Intelligent driving evaluation method and equipment

Abstract

The invention discloses an intelligent driving evaluation method and equipment. The method comprises the following steps: acquiring environment perception data recorded by a real vehicle and a corresponding recorded vehicle pose; performing data transformation on the environment perception data according to a pose difference between the vehicle model pose and the recorded vehicle pose so as to generate transformation perception data corresponding to a vehicle model view angle; generating a prediction track based on the transformation perception data, iteratively updating the pose of the vehicle model according to the prediction track, and recording dynamic evaluation data; the dynamic evaluation data comprises an iteratively updated vehicle model pose, transformation perception data and a prediction track; and determining an evaluation performance index of the intelligent driving system based on the dynamic evaluation data. Therefore, a dynamic closed-loop evaluation mechanism under a real data condition is constructed, self-adaptive association of perception input and a vehicle motion state is realized, and an intelligent driving system can generate continuous decision feedback in an evaluation process, so that the time sequence consistency and stability of the system in dynamic control are accurately reflected.

Description

Technical Field

[0001] This application relates to the field of driver assistance systems technology, and in particular to an intelligent driving evaluation method and device. Background Technology

[0002] In the rapid development of intelligent driving technology, the safety and stability verification of autonomous driving systems has become a crucial step in the research and development process. To evaluate the performance of modules such as perception, decision-making, and control, the industry currently widely adopts a combination of simulation evaluation and real-vehicle testing. Simulation evaluation can reproduce traffic scenarios in a controlled environment, reduce testing risks, and improve testing efficiency; therefore, it is widely used in the algorithm iteration and model training stages.

[0003] Currently, autonomous driving evaluation methods mainly include two categories: open-loop evaluation and virtual scenario-based evaluation. In open-loop evaluation, pre-recorded real-world road datasets are played back to perform discrete frame analysis on the outputs of individual modules such as perception and prediction. While this method boasts high scene coverage efficiency, the lack of dynamic vehicle feedback during the testing process means it can only generate decisions based on static environmental data, failing to verify its actual performance during continuous control. This leads to significant discrepancies between the evaluation results and real-world road tests, especially in interactive scenarios where it struggles to reflect system-level performance defects.

[0004] Furthermore, evaluation methods based on 3D reconstruction use technologies such as LiDAR scanning and photogrammetry to create high-precision digital models of real roads and execute the entire autonomous driving process in a virtual environment. While this method can effectively verify system-level performance, scene acquisition requires specialized hardware, and the reconstruction process involves extensive manual annotation and model optimization, with a single scene construction cycle potentially lasting several weeks. In addition, simulation requires real-time computation of the physics engine and interaction with multiple agents, consuming enormous computational resources and making it difficult to support the high-frequency testing needs of massive scenarios. Summary of the Invention

[0005] This application provides an intelligent driving evaluation method, system, device, storage medium, and program product to at least solve one of the above-mentioned technical problems.

[0006] In a first aspect, embodiments of this application provide an intelligent driving evaluation method, comprising: acquiring environmental perception data recorded by a real vehicle and the corresponding recorded vehicle pose; performing data transformation on the environmental perception data according to the pose difference between the vehicle model pose and the recorded vehicle pose to generate transformed perception data from the perspective of the corresponding vehicle model; generating a predicted trajectory based on the transformed perception data, and iteratively updating the vehicle model pose according to the predicted trajectory, and recording dynamic evaluation data; the dynamic evaluation data includes the iteratively updated vehicle model pose, transformed perception data, and predicted trajectory; and determining the evaluation performance indicators of the intelligent driving system based on the dynamic evaluation data.

[0007] Secondly, embodiments of this application provide a storage medium storing one or more programs including execution instructions, which can be read and executed by electronic devices (including but not limited to computers, servers, or network devices) to perform any of the intelligent driving evaluation methods described above.

[0008] Thirdly, a computer device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any of the intelligent driving evaluation methods described above in this application.

[0009] Fourthly, embodiments of this application also provide a computer program product, the computer program product including a computer program stored on a storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to execute any of the above-mentioned intelligent driving evaluation methods.

[0010] The beneficial effects of the embodiments of this application are as follows: By transforming real-vehicle perception data based on the difference between the vehicle model pose and the recorded vehicle pose, and combining this with the iterative update process of the predicted trajectory, a dynamic closed-loop evaluation mechanism under real-data conditions is constructed. This mechanism achieves adaptive correlation between perception input and vehicle motion state, enabling the intelligent driving system to generate continuous decision feedback during the evaluation process, thereby accurately reflecting the temporal consistency and stability of the system in dynamic control. Therefore, system-level performance verification of the autonomous driving system can be achieved without the need for 3D scene reconstruction, effectively improving the authenticity, continuity, and reproducibility of the evaluation results. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are 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 flowchart illustrating an example of an intelligent driving evaluation method according to an embodiment of this application is shown; Figure 2 A flowchart illustrating an example of generating transformation perception data corresponding to the viewpoint of a vehicle model according to an embodiment of this application is shown. Figure 3Shows an operation flowchart of an example of iteratively updating the pose of a simulation vehicle model according to an embodiment of the present application; Figure 4 Shows a flowchart of another example of an intelligent driving evaluation method according to an embodiment of the present application; Figure 5 Is a schematic structural diagram of an embodiment of an electronic device of the present application. Specific embodiments

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the scope of protection of the present application. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other.

[0014] It should also be noted that in this article, the terms "include" and "comprise" not only include those elements, but also other elements not explicitly listed, or also include elements inherent to such a process, method, article or device. Without further limitation, the elements defined by the statement "including..." do not exclude the existence of additional identical elements in the process, method, article or device including the said elements.

[0015] It should be noted that in the current related technologies, although some open-loop evaluation methods can provide convenient offline verification in the early stage of the algorithm, with the increase in the complexity of intelligent driving algorithms, this method is difficult to meet the requirements of system-level linkage verification. Open-loop evaluation only makes a static comparison between the input data and the algorithm output, ignoring the dynamic impact of the decision output on the vehicle motion state. For example, in scenarios with strong interactivity such as lane-changing obstacle avoidance and intersection passing, the trajectory planning and control commands of the vehicle will act on the spatial distribution of the perception data, thereby affecting subsequent perception and prediction results. However, traditional open-loop methods cannot capture this dynamic feedback relationship, resulting in the evaluation results being unable to truly reflect the stability and safety of the algorithm in continuous time series.

[0016] On the other hand, while 3D scene reconstruction methods can recover environmental geometric features and conduct closed-loop simulations, their engineering implementation faces high hardware dependencies and computational burdens. To obtain high-precision road models, multi-line LiDAR, panoramic cameras, and inertial navigation equipment are typically used for collaborative data acquisition, followed by manual post-processing to generate 3D point clouds and texture maps. This entire process is not only costly but also has a long data update cycle, limiting the diversity and timeliness of test samples. Especially during rapid algorithm training cycles, the reconstructed scene often lags behind the algorithm version, causing evaluation results to become out of sync with the latest model characteristics.

[0017] Furthermore, while virtual reconstruction scenarios can simulate complex traffic environments, they struggle to fully replicate the randomness and diversity of real-world driving data. For example, factors such as road surface irregularities, lighting variations, sensor noise, and pedestrian movement can all affect the algorithm's performance in a real-world environment. These real-world disturbances are often simplified or ignored in simulation models, leading to idealization biases in closed-loop simulation results. Therefore, even with high-fidelity modeling, evaluation results may still exhibit systematic errors compared to real-world road test performance.

[0018] Furthermore, with the increasing trend of parallel development of multiple vehicle models and algorithms, the inadequacy of traditional closed-loop simulation platforms in terms of flexibility has become increasingly prominent. Existing systems mostly adopt a tightly coupled architecture, requiring different versions of autonomous driving algorithms to be re-integrated or the simulation environment to be rebuilt, resulting in high switching costs and easy introduction of compatibility issues. In addition, limited by computing resources and scene scheduling mechanisms, large-scale distributed evaluation often requires long queuing times, which cannot support the high-frequency verification needs of enterprise-level algorithms in massive scenarios. As a result, current evaluation technologies still have significant bottlenecks in terms of efficiency, economy, and scalability.

[0019] It should be understood that the above description of the relevant technologies is intended only to help the public better understand the inventive spirit and motivation of this application, and is not intended to limit this application. Furthermore, the technical solutions described in the above-mentioned relevant technologies are not prior art, and may also be undisclosed technical solutions, such as those under research or in the laboratory stage.

[0020] The technical solutions in this application, including the collection, storage, use, processing, transmission, provision, and disclosure of users' personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0021] Figure 1 A flowchart illustrating an example of an intelligent driving evaluation method according to an embodiment of this application is shown.

[0022] Regarding the execution subject of the method in the embodiments of this application, it can be any controller or processor with computing or processing capabilities. In some examples, the method in the embodiments of this application can be integrated and configured in an electronic device or terminal through software, hardware or a combination of software and hardware, and the type of terminal or electronic device can be diverse, such as mobile phone, tablet computer, desktop computer or vehicle terminal, etc.

[0023] For example, the execution subject of the method in the embodiments of this application can be integrated into the intelligent driving evaluation controller. By transforming perception data driven by position difference and dynamically iteratively updating based on predicted trajectory, an intelligent driving evaluation mechanism that can achieve closed-loop feedback without three-dimensional reconstruction is established. It can achieve continuous control feedback and dynamic performance quantification under real environmental data conditions, taking into account both authenticity and repeatability, and greatly improving the relevance and credibility of the evaluation results to the real vehicle performance.

[0024] like Figure 1 As shown, in step S110, the environmental perception data recorded by the actual vehicle and the corresponding recorded vehicle pose are obtained.

[0025] In some implementations, autonomous driving test vehicles equipped with multiple types of sensors (such as LiDAR, cameras, millimeter-wave radar, inertial measurement units (IMUs), and high-precision RTK-GPS modules) collect data in real-world road traffic scenarios to obtain environmental perception data covering multimodal information. This data reflects the external traffic environment characteristics perceived by the vehicle during operation and maintains a strict spatiotemporal correspondence with the vehicle's motion state.

[0026] In some examples, environmental perception data includes at least one of the following: dynamic object detection data, lane topology information, traffic signal status, road structure information, or traffic sign semantics.

[0027] Specifically, dynamic target detection data may include the semantic category of the target (such as pedestrians, motor vehicles, non-motor vehicles, etc.), the three-dimensional position coordinates, velocity vector, and heading angle of the target, to reflect the motion state of surrounding traffic participants; lane topology information includes lane line equation parameters, lane type (such as main lane, ramp, dashed or solid line lane), and lane connectivity, to characterize the topology of the road network; traffic signal status includes the phase sequence and remaining duration of traffic lights, to indicate the traffic conditions at intersections; road structure information includes roadside boundary point sets, drivable area grid maps, etc., to describe the drivable range and boundary constraints of the road; traffic sign semantics includes sign type (such as speed limit signs, no-entry signs, etc.) and associated parameter values, providing semantic basis for subsequent behavioral decisions.

[0028] Simultaneously, during the data acquisition process, the vehicle's pose information (including position coordinates, heading angle, pitch angle, roll angle, etc.) is acquired in real time by a high-precision positioning system and synchronized with the aforementioned sensing data. To ensure the spatiotemporal consistency of the data, a unified timestamp management and sensor calibration mechanism can be adopted to align the sampling results from different sensors to the same coordinate system. The recorded vehicle pose and environmental perception data obtained in this way together constitute a complete real-vehicle recording dataset, providing high-quality data from real road scenarios.

[0029] In step S120, the environmental perception data is transformed based on the pose difference between the vehicle model pose and the recorded vehicle pose to generate transformed perception data from the corresponding vehicle model perspective.

[0030] In some implementations, by calculating the pose difference between the vehicle model pose and the recorded vehicle pose, geometric transformation techniques (such as coordinate transformation matrices) are used to transform the environmental perception data, converting the recorded vehicle's perception data into data consistent with the current vehicle model's viewpoint and pose. For example, if there is a deviation between the actual recorded vehicle pose and the model vehicle's pose, the transformed perception data is remapped to the vehicle model's viewpoint after calculating the pose difference, thereby ensuring that the vehicle model receives the correct environmental input.

[0031] In step S130, a predicted trajectory is generated based on the transform sensing data, and the vehicle model pose is iteratively updated according to the predicted trajectory. Dynamic evaluation data is recorded, which includes the iteratively updated vehicle model pose, transform sensing data, and predicted trajectory.

[0032] In some implementations, the intelligent driving system processes the transformed environmental perception data to obtain a predicted trajectory under the external environmental conditions perceived from the vehicle model's perspective. This trajectory can be combined with its decision-making algorithms (such as path planning and behavior prediction) to generate an estimate of the vehicle model's possible driving paths over a future period. The generated predicted trajectory is then fed back to the vehicle model's control system to update the vehicle's pose (i.e., position and attitude) and to make corresponding control decisions. During the evaluation cycle, the vehicle model's pose is continuously iterated and updated, generating new perception data and predicted trajectories, and this data is continuously recorded.

[0033] By dynamically feeding back perception data and control decisions, real-time iterative updates of the vehicle model's pose and perception data are achieved, ensuring that the evaluation process reflects the intelligent driving system's behavior in real-world dynamic environments. Therefore, by continuously updating the vehicle model's state using real-vehicle recording data, a more comprehensive assessment of the intelligent driving system's stability, trajectory tracking accuracy, and responsiveness to changes in the dynamic environment can be made.

[0034] In step S140, the performance indicators of the intelligent driving system are determined based on the dynamic evaluation data.

[0035] Here, various evaluation metrics can be used to quantitatively analyze the performance of intelligent driving systems, such as trajectory tracking accuracy (e.g., distance from the target trajectory), system response time, stability (e.g., smoothness of vehicle attitude changes), and safety indicators (e.g., collision risk, spatiotemporal safety margin). By analyzing this dynamic evaluation data, the system can be assessed for various scenarios that may occur in actual operation, such as obstacle avoidance response and decision-making ability when interacting with other vehicles or pedestrians.

[0036] This application's embodiments combine the vehicle model and the actual vehicle's pose difference to transform environmental perception data into the vehicle model's perspective. This allows the vehicle model to make real-time decisions and controls based on the transformed perception data, generating predicted trajectories and continuously iteratively updating the vehicle's state to obtain recorded dynamic evaluation data. Ultimately, through in-depth analysis of this dynamic evaluation data, the overall performance of the intelligent driving system can be accurately quantified, and its performance in real-world road environments can be evaluated. Therefore, while ensuring high-precision evaluation, it avoids the computational costs associated with traditional 3D reconstruction and highly complex simulations, demonstrating strong practical application value and widespread potential.

[0037] In this embodiment of the application, the autonomous driving program performs continuous planning and control when playing back the recorded data of the real vehicle, instead of evaluating each frame of data as discrete data. This control will act on the position and attitude of the autonomous vehicle, so that the vehicle travels according to the predicted trajectory. The conclusions obtained by this evaluation method are similar to those of real vehicle testing.

[0038] In some examples of embodiments of this application, after iteratively updating the vehicle model pose according to the predicted trajectory, the updated vehicle model pose is determined as the vehicle model pose at the next moment, and the pose difference between the vehicle model pose at the next moment and the recorded vehicle pose at the next moment is recalculated; based on the recalculated pose difference, data transformation is performed on the environmental perception data at the next moment to generate the transformed perception data corresponding to the next moment.

[0039] Here, closed-loop evaluation is performed using data recorded from the actual vehicle. No scene reconstruction is required during the evaluation process. During scene playback, the pose difference between the autonomous vehicle after loop closure and the recording vehicle is used to transform some structured data, allowing the reconstruction of the input data required for the planning and control model. The planning and control model then outputs predicted trajectory information, and the pose information of the closed-loop vehicle is updated based on the predicted trajectory. This process is repeated to form continuous planning and control. This method eliminates the need for complex intelligent driving process integration and expensive scene reconstruction, allowing for direct use of data recorded from the actual vehicle for closed-loop evaluation, significantly saving manpower and material costs.

[0040] More specifically, in the dynamic evaluation of autonomous driving systems, the motion state of the vehicle model changes continuously with the output of the decision-making and control algorithms. To realistically simulate the closed-loop behavior of the system during continuous driving, this implementation method uses the vehicle model pose updated in each iteration as the input state for the next moment, thereby achieving continuous pose transfer and temporal correlation.

[0041] Specifically, utilizing the current moment Transform the sensing data to generate a predicted trajectory, and combine it with the previous moment. Output the vehicle model pose, and obtain the updated vehicle model pose through trajectory planning and vehicle dynamics constraint calculation. This pose update also affects the environmental observation conditions of the vehicle model in the next moment. The vehicle model pose needs to be recalculated. Recording vehicle position pose difference between Based on this, geometric coordinate transformations are performed on the recorded perception data, such as three-dimensional coordinate translation, rotation matrix transformation, and view projection correction, to generate transformed perception data corresponding to the vehicle model's viewpoint at the next moment.

[0042] It should be understood that the time in the embodiments of this application refers to the sampling time, and the size of the sampling time interval can be diverse and is not limited herein. Through time-by-time recalculation and data transformation as described in the embodiments of this application, the input data of the vehicle model will always remain consistent with its own dynamic state, thereby realizing real-time feedback on system decision and control output at the perception layer.

[0043] Furthermore, by repeatedly executing a continuous feedback loop of data transformation, predicted trajectory generation, and vehicle model pose update, the vehicle model maintains a continuous closed-loop evolution throughout the entire evaluation cycle.

[0044] Specifically, after generating the transformed perception data for the next moment, the predicted trajectory generation process can be performed again based on this transformed perception data. Subsequently, the vehicle model updates its pose according to the new predicted trajectory, and uses this updated pose as the starting point for the next calculation. By repeatedly executing this closed-loop feedback chain of "perception data transformation - trajectory generation - pose update," the vehicle model continuously evolves throughout the entire evaluation cycle, maintaining a dynamic coupling between its motion state and the perception perspective. Thus, a continuous feedback mechanism is achieved in the vehicle model evaluation process based on recorded data, enabling an approximate simulation of continuous driving behavior on real roads.

[0045] This application provides a closed-loop evaluation method that is closer to real-vehicle testing than open-loop evaluation, while eliminating the high cost of 3D reconstruction. It enables rapid and low-cost testing of a large number of scenarios, balancing efficiency and cost.

[0046] The description of the performance evaluation indicators can be based on various non-restrictive indicator parameter types to comprehensively reflect the safety, stability and control accuracy of the intelligent driving system during closed-loop operation.

[0047] In some implementations, obstacle location information is extracted based on the pose sequence of the simulated vehicle model and the corresponding timestamp transform perception data in the dynamic evaluation data. Specifically, by traversing the transform perception data of each time frame, obstacles that have potential interaction relationships with the vehicle model are extracted, and their motion states are described in a unified spatiotemporal coordinate system.

[0048] Calculate the minimum collision time or minimum safe distance based on the relative motion vector between the pose sequence and the obstacle position.

[0049] Specifically, after obtaining the pose sequence of the simulated vehicle model and the state of the obstacles, the system uses the vehicle model as a reference coordinate system to calculate the relative position vector and relative velocity vector between the vehicle and the obstacles. Let the vehicle model at time... The position is The location of the obstacle is The speeds of the two are respectively and Then the relative position vector and the relative velocity vector can be expressed as: Equation (1) The relative motion vector in equation (1) describes the spatial relationship and motion trend between the obstacle and the vehicle.

[0050] In a continuous time series, the predicted time when the vehicle and the obstacle's potential trajectories intersect in the future is calculated based on their relative positions and speeds, i.e., the time to collision (TTC), to determine the existence of potential collision risks.

[0051] In addition to time metrics, a spatial safety margin, namely the minimum safe distance (MSD), can also be calculated to reflect the minimum spatial separation between the vehicle model and obstacles throughout the entire evaluation period. For example, position points at various times can be sampled along the vehicle's trajectory, and the Euclidean distance to the center point or envelope boundary of the obstacle can be calculated.

[0052] This application's embodiments calculate the minimum collision time (TTC) and minimum safe distance (MSD), transforming abstract driving risks into measurable numerical indicators. The TTC reflects the system's reaction speed and predictive ability to potential risks in a dynamic environment. A higher TTC indicates that the intelligent driving system can perceive risks in advance and take reasonable avoidance measures, while a lower TTC suggests decision delays or insufficient control execution. The MSD reflects the spatial safety margin of the intelligent driving system during continuous control. If the MSD remains stable and above the safety threshold in multiple scenarios, it indicates effective coordination between the perception, planning, and control modules; conversely, it may indicate systemic problems such as control jitter or target recognition errors. This allows for a comprehensive evaluation of the safety decision-making performance of the intelligent driving system.

[0053] In another example, kinematic analysis is performed on the pose sequence of the simulated vehicle model in the dynamic evaluation data to extract the rate of change of acceleration.

[0054] Specifically, the pose sequence information of the simulated vehicle model is extracted from the dynamic evaluation data. The pose sequence records the vehicle model's position coordinates and attitude parameters at consecutive time points, including position, heading angle, and velocity vector. By calculating the position change between consecutive time points, the vehicle's velocity sequence in the longitudinal direction can be obtained. Equation (2) in, Let be the time interval between adjacent frames. Further differentiating the velocity sequence yields the acceleration sequence: Equation (3) To ensure the smoothness and noise resistance of the calculation results, the velocity signal can be filtered (such as by Kalman filtering or moving average method) before acceleration calculation, thereby reducing fluctuations caused by sensor sampling errors or simulation integration errors.

[0055] After obtaining the acceleration sequence, the acceleration signal can be differentiated again to calculate the rate of change of acceleration. : Equation (4) The rate of change of acceleration reflects how quickly a vehicle's acceleration changes per unit time.

[0056] Then, the extracted rate of acceleration change is compared with a preset acceleration change threshold to monitor emergency braking or acceleration events per unit time.

[0057] In some implementations, the acceleration change threshold can be configured according to vehicle type, test conditions, or evaluation objectives. For example, the acceleration threshold for passenger cars can be higher than that for heavy vehicles to ensure safety. When the rate of change of acceleration at any given time exceeds the acceleration change threshold (such as the threshold for a rapid acceleration event or a sudden braking event), it is determined that the vehicle model exhibits a sudden change in behavior during that time period.

[0058] By extracting the rate of change of acceleration and comparing it with a preset threshold, the longitudinal acceleration and deceleration behavior of the vehicle model can be quantitatively evaluated, reflecting abnormal fluctuations in the system at the decision-making or control execution layer. This can identify potential control instability or decision non-smoothness caused by intelligent driving strategies during the simulation phase.

[0059] Figure 2 A flowchart illustrating an example of generating transformation perception data corresponding to the viewpoint of a vehicle model according to an embodiment of this application is shown.

[0060] In step S210, the spatial transformation parameters corresponding to the pose difference are calculated. The spatial transformation parameters include the translation vector and the rotation transformation matrix.

[0061] In some implementations, based on the recorded vehicle pose With the current pose of the vehicle model Calculate the pose difference between the two coordinate systems. The pose difference reflects the difference in spatial position and planar orientation between the two coordinate systems. To achieve accurate spatial mapping, the pose difference can be decomposed into translation vectors. With rotation transformation matrix .

[0062] The translation vector describes the relative displacement of the two coordinate systems in three-dimensional space, while the rotation transformation matrix describes the relative rotation of the two coordinate systems in the attitude direction. In actual calculations, the rotation matrix can be obtained through quaternion transformations, Euler angle transformations, or exponential mappings using Lie groups and Lie algebras to ensure the numerical stability and continuity of the calculations. Therefore, by using the spatial transformation parameters, a high-precision spatial mapping relationship can be established between the recorded vehicle coordinate system and the vehicle model coordinate system.

[0063] In step S220, the target object coordinates in the environmental perception data are rigidly translated according to the translation vector to generate intermediate transformation data.

[0064] In some implementations, the obtained translation vector is used A rigid translation operation is performed on the coordinates of all targets in the recorded environmental perception data to eliminate the spatial differences between the vehicle model and the recorded vehicle, aligning their origin positions.

[0065] Specifically, for each perceived target's three-dimensional coordinate point Applying a translation transformation, we obtain the intermediate transformation result: Equation (5) In some cases, perception data may originate from different sensors (e.g., point clouds, images, or fused detection results), therefore translation operations need to be performed in a unified world coordinate system or vehicle coordinate system. To ensure data consistency, the outputs of each sensor can be converted to a unified coordinate system based on the vehicle calibration matrix before performing the translation.

[0066] This translation operation aligns the spatial distribution of the recorded perception data with the current spatial position of the vehicle model, thereby eliminating positional errors and enabling the vehicle model to perceive the relative spatial relationships of objects within the same scene geometry.

[0067] In step S230, the intermediate transformation data is rotated and aligned according to the rotation transformation matrix, and transformation perception data adapted to the vehicle model's perspective is constructed based on the spatial relationship of the target objects after rotation and alignment.

[0068] Specifically, using the rotation transformation matrix Rotate the target object's coordinates to align the perception direction of the recording vehicle with the observation direction of the vehicle model. The rotation operation can be performed using the following formula: Equation (6) After rotation and alignment, environmental perception data adapted to the vehicle model's perspective can be reconstructed based on the adjusted spatial relationships. In some examples, the environmental perception data not only includes the spatial location information of the target objects but also retains the corresponding semantic attributes, motion information (such as velocity vectors and heading angles), and relative spatial relationships with lane structures and traffic signals. Therefore, the perception input received by the vehicle model during evaluation will accurately correspond to the real visual environment under its own position and orientation, completing a full spatial transformation from the actual vehicle recording perspective to the vehicle model's perspective, ensuring a perfect match between the perception data and the vehicle model's orientation and perspective.

[0069] Figure 3 A flowchart illustrating an example of iteratively updating the pose of a simulated vehicle model according to an embodiment of this application is shown.

[0070] like Figure 3 As shown, in step S310, at least one initial trajectory is predicted based on the transformation sensing data.

[0071] In some implementations, the planning and control model (or decision and planning module) of the intelligent driving system is used to predict the possible driving trajectory of the vehicle model. The planning and control model used for trajectory prediction can be diverse, such as rule-based path planning algorithms or machine learning-based behavior prediction models, etc., and should not be limited here, and can be adjusted according to the evaluation requirements of the intelligent driving system. In some implementations, multiple initial trajectories are generated based on the perceived environment from the current vehicle model's perspective. Each initial trajectory corresponds to a different driving behavior intention (e.g., keeping the lane, changing lanes to overtake, slowing down to follow, or avoiding obstacles, etc.) or driving preference (e.g., lower highway tolls, smoother roads, or shorter distances, etc.).

[0072] In step S320, a collision risk assessment and / or comfort assessment are performed on each initial trajectory to determine the optimized predicted trajectory.

[0073] Here, safety and comfort assessments can be combined to comprehensively evaluate multiple initial predicted trajectories in order to determine the optimal trajectory.

[0074] For example, in collision risk assessment, safety indicators such as the spatial distance between the predicted trajectory and surrounding traffic participants, collision response time, and minimum relative distance can be calculated based on the dynamic target information contained in the transformation sensing data. If the minimum distance between a trajectory and any target is lower than a set threshold within the planning period, the trajectory is determined to have a potential collision risk.

[0075] In comfort assessment, dynamic ride comfort indices of the trajectory can be calculated, including longitudinal acceleration, lateral acceleration, and angular velocity, to ensure that the trajectory provides a good riding experience while meeting safety requirements.

[0076] By conducting safety and comfort assessments on multiple candidate trajectories, the selected trajectory was ensured to be both safe and feasible while also taking into account vehicle control stability and passenger experience, thus truly reflecting the comprehensive decision-making capabilities of the intelligent driving system.

[0077] In step S330, the optimized trajectory is converted into vehicle control commands to drive iterative updates of the vehicle model pose.

[0078] Here, the trajectory transformation process is based on the discretization solution of the vehicle dynamics model or control model. Specifically, the trajectory is decomposed into several time-series control parameters, including longitudinal control variables (desired speed, acceleration, throttle / brake opening) and lateral control variables (desired steering angle, steering rate, etc.). These parameters are input to the controller module of the vehicle model, and the vehicle's state variables at the next moment, i.e., the updated vehicle model pose, are calculated through the vehicle kinematic equations or a dynamics simulation engine. Thus, by transforming the optimized trajectory into vehicle control commands and updating the vehicle model pose in real time, the dynamic evolution of the vehicle model in closed-loop evaluation is realized.

[0079] In some examples of embodiments of this application, transform perception data is sent to the planning and control model, enabling the planning and control model to generate a predicted trajectory based on the transform perception data. Specifically, in the evaluation process, the generated transform perception data is sent to the planning and control model. This transform perception data includes environmental information from the current perspective of the vehicle model, such as lane topology, the position and motion state of dynamic objects, traffic signal status, road boundaries, and the distribution of static obstacles.

[0080] Here, to achieve efficient data interaction between the planning and control module and the evaluation master control system, a service interface mechanism based on a remote communication protocol is adopted. Specifically, a standardized communication channel can be established between the evaluation platform and the planning and control model, supporting communication methods such as HTTP, gRPC, WebSocket, or ROS services. By defining a unified message format (such as JSON, Protobuf, or ROS messages) within the system, it is ensured that different algorithm models can receive sensing input under a unified data structure. This allows the evaluation master control terminal to proactively encapsulate the transformed sensing data into network messages and send them to the target service address according to the current test configuration, thereby triggering trajectory generation calculations of the remote planning and control model.

[0081] In some implementations, remote communication is established between the service address and the planning and control model to send change-aware data to the model. During the evaluation process, different service addresses are switched to dynamically and hot-swap the planning and control model. Specifically, a remote communication connection is established between the service address and the planning and control model. Each planning and control model corresponds to an independent service instance, and its communication entry point can be identified by a network port, IP address, or container service identifier. When the evaluation platform needs to switch between different versions or types of planning algorithms, it only needs to dynamically change the service address during operation to achieve hot-swap without interrupting the evaluation process.

[0082] For example, during a certain evaluation, the current service address can be switched from model_v1.service.local to model_v2.service.local via configuration file or runtime instructions. The evaluation framework automatically re-establishes the communication channel and sends transformation-aware data to the new model instance. During the switching process, the main control system maintains temporal continuity and vehicle model state caching, thereby achieving seamless replacement and online comparison of algorithm modules.

[0083] Then, the predicted trajectory is received from the planning and control model, and the pose of the simulated vehicle model is iteratively updated based on the predicted trajectory.

[0084] In some implementations, after receiving trajectory results from a remote model, the evaluation system analyzes parameters such as pose, velocity, acceleration, and steering angle, and uses these as control inputs for the vehicle model. A new vehicle model pose is then calculated using the vehicle dynamics model. This updated pose is written to the simulation state cache, and the pose difference and transformed sensing data are recalculated at the next time step, thus completing a full closed-loop feedback cycle.

[0085] Through the service address-based remote communication mechanism provided in this application, the planning and control model is no longer fixed to the local simulation platform, but can be independently deployed on a remote server, container environment, or cloud cluster. This significantly improves the modularity and resource utilization of the evaluation system, allowing different algorithm teams to independently debug and update the model without affecting the overall evaluation framework. Furthermore, by dynamically switching service addresses during the evaluation process, online hot-swapping of the planning and control model is achieved, enabling the replacement of algorithm modules without stopping or restarting the simulation. This greatly improves testing efficiency and is particularly suitable for regression verification, A / B testing, or performance comparison analysis scenarios involving multiple algorithm versions.

[0086] In some examples of embodiments of this application, behavioral event annotation information is configured for environmental perception data recorded by a real vehicle. This behavioral event annotation information includes timestamped driving behavior events, which describe the driving actions of the recorded vehicle at different times, thereby further improving the accuracy of trajectory prediction. By introducing behavioral event information into the trajectory prediction process, the behavioral intentions of a real driver in a specific environment can be reproduced in the time dimension, thereby enhancing the semantic consistency and dynamic rationality of trajectory prediction.

[0087] In some implementations, during the real vehicle data acquisition or post-processing stage, the recorded environmental perception data is labeled with behavioral events. The labeling information includes the type of driving behavior event and its corresponding timestamp. For example, driving behavior events include any one or more of the following: lane change, braking, acceleration, steering, or following distance adjustment.

[0088] In one example, event labeling information can be obtained with human assistance, such as manually labeling or algorithmically labeling collected driving videos and perception data to identify the time period and event type corresponding to the driver's behavior. In another example, event labeling information can be automatically extracted based on real vehicle control signals, such as extracting signals like steering angle, throttle opening, braking force, and target speed from the vehicle bus (CAN) or control records, and automatically detecting driving behavior events using threshold judgment or pattern recognition algorithms.

[0089] Furthermore, when the timestamps of the system clock information and the behavioral event annotation information are detected to match, a predicted trajectory is generated based on the corresponding driving behavior events in the transformation perception data and the behavioral event annotation information.

[0090] Here, when the system detects a match between the current system clock and the timestamp of a certain behavioral event, the system generates a predicted trajectory with a specific driving intention based on the transform perception data and the corresponding behavioral event type. For example, when the detected behavioral event is "lane change," the system plans a lateral translation trajectory that crosses adjacent lane lines based on the lane topology information in the transform perception data, and generates a safe and feasible lane change trajectory by considering the speed and spacing constraints of surrounding target vehicles; when the behavioral event is "braking," the system generates a longitudinal deceleration trajectory with a desired deceleration distribution based on the status of obstacles or traffic signals ahead in the transform perception data, in order to achieve smooth braking behavior.

[0091] By incorporating timestamped behavioral event annotation information into trajectory prediction through the embodiments of this application, it is possible to generate predictions that are consistent with the actual driver's intent in different scenarios. Figure 1 The predictive trajectory is more consistent with the actual driving behavior, rather than being generated solely based on environmental geometry or physical constraints. The generated predictive trajectory has stronger semantic interpretability and consistency with real driving.

[0092] In some examples of embodiments of this application, the environmental perception data recorded by the real vehicle and the corresponding recorded vehicle pose originate from scene fragment data packets. These scene fragment data packets are generated by splitting the recorded scene of the real vehicle to be evaluated, and each scene fragment data packet has a corresponding scene fragment identifier. Through this structured scene management mechanism, segmented processing and independent evaluation of large-scale real road scenes can be achieved. Furthermore, the evaluation performance indicators and scene fragment identifiers corresponding to the scene fragment data packets can be sent to the evaluation management terminal, enabling the evaluation management terminal to aggregate the evaluation performance indicators corresponding to all scene fragment data packets, thereby achieving automated, distributed closed-loop performance statistics and performance attribution.

[0093] In this embodiment, relying on distributed cloud computing tasks, a large number of scene fragment data packets can be evaluated in parallel, thereby achieving high-efficiency and low-cost closed-loop verification of massive scenes. Specifically, the closed-loop evaluation does not target a single driving scenario, but simultaneously covers tens of thousands of real-vehicle recorded scene fragments. To complete such a large-scale evaluation task within a reasonable time cost, a distributed computing architecture can be adopted, distributing the evaluation tasks of different scene fragments to multiple computing nodes for parallel execution. Through the dynamic scheduling mechanism of elastic computing resources, computing instances can be automatically expanded or released according to the task load, enabling each node to complete the evaluation calculation of more scenarios in parallel within the same time. This not only significantly improves the evaluation throughput and resource utilization, but also effectively reduces the overall testing time and hardware costs, thereby constructing a distributed intelligent driving evaluation system with high scalability and high concurrency capabilities.

[0094] Figure 4 A flowchart illustrating another example of an intelligent driving evaluation method according to an embodiment of this application is shown.

[0095] like Figure 4 As shown, this demonstrates the complete process of achieving closed-loop evaluation of the autonomous driving planning and control model through data transformation based on real vehicle recording data. The entire process can complete dynamic closed-loop inference and performance verification based on real data without the need for 3D scene reconstruction.

[0096] At the beginning of the evaluation phase, environmental perception data and corresponding vehicle pose information recorded by the actual vehicle are read from the recorded dataset. The recorded data includes structured information collected by the vehicle in the actual road environment, such as obstacle positions, lane topology, road boundaries, traffic signals, and target vehicle status. To enable the planning and control model to correctly parse the input information, the model can first undergo input-output adaptation by converting the actual vehicle recorded data into the data format required for model inference. This format can adopt a unified structured data description (such as TDR format). An aggregation script parses, extracts, and encapsulates the raw perception data into standard input, thereby ensuring the consistency between the model inference data and the closed-loop process data in the spatiotemporal dimensions.

[0097] During the data transformation phase, based on the pose difference between the current closed-loop vehicle model's pose and the recorded vehicle's pose, corresponding spatial transformation parameters are calculated, and these parameters are used to perform real-time geometric transformations on the recorded environmental perception data. Through translation and rotation mapping, scene elements in the original recorded data are transformed from the recording vehicle's perspective to the closed-loop vehicle's perspective, ensuring that the closed-loop system receives perceptual input consistent with its own motion state at every moment. This pose-difference-based data transformation mechanism allows for dynamic reconstruction of the real environment without the need to construct a virtual 3D scene during the evaluation process.

[0098] The transformed perception data is input into the planning and control model to generate the vehicle's predicted trajectory. The planning and control model typically comprises three parts: model prediction, safety fallback, and control. The model prediction module calculates the expected driving trajectory over a future period based on the current perception data; the safety fallback program performs risk assessment and constraint correction on this trajectory to ensure its executability and safety; and the control module translates the optimized trajectory into vehicle control commands, such as target speed, steering angle, and acceleration, to drive the vehicle model's motion simulation.

[0099] Based on the predicted trajectory output by the planning and control model and combined with the set planning frequency, motion calculations are performed on the closed-loop vehicle model, and its pose information is updated. The updated vehicle pose is recorded and fed back to the data transformation module for perspective correction of the perception data in the next moment, thus forming a continuous feedback loop. Throughout the entire evaluation cycle, the state of the closed-loop vehicle continues to evolve, and its dynamic pose always affects the generation of the next round of perception data, thereby realistically simulating the behavioral logic of the autonomous driving system in the continuous control process.

[0100] In implementation, the planning and control model is encapsulated as an independent network inference service. The closed-loop evaluation process communicates with this service through a pre-defined network protocol. The planning and control model can be hot-swapped by switching service addresses according to different testing needs, allowing for dynamic replacement of different versions or strategies during the evaluation process. Algorithm comparison and performance regression testing can be completed without re-integrating the entire system. This model-as-a-service and hot-swapping mechanism significantly improves the flexibility of the evaluation and the maintainability of the engineering.

[0101] During the closed-loop execution process, key operational data can be continuously recorded, including vehicle pose sequence, predicted trajectory, control output, and safety assessment results. This dynamic evaluation data will be used to calculate performance indicators, such as trajectory deviation, minimum collision time, and longitudinal acceleration change rate, after the evaluation is completed, to quantify the system's performance in terms of safety, ride comfort, and stability.

[0102] The intelligent driving closed-loop evaluation method provided in this application, based on pose difference data transformation driven by real vehicle recording data, achieves high-precision closed-loop verification of the autonomous driving planning and control model without 3D scene reconstruction. This method, while ensuring evaluation authenticity, possesses advantages such as high efficiency, low computational cost, and strong scalability. It can support rapid iterative evaluation of multiple models and versions, providing a high-fidelity evaluation environment for the verification and optimization of intelligent driving algorithms.

[0103] It should be noted that while traditional open-loop evaluation methods offer high testing efficiency and can cover a large number of scenarios in a short time, the lack of vehicle dynamic feedback during testing often results in evaluations that significantly differ from real-world road test performance. This makes it difficult to accurately reflect the system's performance during continuous control, thus limiting their effectiveness. In contrast, closed-loop evaluation methods based on scene reconstruction can achieve simulations closer to real-world vehicle performance, but their implementation relies on complex link adaptation and high-precision scene modeling. This typically requires modelers to spend weeks digitally reconstructing street and traffic elements, resulting in high evaluation costs and hindering large-scale deployment. The evaluation method provided in this application, however, eliminates the need for 3D scene reconstruction and can directly utilize environmental perception data recorded by the real vehicle for closed-loop evaluation, thereby achieving evaluation results highly consistent with real-world road test results at a lower cost.

[0104] Furthermore, to further improve the accuracy and practicality of closed-loop evaluation, this application also proposes to introduce a short-cycle evaluation and precise labeling mechanism into the evaluation system, and establish a multi-dimensional performance index system to more comprehensively reflect the overall capabilities of the intelligent driving system while ensuring the reliability of the evaluation.

[0105] Regarding the annotation mechanism, it supports more refined behavioral event annotations to control specific actions of intelligent driving vehicles in closed-loop testing. For example, control annotations with precise timestamps can be added to the recorded data to trigger corresponding driving commands during the evaluation process, such as "sending a lane change signal at time t" or "executing an emergency braking command." In this way, real driving behavior events can be accurately reproduced in closed-loop evaluation, thereby improving the temporal consistency and behavioral controllability of the evaluation.

[0106] To address the limitations caused by the lack of reconstruction of perception data, a time window constraint mechanism is proposed. Since the perception data in the closed-loop evaluation process comes from the raw data recorded by the actual vehicle, when the pose of the closed-loop vehicle deviates significantly from that of the recorded vehicle, data transformation may lead to perception distortion, thus affecting the accuracy of the evaluation. Therefore, in this embodiment, the duration of a single closed-loop evaluation is limited to a short interval (e.g., 10 seconds). Within this time window, the spatial deviation between the closed-loop vehicle and the recorded vehicle remains within a reasonable range, effectively controlling perception errors. This time length has been verified to be sufficient to cover most key driving scenarios, while ensuring the authenticity of the evaluation data and the reliability of the results.

[0107] Regarding performance evaluation metrics, this application constructs a multi-dimensional indicator system to comprehensively evaluate autonomous driving algorithms from multiple perspectives, including safety, comfort, and system stability. Safety indicators may include collision rate and minimum collision time, used to measure the system's risk response capability; comfort indicators may include emergency braking rate, gradual braking rate, and rapid acceleration rate, used to evaluate the smoothness of the system's control behavior and the passenger experience. Through comprehensive analysis of indicators across different dimensions, the performance of the algorithm in real-world traffic scenarios can be more fully reflected.

[0108] For example, in a typical scenario recorded in a real vehicle, a vehicle is entering a highway merging point when a high-speed vehicle is approaching from behind. The original driver failed to slow down in time, resulting in a collision. Directly reproducing this scenario in a real vehicle to verify the algorithm's repair effect would pose a high risk. However, the closed-loop evaluation method of this application allows for the safe reproduction of this scenario in a software simulation environment, verifying whether the planning and control model can correctly perceive the approaching vehicle and trigger braking commands when danger approaches, thereby avoiding a collision. Such test results can be directly correlated with safety indicators (such as collision rate and minimum collision time), providing a quantitative basis for algorithm optimization.

[0109] Therefore, by introducing short-cycle evaluation and precise labeling mechanisms, perception errors are controlled in the spatiotemporal dimensions to ensure the accuracy of evaluation results. Furthermore, a multi-dimensional indicator system is used to achieve a comprehensive performance evaluation of the intelligent driving system, providing a scientific and reproducible technical means for the safety verification and performance improvement of the intelligent driving system.

[0110] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of combined actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Secondly, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0111] In some embodiments, this application also provides a computer program product, the computer program product including a computer program stored on a non-volatile computer-readable storage medium, the computer program including program instructions, which, when executed by a computer, cause the computer to perform any of the above-described intelligent driving evaluation methods.

[0112] In some embodiments, this application also provides an electronic device, which includes: at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform an intelligent driving evaluation method.

[0113] The apparatus described in the embodiments of this application can be used to execute the intelligent driving evaluation method of the embodiments of this application, and accordingly achieve the technical effects achieved by the intelligent driving evaluation method of the embodiments of this application, which will not be elaborated further here. In the embodiments of this application, the relevant functional modules can be implemented by a hardware processor.

[0114] Figure 5 This is a schematic diagram of the hardware structure of an electronic device for executing an intelligent driving evaluation method according to another embodiment of this application, as shown below. Figure 5 As shown, the device includes: One or more processors 510 and memory 520, Figure 5 Take the 510 processor as an example.

[0115] The device for performing the intelligent driving evaluation method may also include: an input device 530 and an output device 540.

[0116] The processor 510, memory 520, input device 530, and output device 540 can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0117] The memory 520, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the intelligent driving evaluation method in the embodiments of this application. The processor 510 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 520, thereby implementing the intelligent driving evaluation method of the above-described method embodiments.

[0118] The memory 520 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the device, etc. Furthermore, the memory 520 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 520 may optionally include memory remotely located relative to the processor 510, and these remote memories may be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0119] Input device 530 can receive input digital or character information and generate signals related to user settings and function control of the device. Output device 540 may include display devices such as a display screen.

[0120] The one or more modules are stored in the memory 520, and when executed by the one or more processors 510, they execute the intelligent driving evaluation method in any of the above method embodiments.

[0121] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0122] The electronic devices in this application embodiments exist in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include: smartphones (e.g., iPhones), multimedia phones, feature phones, and low-end phones, etc.

[0123] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, such as the iPad.

[0124] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes audio and video players (such as iPods), handheld game consoles, e-book readers, as well as smart toys and portable car navigation devices.

[0125] (4) Server: A device that provides computing services. The components of a server include a processor, hard disk, memory, system bus, etc. Servers are similar to general computer architectures, but because they need to provide highly reliable services, they have higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0126] (5) Other electronic devices with data interaction functions.

[0127] In some embodiments, this application also provides a mobile platform on which the computer device described in any embodiment of this application is installed. The mobile platform includes, but is not limited to, vehicles, tracked robots, bipedal robots, quadrupedal robots, etc., wherein the vehicle can be a passenger car, pickup truck, truck, etc. It should be noted that the above are merely examples, and this application does not limit the specific form of the mobile platform.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for evaluating intelligent driving, comprising: Acquire environmental perception data and corresponding vehicle pose recorded from the actual vehicle; The environmental perception data is transformed based on the pose difference between the vehicle model pose and the recorded vehicle pose to generate transformed perception data from the corresponding vehicle model perspective. A predicted trajectory is generated based on the transformation sensing data, and the vehicle model pose is iteratively updated according to the predicted trajectory, while recording dynamic evaluation data; the dynamic evaluation data includes the iteratively updated vehicle model pose, transformation sensing data, and predicted trajectory. Based on the dynamic evaluation data, the evaluation performance indicators of the intelligent driving system are determined.

2. The method according to claim 1, wherein, After iteratively updating the vehicle model pose based on the predicted trajectory, the method further includes: The updated vehicle model pose is determined as the vehicle model pose at the next moment, and the pose difference between the vehicle model pose at the next moment and the recorded vehicle pose at the next moment is recalculated. Based on the recalculated pose difference, perform data transformation on the environmental perception data at the next moment to generate transformed perception data corresponding to the next moment. In this process, by repeatedly executing a continuous feedback loop of data transformation, predicted trajectory generation, and vehicle model pose update, the vehicle model maintains a continuous closed-loop evolution throughout the entire evaluation cycle.

3. The method according to claim 1, wherein, The step of transforming the environmental perception data based on the pose difference between the vehicle model pose and the recorded vehicle pose to generate transformed perception data corresponding to the vehicle model's perspective includes: Calculate the spatial transformation parameters corresponding to the pose difference, wherein the spatial transformation parameters include a translation vector and a rotation transformation matrix; The target object coordinates in the environmental perception data are rigidly translated according to the translation vector to generate intermediate transformation data. The intermediate transformation data is rotated and aligned according to the rotation transformation matrix, and transformation perception data adapted to the perspective of the vehicle model is constructed based on the spatial relationship of the target objects after rotation and alignment.

4. The method according to claim 1, wherein, The step of generating a predicted trajectory based on the transformed sensing data and iteratively updating the pose of the simulated vehicle model according to the predicted trajectory includes: Based on the transformed sensing data, predict at least one corresponding initial trajectory; Collision risk assessment and / or comfort assessment are performed on each initial trajectory to determine the optimized predicted trajectory; The optimized predicted trajectory is converted into vehicle control commands to drive iterative updates of the vehicle model pose.

5. The method according to claim 1 or 4, wherein, The method further includes: Configure behavioral event annotation information for the environmental perception data recorded by the real vehicle, the behavioral event annotation information including driving behavior events with timestamps; The generation of the predicted trajectory based on the transformation sensing data includes: When the timestamps of the system clock information and the behavior event annotation information are detected to match, a predicted trajectory is generated based on the corresponding driving behavior events in the transformed perception data and the behavior event annotation information; the driving behavior events include any one or more of the following: lane change, braking, acceleration, steering, or following distance adjustment.

6. The method according to claim 1, wherein, The step of generating a predicted trajectory based on the transformed sensing data and iteratively updating the pose of the simulated vehicle model according to the predicted trajectory includes: The transformation sensing data is sent to the planning and control model, so that the planning and control model generates a predicted trajectory based on the transformation sensing data; The predicted trajectory is received from the planning and control model, and the pose of the simulated vehicle model is iteratively updated based on the predicted trajectory.

7. The method according to claim 6, wherein, Sending the transformation sensing data to the planning and control model includes: Remote communication is established between the service address and the planning and control model to send the transformation sensing data to the planning and control model; In the evaluation process, different service addresses are switched to dynamically hot-swap the planning and control model.

8. The method according to claim 1, wherein, The process of determining the performance indicators of the intelligent driving system based on the dynamic evaluation data includes: Based on the pose sequence of the simulated vehicle model and the transformation perception data with corresponding timestamps in the dynamic evaluation data, obstacle position information is extracted; Based on the relative motion vector between the pose sequence and the obstacle position, calculate the minimum collision time or the minimum safe distance; the minimum collision time is the shortest predicted time required for the simulated vehicle model to collide with the obstacle, and the minimum safe distance is the minimum spatial interval between the simulated vehicle model and the obstacle.

9. The method according to claim 1, wherein, The process of determining the performance indicators of the intelligent driving system based on the dynamic evaluation data includes: Kinematic analysis was performed on the pose sequence of the simulated vehicle model in the dynamic evaluation data to extract the rate of change of acceleration; The extracted rate of acceleration change is compared with a preset acceleration change threshold to monitor sudden braking or acceleration events per unit time.

10. The method according to claim 1, wherein, The environmental perception data and corresponding recorded vehicle pose recorded by the actual vehicle originate from scene fragment data packets. These scene fragment data packets are generated by splitting the recorded scene of the actual vehicle to be evaluated, and each scene fragment data packet has a corresponding scene fragment identifier. The method further includes: The evaluation performance metrics and scene segment identifiers corresponding to the scene segment data packets are sent to the evaluation management terminal, so that the evaluation management terminal aggregates the evaluation performance metrics corresponding to all scene segment data packets.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein, The processor executes the computer program to implement the steps of the method according to any one of claims 1-10.

Images