajectory prediction and control method for unmanned vehicles as described in claim 1, characterized in that, The method for performing predictive control optimization on the multimodal predicted trajectory set to construct a target predictive control optimization dataset includes: The multimodal predicted trajectory set is traversed for spatiotemporal alignment to generate a dynamic obstacle probability occupancy grid map; By setting kinematic constraints for the autonomous vehicle and combining them with the multimodal predicted trajectory set for statistical analysis, an reachability state space for the autonomous vehicle is constructed. The target prediction control optimization dataset is constructed by structurally encoding the dynamic obstacle probability occupancy grid map and the reachable state space.

6. The driving trajectory prediction and control method for unmanned vehicles as described in claim 5, characterized in that, The method for solving the objective optimization control sequence based on the aforementioned objective predictive control optimization dataset includes: A hierarchical parallel optimization solver is constructed, which includes an upper discrete search layer and a lower continuous optimization layer. Synchronize the target prediction control optimization dataset to the hierarchical parallel optimization solver: S1: Through the upper discrete search layer, a graph search is performed on the reachable state space to generate a candidate state sequence; S2: Through the lower continuous optimization layer, the candidate state sequence is used as the initial guess solution for control optimization to obtain the target optimization control sequence.

7. The driving trajectory prediction and control method for unmanned vehicles as described in claim 6, characterized in that, The method includes performing trajectory tracking control on an autonomous vehicle according to the aforementioned target optimization control sequence to generate driving trajectory control results. The target optimized control sequence is analyzed according to the control period to determine the desired control quantity; Based on the target optimized control sequence, the autonomous vehicle is periodically tracked and controlled to obtain periodic control quantities. The periodic control quantity is compared with the desired control quantity in multiple dimensions to generate a control comparison result. When the control comparison result is greater than the preset control deviation threshold, a control compensation command is generated; The target optimized control sequence is compensated for trajectory by the control compensation command, and the target optimized control sequence is updated backtrackingly to continuously track and control the unmanned vehicle, thereby generating the driving trajectory control result.

8. A driving trajectory prediction and control system for unmanned vehicles, characterized in that, The system is used to implement the driving trajectory prediction and control method for autonomous vehicles according to any one of claims 1-7, the system comprising: Data acquisition component: Collects real-time driving scenario data through a multi-source heterogeneous sensor network to obtain multimodal driving data; Trajectory prediction component: The multimodal driving data is input into the hybrid prediction network to perform trajectory prediction and obtain a multimodal predicted trajectory set; Optimization component: Perform predictive control optimization on the multimodal predicted trajectory set to construct a target predictive control optimization dataset; Trajectory control component: Based on the target predictive control optimization dataset, solve for the target optimization control sequence, perform trajectory tracking control on the autonomous vehicle according to the target optimization control sequence, and generate driving trajectory control results.

Images