Intelligent driving control method of vehicle and vehicle

Abstract

The invention relates to an intelligent driving control method of a vehicle and the vehicle, and belongs to the technical field of intelligent driving, and the method comprises the following steps: obtaining an original data stream of a vehicle sensor network, and carrying out feature extraction on a dynamic feature vector; state vectors of the traffic participants are extracted according to the dynamic feature vectors, edges between the two traffic participants are constructed, a traffic participant interaction graph is constructed, and an asymmetric factor matrix is calculated; acquiring a historical scene data set of the vehicle, calculating correction similarity, and calculating a weight coefficient; performing weighted fusion on the weight coefficient and the historical scene data set to obtain an enhanced data set; and performing updating training on the pre-trained intelligent driving decision model according to the enhanced data set to obtain an updated intelligent driving decision model, generating an intelligent driving decision, and controlling vehicle operation, thereby realizing more accurate characterization of a dynamic game relationship in a complex traffic scene, and improving the accuracy of the dynamic game relationship. And the intelligent driving decision model can continuously adapt to an asymmetric interaction mode effect in a real scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving, and in particular to an intelligent driving control method of a vehicle and the vehicle. BACKGROUND

[0002] With the evolution of intelligent driving technology to high-order automatic driving, the intelligent driving decision of the vehicle faces unprecedented complexity, and the data-driven decision model becomes the key to realizing safe and human-like driving. In this process, the vehicle middleware, as a bridge connecting the bottom layer of heterogeneous hardware and the upper layer of application algorithms, is responsible for the synchronization, scheduling and communication of massive multi-source sensor data, providing stable and efficient data flow support for the optimization of the decision model.

[0003] Currently, the iterative optimization of driving decisions is usually achieved by collecting scene data from mass-produced vehicles and returning it to the cloud, using traditional dynamic time warping algorithms to measure the similarity between the current scene and the historical scene library, fusing historical data based on similarity weights to construct an enhanced data set, and finally driving the incremental update of the decision model.

[0004] In the above prior art, the scene similarity measurement method based on the traditional dynamic time warping algorithm cannot accurately represent the asymmetric interaction relationship between traffic participants, resulting in distorted scene matching, similarity calculation that does not match the real road interaction logic, and insufficient reliability of the driving decision model in complex interaction scenarios, making it difficult to meet the safety needs of high-order automatic driving. SUMMARY

[0005] The present application at least partly solves one of the problems in the related art.

[0006] To this end, the present application aims to provide an intelligent driving control method of a vehicle and the vehicle, which quantifies the asymmetry between traffic participants, filters historical scene data sets, and performs incremental training on pre-trained intelligent driving decision models to generate intelligent driving decisions and control the operation of the vehicle.

[0007] To achieve the above-mentioned purpose, the first aspect of the present application provides an intelligent driving control method of a vehicle and the vehicle, comprising: obtaining an original data stream of a vehicle sensor network, and performing feature extraction on the original data stream to obtain a dynamic feature vector; extracting a state vector of a traffic participant according to the dynamic feature vector, constructing an edge between any two traffic participants according to the state vector, constructing a traffic participant interaction graph according to the state vector and the edge, and calculating an asymmetric factor matrix; wherein the asymmetric factor matrix is used to represent the asymmetric strength of the interaction influence between any two traffic participants; obtain a historical scene data set of the vehicle, calculate a revised similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculate a weight coefficient according to the revised similarity; weight fuse the weight coefficient and the historical scene data set to obtain an enhanced data set; update train a pre-trained intelligent driving decision model according to the enhanced data set to obtain an updated intelligent driving decision model, generate an intelligent driving decision according to the updated intelligent driving decision model, and control the vehicle to run.

[0008] In the technical solution, the traditional rule-based decision logic is difficult to cover all long-tail scenarios, and a data-driven decision model becomes the key to realizing safe and humanized driving. However, the traditional dynamic time warping algorithm cannot accurately depict the asymmetric interaction game relationship between traffic participants, and the traffic participants are regarded as independent entities, ignoring the mutual influence between their decisions. The symmetry assumption of the dynamic time warping algorithm conflicts with the asymmetric nature of the real road game, resulting in the failure of the entire data closed loop in key complex scenarios, and it is difficult to meet the stringent requirements of high-level autonomous driving on safety and reliability.

[0009] And through the above-mentioned scheme, the application obtains the original data stream of the vehicle sensor network and extracts the dynamic feature vector, constructs the interaction graph of the traffic participants and calculates the asymmetric factor matrix, calculates the revised similarity and the weight coefficient based on the asymmetric factor matrix and the historical scene data set, realizes the accurate screening and weighted fusion of the historical data, and generates the enhanced data set. Therefore, compared with the prior art, the asymmetric strength of the interaction influence between the traffic participants can be quantified, so that the dynamic game relationship in the complex traffic scene can be more accurately represented, and the intelligent driving decision model can adapt to the asymmetric interaction mode in the real scene.

[0010] On this basis, the pre-trained decision model is incrementally optimized through the enhanced data set, and compared with the assumption of symmetry and monotonic continuous path in the prior art, the safety, humanization and reliability of the intelligent driving decision in the complex game environment can be improved.

[0011] In some embodiments of the application, the method for extracting features from the original data stream to obtain the dynamic feature vector includes: perform multi-modal fusion on the vehicle operating environment data in the original data stream through a space-time encoder to obtain an environment feature matrix; perform segmentation processing on the control instruction data in the original data stream through a pre-set length sliding window to extract a statistical feature vector; record message latency of middleware running data in the original data stream and calculate central processing unit occupancy, combine the message latency and the central processing unit occupancy to build a system load vector; combine the environment feature matrix, the statistical feature vector and the system load vector to obtain the dynamic feature vector.

[0012] In the technical scheme, the vehicle running environment data is fused by the space-time encoder, which can unify the feature representation of different sensor data, and the spatial structure and time sequence correlation of the environment are retained; The statistical feature vector is extracted by segmenting the control instruction data by a sliding window, so that the dynamic change trend of the vehicle motion is effectively captured; The system load vector is built by recording the message latency of the middleware running data and calculating the central processing unit occupancy, so as to reflect the running state of the computing platform in real time; Finally, the environment feature matrix, the statistical feature vector and the system load vector are combined to obtain the dynamic feature vector, which provides a comprehensive and multi-dimensional scene representation basis for subsequent interaction graph construction and similarity calculation, and improves the accuracy and robustness of subsequent decision optimization.

[0013] In some embodiments of the present application, the method for constructing an interaction graph of a traffic participant according to the state vector and the edge comprises: The target detection and tracking module is used to identify and locate the traffic participant in the environment feature matrix, and a state vector is extracted from the position coordinates, current speed, acceleration and heading angle of the traffic participant; The first Euclidean distance and relative speed between any two traffic participants are calculated according to the state vector, and when the first Euclidean distance and relative speed between the two traffic participants meet a preset judgment condition, an edge is constructed between the two traffic participants; The interaction graph of the traffic participant is constructed according to the state vector of each traffic participant and the edge.

[0014] In the technical scheme, the target detection and tracking module is used to identify and locate the environment feature matrix, and the accurate state vector of the traffic participant is extracted, including position, speed, acceleration and heading angle, which provides a fine-grained physical state input for interaction relationship modeling; By calculating the Euclidean distance and relative speed between the traffic participants and constructing the edge based on the preset judgment condition, the potential interaction relationship between the traffic participants can be accurately captured, and the interference of irrelevant objects is avoided; Finally, the interaction graph of the traffic participant is constructed based on the state vector and the edge, which abstracts the complex traffic scene into a calculable graph structure, provides a basis for subsequent asymmetric factor calculation, and improves the accuracy and effectiveness of interaction modeling.

[0015] In some embodiments of the present application, the method of calculating the asymmetric factor matrix comprises: performing node encoding on the state vector using a graph attention network, calculating the weight of the edge according to the node-encoded state vector, and obtaining the edge weight; calculating the asymmetric factor according to the edge weight corresponding to the second traffic participant in different directions, and generating the asymmetric factor matrix according to the asymmetric factor.

[0016] In the technical solution, the information of neighbor nodes can be adaptively aggregated by performing node encoding on the state vector using a graph attention network, and a node representation rich in context semantics can be generated; The weight of the edge is calculated based on the encoded node representation, and the mutual influence strength between traffic participants is accurately quantified; The asymmetric factor is obtained by calculating the difference between the edge weights in different directions, effectively depicting the directionality and asymmetry of the interaction relationship; finally, the asymmetric factor matrix is generated, providing a quantitative interaction strength basis for subsequent similarity calculation, overcoming the defects of traditional methods that ignore the asymmetry of interaction, and improving the discrimination ability and reliability of scene matching.

[0017] In some embodiments of the present application, the method of calculating the modified similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculating the weight coefficient according to the modified similarity comprises: extracting a real-time scene feature sequence of the dynamic feature vector in chronological order, and selecting a target historical scene in the historical scene data set; calculating a second Euclidean distance between the real-time scene feature sequence and the target historical scene, calculating a modified dynamic time warping distance by a dynamic time warping algorithm according to the second Euclidean distance and the asymmetric factor matrix, and obtaining the modified similarity according to the modified dynamic time warping distance; calculating the weight coefficient by a decay-type mapping function according to the modified similarity.

[0018] In the technical solution, a comparison basis is provided for similarity calculation by extracting a real-time scene feature sequence of a dynamic feature vector and a target historical scene in a historical scene data set; By calculating the Euclidean distance between the real-time scene and the historical scene, and combining the asymmetric factor matrix to modify the dynamic time warping algorithm, the difference between scenes can be accurately quantified in asymmetric interaction scenes, avoiding the matching distortion caused by the symmetric assumption of the traditional DTW algorithm; The weight coefficient is calculated by a decay-type mapping function based on the modified dynamic time warping distance, so that high-similarity historical scenes obtain higher weights, ensuring the representativeness and relevance of the enhanced data set, and improving the efficiency and effectiveness of subsequent model optimization.

[0019] In some embodiments of the present application, the method for updating and training the pre-trained intelligent driving decision-making model according to the enhanced data set comprises: calculating a knowledge distillation loss by the pre-trained intelligent driving decision-making model and the updated intelligent driving decision-making model according to the enhanced data set and preset model parameters; generating edge cases by the generative adversarial network, and calculating a decision boundary loss by the updated intelligent driving decision-making model according to the edge cases and the enhanced data set; obtaining the updated model parameters according to the knowledge distillation loss and the decision boundary loss.

[0020] In the technical solution, the knowledge distillation loss is calculated by the enhanced data set and the preset model parameters, so that the updated model retains historical knowledge while adapting to new scenarios, effectively alleviating the catastrophic forgetting problem; The edge cases are generated by the generative adversarial network, and the decision boundary loss is calculated, which actively exposes the weak links of the model under critical conditions and promotes the model to optimize the decision boundary. The updated model parameters are obtained by jointly optimizing the knowledge distillation loss and the decision boundary loss, so that the model has both the stability of historical experience and strong adaptability to new scenarios, improving the robustness and safety of intelligent driving decision-making in extreme conditions.

[0021] In some embodiments of the present application, the intelligent driving control method further comprises: performing scenario simulation on the updated intelligent driving decision-making model according to the updated model parameters, outputting a prediction result, and calculating a safety margin and a misjudgment rate according to the prediction result; comparing the safety margin or the misjudgment rate with a preset threshold, generating a feedback signal when the safety margin or the misjudgment rate meets the preset threshold condition, updating the parameters of the graph attention network, and recalculating the asymmetric factor matrix according to the updated parameters of the graph attention network.

[0022] In the technical solution, the performance of the optimized model in extreme scenarios can be quantitatively evaluated by performing scenario simulation according to the updated model parameters and calculating the safety margin and the misjudgment rate; By comparing the safety margin or the misjudgment rate with the preset threshold, it can be automatically determined whether the model meets the deployment requirements; When the indicators do not meet the conditions, a feedback signal is generated and the parameters of the graph attention network are updated, realizing closed-loop optimization of the asymmetric factor calculation module and forming an iterative mechanism for self-correction, which continuously improves the accuracy of interactive perception and ultimately ensures that the entire system can continuously evolve towards a safer and more reliable direction.

[0023] In some embodiments of the present application, the method of comparing the safety margin or the false positive rate with the size of the preset threshold value further comprises: When the safety margin or the false positive rate does not meet the preset threshold condition, a verification pass flag is generated and output.

[0024] In the technical solution, by setting the output condition of the verification pass flag, a pass signal is automatically generated when the safety margin and the false positive rate both meet the preset threshold requirement, providing an explicit and objective decision basis for model deployment; This mechanism can avoid the subjectivity and hysteresis of manual intervention, ensure that only high-quality models that have been fully verified can enter the actual application stage, thereby improving the safety and reliability of intelligent driving system upgrade, and providing reliable guarantee for the final decision of the vehicle control system.

[0025] In some embodiments of the present application, obtaining the original data stream of the vehicle sensor network comprises: Obtaining multi-source vehicle data of the vehicle sensor network, using a time synchronization module to align the unified clock of the multi-source vehicle data, and performing resampling and interpolation processing according to the timestamp to obtain the original data stream.

[0026] In the technical solution, the time synchronization module is used to align the unified clock of the multi-source vehicle data and perform resampling and interpolation processing, effectively solving the timing misalignment problem caused by inconsistent sampling frequency and transmission delay of different sensors; Generating a spatiotemporally aligned original data stream provides a consistent time reference for subsequent feature extraction and fusion, ensures the synchronization of the environment feature matrix, the control feature vector, and the system state vector in the time dimension, and lays a solid foundation for accurately constructing the traffic participant interaction graph and calculating the asymmetric factor, thereby improving the accuracy and reliability of the entire decision optimization process from the data source.

[0027] In a second aspect, the present application provides a vehicle comprising: A vehicle sensor network for collecting multi-source vehicle data and historical scene data sets; A data acquisition module in communication with the vehicle sensor network for acquiring environment perception data in the multi-source vehicle data; A middleware module in communication with the vehicle sensor network for acquiring middleware running status in the multi-source vehicle data; A vehicle control system in communication with the vehicle sensor network for acquiring real-time instruction signals in the multi-source vehicle data; A time synchronization module in communication with the vehicle sensor network for aligning the unified clock of the multi-source vehicle data, and performing resampling and interpolation processing according to the timestamp to obtain an original data stream; A time-space encoder, which is communicatively connected to the time synchronization module, is used to perform multimodal fusion on the vehicle operating environment data in the original data stream to obtain an environmental feature matrix; The target detection and tracking module is communicatively connected to the spatiotemporal encoder and is used to identify and locate traffic participants in the environmental feature matrix and extract the state vectors of the traffic participants. A processor, which is communicatively connected to the time synchronization module, the spatiotemporal encoder, and the target detection and tracking module, is configured to: Construct the edges between the two traffic participants based on the state vectors, construct the traffic participant interaction graph based on the state vectors and the edges, and calculate the asymmetric factor matrix. The corrected similarity is calculated based on the historical scene dataset, the asymmetric factor matrix, and the dynamic feature vector, and the weight coefficient is calculated based on the corrected similarity. The weighted coefficients and the historical scene dataset are weighted and fused to obtain the enhanced dataset; The pre-trained intelligent driving decision model is updated and trained based on the augmented dataset to obtain an updated intelligent driving decision model. Intelligent driving decisions are then generated based on the updated intelligent driving decision model to control vehicle operation.

[0028] In the technical solution, a complete multi-source data acquisition system is constructed through the collaborative work of the vehicle sensor network, data acquisition module, middleware module and vehicle control system; Through the step-by-step processing of the time synchronization module, the spatiotemporal encoder, and the target detection and tracking module, an efficient conversion from raw data to traffic participant state vectors is achieved. Ultimately, the processor completes the entire process of interaction graph construction, asymmetric factor calculation, similarity correction, and model optimization, materializing the innovative algorithm in the method claims into the specific functional architecture of the vehicle. This enables the vehicle itself to have the ability to continuously learn and self-optimize, realizing the engineering implementation and product protection of the intelligent driving decision optimization method.

[0029] As can be seen from the above technical solutions, additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0030] Figure 1 This is a schematic flowchart of an intelligent driving control method for a vehicle provided in an embodiment of this application.

[0031] Figure 2 This is a schematic flowchart of another intelligent driving control method for a vehicle provided in an embodiment of this application.

[0032] Figure 3 A flowchart of an intelligent driving decision model update training process provided by an embodiment of the present application.

[0033] Figure 4 A flowchart of an intelligent driving decision model optimization verification process provided by an embodiment of the present application.

[0034] Figure 5 A structural diagram of a vehicle provided by an embodiment of the present application. DETAILED DESCRIPTION

[0035] In the description of the present application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", "axial", "radial", "circumferential", and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application. In the present application, unless specifically defined and limited otherwise, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral; it can be mechanical connection, or electrical connection or communication with each other; it can be directly connected, or indirectly connected through an intermediate medium, or the internal communication of two elements or the interaction relationship between two elements, unless otherwise specifically limited. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In the present application, unless specifically defined and limited, the first feature is "on" or "under" the second feature. The first and second features can be in direct contact, or the first and second features can be in indirect contact through an intermediate medium. Moreover, the first feature "above", "above" and "above" the second feature can be directly above or obliquely above the first feature, or only indicate that the horizontal height of the first feature is higher than that of the second feature. The first feature "below", "below" and "below" the second feature can be directly below or obliquely below the first feature, or only indicate that the horizontal height of the first feature is less than that of the second feature. In this application, the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" mean that a particular feature, structure, material, or characteristic is included in at least one embodiment or example of the present application. Exemplary representations of the above terms in this specification are not necessarily directed to the same embodiment or example. Moreover, the described specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples. In addition, different embodiments or examples described in this specification and features of different embodiments or examples can be combined and combined by those skilled in the art without contradiction.

[0036] In the following, the present application will be specifically described by exemplary embodiments. However, it should be understood that the elements, structures and features in one embodiment can also be beneficially combined into other embodiments without further description. With the rapid development of automatic driving technology, the perception of drivable area plays a key role in automatic parking function, and high-precision and stable perception of drivable area can effectively improve the safety and efficiency of automatic parking function.

[0037] In the current automobile industry, intelligentization and networking are becoming the core driving force to promote industrial transformation. With the rapid development of technology, intelligent driving has gradually moved from concept to reality, becoming a key force to reshape the future travel pattern. The intelligent driving market presents the characteristics of large scale and significant proportion, and is expected to maintain high-speed growth in the future. Advanced auxiliary driving systems have reached a high penetration rate in the passenger vehicle market, especially in new energy vehicle models. At the same time, high-level automatic driving technology is expected to realize large-scale commercial application in the near future.

[0038] To realize higher level of automatic driving, efficient processing and utilization of massive data become the key. The traditional V-shaped development process, from requirement analysis to design, development, testing and verification, has been difficult to meet the dynamic needs of continuous iteration and upgrading of automatic driving systems.

[0039] Especially in complex urban traffic scenarios, vehicles need to process multi-source heterogeneous data in real time, accurately perceive and make decisions on the surrounding environment, which requires automatic driving systems to have strong data processing and model optimization capabilities. Data-driven algorithm iteration has thus become the core path to promote the progress of automatic driving technology, by collecting, analyzing and utilizing various data in the vehicle driving process, continuously optimizing perception, planning and decision-making algorithms, and improving the intelligent level of the system.

[0040] In the architecture of autonomous driving systems, middleware plays a crucial role. As a bridge connecting the underlying hardware and the upper-layer applications, it undertakes key functions such as data interaction, resource management, and task scheduling. In the field of intelligent driving, middleware not only needs to process massive data from multiple sensors such as cameras, lidar, and millimeter-wave radar, but also needs to achieve efficient communication and collaboration between different functional modules.

[0041] Taking the communication middleware as an example, in the current mainstream electronic and electrical architecture, there are both high-performance SOC chips and traditional MCU chips in the vehicle, involving multiple operating systems such as QNX and Linux. This requires the communication middleware to have cross-platform compatibility and be able to achieve stable and efficient communication in a multi-chip, multi-core, and heterogeneous environment.

[0042] Among them, the SOC (System on a Chip) chip is the main chip that provides strong AI computing power for intelligent driving, integrating a complete information system onto a single chip; The MCU (Microcontroller Unit) chip is a relatively simple and low-power chip that integrates a microprocessor, memory, input / output interface, etc. on a chip, mainly used for implementing specific control functions; QNX is a Unix-like commercial operating system that complies with the POSIX specification; Linux is a Unix-like operating system commonly used in automotive infotainment systems and intelligent driving domains that require complex computing functions; Currently, the communication middleware for intelligent driving mainly has two technical routes: SOME / IP and DDS. Among them, DDS is gradually becoming the industry mainstream due to its high bandwidth, low latency, and completely decoupled software architecture, etc. It has shown significant performance advantages in dealing with the high-performance transmission needs of diversified data streams in the era of full-scale production of autonomous driving; SOME / IP (Scalable service-Oriented MiddlewarE over IP) is a communication protocol standard in the automotive electronics field; DDS (Data Distribution Service) is an international standard communication protocol published by the Object Management Group; With the increasing penetration rate of high-level intelligent driving systems in mass-produced vehicles, the massive data generated by mass-produced vehicles during daily driving has become a valuable resource. Compared to the early data collection method relying on research and development collection vehicles, mass-produced vehicle data collection has the advantages of wide range, strong real-time performance, and closer to real user driving behavior.

[0043] By returning these data to the cloud and combining AI algorithms for deep mining and analysis, car companies and autonomous driving technology providers can achieve data closure, that is, from data collection, transmission, storage, processing to model training and optimization, forming a complete feedback iteration system. In this system, newly collected data is constantly used to train and optimize neural network models, driving the continuous evolution of autonomous driving systems.

[0044] Wherein, AI (Artificial Intelligence) refers to the intelligence exhibited by machines, and its core is to enable computer systems to perform tasks that usually require human intelligence; For example, at the perception level, training the BEV + Transformer model with massive data can effectively improve the recognition accuracy of the system to the surrounding environment; at the planning level, Tesla uses a data-driven decision tree generation model to significantly reduce the calculation time of the planned path, realizing the iterative upgrade of the planning algorithm; Wherein, BEV (Bird's-Eye View) + Transformer model is a key technology in the field of autonomous driving perception, which generates a 2D or 3D environment representation map from above the vehicle by fusing images from multiple cameras around the vehicle. After the deep learning model architecture Transformer was introduced into the field of computer vision and autonomous driving, it can handle the relationship between elements in the BEV space and understand their interactions, greatly improving the upper limit of perception capability.

[0045] In known prior art, the traditional DTW algorithm forces a monotonous and continuous path for time series alignment, and cannot represent the asymmetry of traffic participant decision-making. In vehicle rush and pedestrian avoidance game scenarios, trajectory mutation points are incorrectly matched, causing similarity weight deviation from the real interaction logic; Wherein, DTW (Dynamic Time Warping) algorithm is a classic algorithm for measuring the similarity between two time series of different lengths, which can "warp" the time axis to find the best alignment between two sequences, thus overcoming the difference in time speed; Distortion weight causes high-risk scenario data to be diluted during fusion, leading to decision boundary deviation after multiple iterations. For example, after 100 closed-loop iterations, the existing autonomous driving system's misjudgment rate at the intersection game scenario increased by 23%; And in the prior art, the safety margin evaluation and similarity calculation modules are decoupled, and cannot dynamically correct the path constraint defects of the DTW algorithm, resulting in systematic error accumulation.

[0046] Moreover, the existing scheme regards traffic participants as independent entities, ignoring the essential influence of their interactive game on the similarity of the scene. The symmetry assumption of the DTW algorithm ( ) conflicts with the asymmetry of real road games, and this fundamental contradiction leads to the failure of data closure in complex scenes.

[0047] Based on this, the present application proposes an intelligent driving control method for a vehicle and a vehicle. By quantifying the asymmetry between traffic participants, filtering the historical scene data set, and incrementally training the pre-trained intelligent driving decision model, the intelligent driving decision is optimized, which more accurately represents the dynamic game relationship in complex traffic scenes, enables the intelligent driving decision model to continuously adapt to the effect of the asymmetric interaction mode in real scenes, and solves the problem that the traditional DTW algorithm cannot accurately depict the asymmetric interactive game relationship between traffic participants based on the assumption of symmetry and monotonic continuous path, regards traffic participants as independent entities, and ignores the mutual influence between their decisions.

[0048] The intelligent driving decision model makes driving decisions based on multi-modal perception input, and its structure is a multi-modal fusion encoder and decoder architecture, including a sensor feature encoder for processing image, point cloud and other raw data, a multi-modal feature fusion module for fusing information from different sources, a time series or graph neural network for modeling the interactive relationship between traffic participants, and a decision head for finally outputting specific decisions.

[0049] The construction of the intelligent driving decision model is a systematic project. The initial construction usually adopts supervised learning or reinforcement learning, and is trained on an offline super large data set to establish a baseline model with basic driving ability. Thereafter, the intelligent driving decision model enters the continuous integration optimization stage, and continuously performs incremental learning on new data collected by mass-produced vehicles, so that the intelligent driving decision model can adapt to new scenes and optimize weak links, thereby realizing the continuous evolution of performance.

[0050] In the following, the embodiments of the present application will be described in detail with reference to the accompanying drawings. As shown in FIGS. 1 to 3, Figure 1 and FIGS. 4 to 6, Figure 2 in one illustrative embodiment of the intelligent driving control method for a vehicle of the present application, the decision optimization method comprises: S1: obtaining the raw data stream of the vehicle sensor network, and extracting the dynamic feature vector from the raw data stream.

[0051] In some embodiments, the multi-source vehicle-mounted data of the vehicle sensor network is first obtained, the time synchronization module is used to align the unified clock of the multi-source vehicle-mounted data, and the resampling and interpolation processing is performed according to the time stamp to obtain the raw data stream; The multi-source vehicle-mounted data includes environment perception data, middleware running state, and real-time instruction signal. The environment perception data includes three-dimensional space point cloud data, traffic scene continuous image frame data, and target detection data. The three-dimensional space point cloud data is obtained by a laser radar and is used to accurately describe the spatial geometric structure of the surrounding environment. The traffic scene continuous image frame data is obtained by a vehicle-mounted camera and is used to obtain the visual texture and color information of the traffic scene. The target detection data is obtained by a millimeter wave radar and includes the radial distance, relative speed, and angle information of the target, and is used to enhance the environment perception ability in low visibility conditions. The middleware running state is obtained by a middleware module and records the task scheduling delay, process resource occupation, and communication link quality, and is used for subsequent delay index extraction. The real-time instruction signal is obtained by a vehicle control system and covers parameters such as acceleration, steering angle, and brake pressure, and is used to reflect the dynamic changes of driving behavior and vehicle state.

[0052] Further, the multi-source vehicle-mounted data is wherein, is the three-dimensional space point cloud data; is the traffic scene continuous image frame data; is the target detection data; is the real-time instruction signal; is the middleware running state.

[0053] The time synchronization module aligns the unified clock reference of each data source in the multi-source vehicle-mounted data resamples and interpolates the original data output by each sensor according to the timestamp to match signals with different sampling frequencies and delay characteristics under the same time reference, forming a spatio-temporally aligned original data stream wherein is the i-th sampling moment under the unified time reference, is the multi-modal data packet at this moment from the laser radar, vehicle-mounted camera, millimeter wave radar, vehicle control system, and middleware module.

[0054] Further, the time synchronization module aligns the unified clock of the environment perception data, middleware running state, and real-time instruction signal in the multi-source vehicle-mounted data and resamples and interpolates them according to the timestamp to obtain vehicle operating environment data, control instruction data, and middleware running data.

[0055] ​The time synchronization module is used for aligning and resampling and interpolating the multi-source vehicle data, so as to effectively solve the time sequence dislocation problem caused by the inconsistent sampling frequency and transmission delay of different sensors. The original data stream is generated in time and space alignment, a consistent time reference is provided for subsequent feature extraction and fusion, the synchronization of the environmental feature matrix, the control feature vector and the system state vector in the time dimension is ensured, a solid foundation is laid for accurately constructing the traffic participant interaction graph and calculating the asymmetric factor, and the accuracy and reliability of the entire decision optimization process are improved from the data source.

[0056] In some embodiments, the method for extracting features from the original data stream to obtain a dynamic feature vector comprises: The vehicle operating environment data in the original data stream is fused by a space-time encoder to obtain an environmental feature matrix; The control instruction data in the original data stream is segmented by a preset length sliding window to extract a statistical feature vector; The message delay of the middleware running data in the original data stream is recorded, and the central processor occupancy rate is calculated, and the message delay and the central processor occupancy rate are combined to construct a system load vector; The environmental feature matrix, the statistical feature vector and the system load vector are combined to obtain a dynamic feature vector.

[0057] Further, the vehicle operating environment data is fused by a space-time encoder to map the spatial geometric structure, visual texture information and motion target physical attributes to the same feature space to generate an environmental feature matrix , wherein and respectively represent the height and width of the feature map in the spatial dimension, represents the number of fused feature channels, and the environmental feature matrix maintains the consistency of the road topological structure and the obstacle motion trajectory in the spatial distribution.

[0058] The acceleration sequence and the steering angle change rate sequence are segmented by a sliding time window with a length of in the control instruction data to extract a statistical feature vector ; , wherein is the mean value of the acceleration in the window, reflecting the stability of the vehicle longitudinal power output; is the standard deviation of the acceleration in the window, depicting the amplitude of the longitudinal control fluctuation; is the mean value of the steering angle change rate in the window, used to describe the stability of the vehicle lateral dynamic adjustment.

[0059] The time interval from the generation of the message in the middleware running data to the completion of the processing is recorded as the message delay Reflecting the real-time of communication link and scheduling mechanism, while calculating the central processor occupancy rate To measure the computing load pressure of the middleware running; Combine the message latency and the central processor occupancy rate into a system load vector ; Finally, combine the environmental feature matrix , the statistical feature vector and the system load vector into a dynamic feature vector .

[0060] In the technical solution, the vehicle operating environment data is fused by the space-time encoder, which can unify the feature representation of different sensor data and preserve the spatial structure and time sequence correlation of the environment; The statistical feature vector is extracted by segmenting the control instruction data with a sliding window, effectively capturing the dynamic change trend of vehicle motion; The system load vector is constructed by recording the message latency and calculating the central processor occupancy rate of the middleware running data, reflecting the running state of the computing platform in real time; Finally, the environmental feature matrix, the statistical feature vector and the system load vector are combined to obtain the dynamic feature vector, which provides a comprehensive and multi-dimensional scene representation basis for subsequent interaction graph construction and similarity calculation, and improves the accuracy and robustness of subsequent decision optimization.

[0061] S2: Extract the state vector of the traffic participant according to the dynamic feature vector, construct the edge between any two traffic participants according to the state vector, construct the traffic participant interaction graph according to the state vector and the edge, and calculate the asymmetric factor matrix; wherein the asymmetric factor matrix is used to represent the asymmetric strength of the interaction between any two traffic participants.

[0062] In some embodiments, the method of constructing the traffic participant interaction graph according to the state vector and the edge comprises: Use the target detection and tracking module to identify and locate the traffic participants in the environmental feature matrix, and extract the position coordinates, current speed, acceleration and heading angle of the traffic participants to form the state vector; According to the state vector, calculate the first Euclidean distance and relative speed between any two traffic participants, and when the first Euclidean distance and relative speed between the two traffic participants meet the preset judgment condition, construct the edge between the two traffic participants; Construct the traffic participant interaction graph according to the state vector of each traffic participant and the edge.

[0063] Further, all traffic participants in the environmental feature matrix are identified and located by the target detection and tracking module, and the state vector of each traffic participant is extracted ; wherein, and are the position coordinates of the traffic participants in the same world coordinate system, used to determine the absolute spatial position of the traffic participants in the road topology; is the current speed of the traffic participant, reflecting the longitudinal motion state; is the acceleration, depicting the speed change trend and driving intention; is the heading angle, indicating the orientation of the traffic participant relative to the reference direction of the road.

[0064] In the process of constructing edges, the first Euclidean space distance between each pair of traffic participants and the relative speed are calculated, wherein, represents the modulus of the difference between the two speed vectors; When the first Euclidean distance between the two traffic participants and the relative speed satisfy the preset judgment condition, an edge is constructed between the two traffic participants; the calculation model of the preset judgment condition is:

[0065] wherein, is a preset reaction time threshold value, used to simulate the operation delay of vehicles or pedestrians, is an additional safety space compensation, used to improve robustness and cover sensor positioning errors and environmental uncertainties; through the preset judgment condition, it can be ensured that the interaction relationship is established only when the two can produce mutual influence within time and the distance is less than the safety margin .

[0066] The traffic participant interaction graph is constructed by the state vector of each traffic participant and the edge, wherein the traffic participant interaction graph is composed of a node set and an edge set ; wherein, the node set represents the motion and state information of each traffic participant, and the edge set represents the participant pairs that may have direct interaction relationship.

[0067] By using the target detection and tracking module to identify and locate the environmental feature matrix, the accurate state vector of the traffic participant is extracted, including position, speed, acceleration and heading angle, providing fine-grained physical state input for interaction relationship modeling; ​By calculating the Euclidean distance and relative speed between traffic participants, and constructing edges based on preset judgment conditions, the potential interaction relationship between traffic participants can be accurately captured, and the interference of irrelevant objects can be avoided. Finally, based on the state vector and the edge, an interaction graph of traffic participants is constructed, which abstracts the complex traffic scene into a computable graph structure, provides a basis for subsequent asymmetric factor calculation, and improves the accuracy and effectiveness of interaction modeling.

[0068] In some embodiments, the method of calculating the asymmetric factor matrix comprises: Using a graph attention network to perform node encoding on the state vector, calculating the weight of the edge according to the node-encoded state vector, and obtaining the edge weight; According to the edge weight corresponding to the two traffic participants in different directions, the asymmetric factor is calculated, and the asymmetric factor matrix is generated according to the asymmetric factor.

[0069] Further, the node set of the traffic participant interaction graph is node-encoded by the node encoding layer of the graph attention network, and a node representation capable of capturing the motion pattern and behavior characteristics of the neighbor nodes is generated through adaptive feature aggregation; For each directed edge , the edge weight is calculated using the edge weight calculation function of the graph attention network , and the edge weight represents the behavior influence strength of the traffic participant on the traffic participant in the current scene feature space; Since the interaction relationship in the traffic scene often has directional differences, the influence degree of the same pair of traffic participants in different directions is not equal, so the asymmetric factor is defined; The asymmetry is quantified by calculating the absolute value of the difference between the edge weights of the two opposite directions, and the asymmetric factor is greater, indicating that the influence degree of the traffic participant on the traffic participant is more obvious than the influence degree of the traffic participant on the traffic participant .

[0070] All asymmetric factors that satisfy are organized into an asymmetric factor matrix according to the index of the traffic participant, and the calculation model of the asymmetric factor matrix is:

[0071] wherein, is the asymmetric factor matrix.

[0072] By using the graph attention network to encode the state vector, the information of the neighbor nodes can be adaptively aggregated, and a node representation rich in context semantics is generated. The weight of the edge is calculated based on the encoded node representation, and the mutual influence strength between the traffic participants is accurately quantified. By calculating the difference of the edge weight in different directions, the asymmetric factor is obtained, which effectively describes the directionality and asymmetry of the interaction relationship. Finally, the asymmetric factor matrix is generated, which provides a quantitative basis for the subsequent similarity calculation, overcomes the defect of ignoring the interaction asymmetry in the traditional method, and improves the discrimination ability and reliability of the scene matching.

[0073] S3: Obtain the historical scene data set of the vehicle, calculate the corrected similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculate the weight coefficient according to the corrected similarity; In some embodiments, the method of calculating the corrected similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculating the weight coefficient according to the corrected similarity includes: Extract the real-time scene feature sequence of the dynamic feature vector in chronological order, and select the target historical scene in the historical scene data set; Calculate the second Euclidean distance of the real-time scene feature sequence and the target historical scene, calculate the corrected dynamic time warping distance by the dynamic time warping algorithm according to the second Euclidean distance and the asymmetric factor matrix, and obtain the corrected similarity according to the corrected dynamic time warping distance; According to the corrected similarity, calculate the weight coefficient by the decay type mapping function.

[0074] Further, the real-time scene feature sequence is extracted from the dynamic feature vector in chronological order ; Wherein, each is a multi-modal fusion feature vector at a uniform time, which includes the road topology, obstacle trajectory information and control and system state features encoded in the environment feature matrix; is a time step number, from 1 to , indicating that the real-time scene feature sequence contains continuous time points; Meanwhile, the historical data set is obtained, wherein each is derived from the closed-loop verified historical operation record, and each is composed of a pair of samples , is a multi-modal input feature, is a corresponding supervision label, is the a number of data samples of the historical scene, the target historical scene is selected from the historical data set ; wherein, is a multi-modal feature vector corresponding to a time step; is a similarity between the real-time scene feature sequence and the target historical scene under time warping, a warping path set is constructed ; wherein, each element in the path set represents a pair relationship between the real-time scene time step and the historical scene time step .

[0075] For each pair of pair relationships, the second Euclidean distance between the real-time scene feature sequence and the target historical scene in the feature space is calculated to measure the static similarity, and the interactive directionality difference provided by the asymmetric factor matrix is introduced to dynamically adjust the matching cost by the term to enlarge or reduce the matching cost in the calculation of the modified dynamic time warping distance, and to ensure that the warping path with the minimum cumulative weighted distance is selected to obtain the best alignment effect by minimizing operation; the calculation model of the modified dynamic time warping distance is:

[0076] wherein, is a game intensity coefficient, used to control the influence degree of interactive asymmetry on matching distance.

[0077] Let the modified dynamic time warping distance be , the modified similarity is obtained, which describes the time sequence matching difference between the real-time scene and the first historical scene after considering the environmental features, control features, system state features and interactive asymmetry factors; The modified similarity is converted into a normalized weight coefficient by defining a decay type mapping function, the weight coefficient represents the similarity contribution coefficient of the real-time scene and the first historical scene, the greater the weight value, the higher the influence of the scene data in the fusion result, and the calculation formula of the weight coefficient is:

[0078] wherein, is a decay factor, used to control the decay speed between similarity and weight; when ​When the value is large, a small similarity difference will be significantly amplified, and the weight of a low-similarity scene will decay faster. When the value is small, the weight changes more smoothly.

[0079] Through the calculation formula of the weight coefficient, a small will get a weight value close to 1, while a large will get a weight value significantly lower than 1, thereby realizing the priority ranking of scenes based on matching accuracy. Calculate the weight coefficients of all historical scenes in batches according to the index from 1 to to form a weight vector , where is the number of historical scenes, and the weight vector is used to give higher contribution to high-similarity historical scenes in the fusion process, thereby improving the accuracy and effectiveness of subsequent scene-enhanced training sample construction.

[0080] By extracting the real-time scene feature sequence of the dynamic feature vector and the target historical scene in the historical scene dataset, a comparison basis is provided for similarity calculation. By calculating the Euclidean distance between the real-time scene and the historical scene and combining the asymmetric factor matrix to modify the dynamic time warping algorithm, the difference between scenes can be accurately quantified in asymmetric interactive scenes, avoiding the matching distortion caused by the symmetric assumption of traditional DTW algorithms. Based on the modified dynamic time warping distance, the weight coefficient is calculated through a decay-type mapping function, so that high-similarity historical scenes obtain higher weights, ensuring the representativeness and relevance of the enhanced dataset and improving the efficiency and effectiveness of subsequent model optimization.

[0081] S4: Weighted fusion of the weight coefficient and the historical scene dataset to obtain an enhanced dataset; Further, the calculation model of the enhanced dataset is:

[0082] wherein, is the enhanced dataset; in the calculation model of the enhanced dataset, the weighted operation on the historical scene dataset is realized by applying a weight label to each sample or repeating or scaling the sample according to the weight ratio in the sampling stage, thereby forming an enhanced dataset that is closer to the current scene features in sample distribution; The enhanced dataset takes into account the diversity of historical scenes and the relevance of real-time scenes, and is highly consistent with the target running scene in statistical characteristics, thereby improving the adaptability and generalization performance of the intelligent driving decision-making model.

[0083] S5: updating and training the pre-trained intelligent driving decision model according to the enhanced data set to obtain an updated intelligent driving decision model, and generating an intelligent driving decision according to the updated intelligent driving decision model to control the vehicle to run.

[0084] In the technical solution, the traditional rule-based decision logic is difficult to cover all long-tail scenarios, and a data-driven decision model becomes the key to realizing safe and humanized driving. However, the traditional DTW algorithm cannot accurately depict the asymmetric interaction game relationship between traffic participants, and the traffic participants are regarded as independent entities, ignoring the mutual influence between their decisions. The symmetry assumption of DTW conflicts with the asymmetric nature of real road games, resulting in the failure of the entire data closed loop in key complex scenarios, and it is difficult to meet the stringent requirements of high-level autonomous driving on safety and reliability.

[0085] And through the above-mentioned scheme, the original data stream of the vehicle sensor network is obtained and the dynamic feature vector is extracted, the traffic participant interaction graph is constructed and the asymmetric factor matrix is calculated, the modified similarity and the weight coefficient are calculated based on the asymmetric factor matrix and the historical scene data set, the precise screening and weighted fusion of the historical data are realized, and the enhanced data set is generated. Therefore, compared with the prior art, the asymmetric strength of the interaction influence between the traffic participants can be quantified, so that the dynamic game relationship in the complex traffic scene can be more accurately represented, and the intelligent driving decision model can continuously adapt to the asymmetric interaction mode in the real scene.

[0086] On this basis, the pre-trained decision model is incrementally optimized through the enhanced data set, and compared with the assumption of only realizing symmetry and monotonic continuous path in the prior art, the safety, humanization and reliability of the intelligent driving decision in the complex game environment can be improved.

[0087] In some embodiments, as shown in Figure 3 The method of updating and training the pre-trained intelligent driving decision model according to the enhanced data set includes: calculating a knowledge distillation loss through the pre-trained intelligent driving decision model and the updated intelligent driving decision model according to the enhanced data set and preset model parameters; generating an edge case through the generative adversarial network, and calculating a decision boundary loss through the updated intelligent driving decision model according to the edge case and the enhanced data set; obtaining updated model parameters according to the knowledge distillation loss and the decision boundary loss.

[0088] Further, the pre-trained intelligent driving decision model is updated and trained according to the enhanced data set and the preset model parameters, which adapts to the new scene characteristics while retaining the existing capabilities; The preset model parameters are obtained through iterative inheritance, and are originally derived from an intelligent driving decision model pre-trained on a large-scale and diversified historical driving data set. The intelligent driving decision model serves as the starting point of the entire closed-loop system. In each subsequent data closed-loop iteration, the parameters of the current intelligent driving decision model are the saved model state after incremental updating and verification of the previous round. Therefore, the preset model parameters are a dynamically evolving and continuously optimized parameter set.

[0089] First, the knowledge distillation loss is calculated , and the calculation model is:

[0090] wherein, is the prediction result of the intelligent driving decision model before updating on the input sample in the augmented data set; is the prediction result of the intelligent driving decision model after updating on the input sample in the augmented data set; The knowledge distillation loss achieves historical knowledge transfer and measures the difference between the intelligent driving decision model before and after updating in the output space. Minimizing the knowledge distillation loss enables the intelligent driving decision model after updating to adapt to the feature distribution of the augmented data set while maintaining similar decision-making behavior as the intelligent driving decision model before updating, thereby avoiding forgetting the patterns learned in historical scenarios. Subsequently, the edge case is generated by using the generative adversarial network , and the edge case is located near the decision boundary and can expose the weaknesses of the intelligent driving decision model in the uncertain region. According to the edge case decision boundary loss, the calculation model is:

[0091] wherein, is the expected minimum output interval threshold; is the output difference of the intelligent driving decision model after updating on the input sample and the edge case . When the difference is less than , a positive loss is generated, which pushes the intelligent driving decision model to expand the interval between classes in this region and improve the discriminability.

[0092] The updated model parameters are obtained according to the knowledge distillation loss and the decision boundary loss, and the model parameters have both stability and robustness.

[0093] The knowledge distillation loss is calculated by enhancing the data set and preset model parameters, so that the updated model retains historical knowledge while adapting to new scenarios, effectively alleviating the catastrophic forgetting problem. The decision boundary loss is calculated by generating edge cases through the adversarial generation network, actively exposing the weak links of the model under critical conditions, and promoting the optimization of the decision boundary of the model. The updated model parameters are obtained by jointly optimizing the knowledge distillation loss and the decision boundary loss, so that the model has both the stability of historical experience and strong adaptability to new scenarios, improving the robustness and safety of intelligent driving decisions in extreme conditions.

[0094] In some embodiments, the intelligent driving control method further comprises: According to the updated model parameters, the updated intelligent driving decision model is simulated in a scene, and a prediction result is output, and a safety margin and a misjudgment rate are calculated according to the prediction result; The safety margin or the misjudgment rate is compared with the size of the preset threshold value, and when the safety margin or the misjudgment rate meets the preset threshold condition, a feedback signal is generated, the parameters of the graph attention network are updated, and the asymmetric factor matrix is recalculated according to the updated parameters of the graph attention network.

[0095] By simulating a scene according to the updated model parameters and calculating the safety margin and the misjudgment rate, the performance of the optimized model in extreme scenarios can be quantitatively evaluated; By comparing the safety margin or the misjudgment rate with the size of the preset threshold value, it can be automatically determined whether the model meets the deployment requirements; When the indicators do not meet the conditions, a feedback signal is generated and the parameters of the graph attention network are updated, realizing the closed-loop optimization of the asymmetric factor calculation module, forming an iterative mechanism for self-correction, continuously improving the accuracy of interactive perception, and ultimately ensuring that the entire system can continuously evolve towards a safer and more reliable direction.

[0096] In some embodiments, the method of comparing the safety margin or the misjudgment rate with the size of the preset threshold value further comprises: When the safety margin or the misjudgment rate does not meet the preset threshold condition, a verification pass flag is generated and output.

[0097] By setting the output condition of the verification pass flag, a pass signal is automatically generated when the safety margin and the misjudgment rate both meet the preset threshold requirements, providing a clear and objective decision basis for model deployment; This mechanism can avoid the subjectivity and lag of manual intervention, ensuring that only high-quality models that have been thoroughly verified can enter the actual application stage, thereby improving the safety and reliability of the intelligent driving system upgrade, and providing reliable protection for the final decision of the vehicle control system.

[0098] Further, as shown in Figure 4 the prediction result is output, and a safety margin and a misjudgment rate are calculated according to the prediction result; Stability and safety evaluation in a dual verification environment; First, the updated intelligent driving decision model injects disturbance to the acceleration vector of the traffic participants in the simulation platform To simulate dynamic scenarios such as sudden speed change, emergency braking and acceleration overtaking, the updated intelligent driving decision model outputs the prediction results, and the safety margin is calculated by the prediction results and the true trajectory, and the calculation model is:

[0099] Among them, is the actual minimum safety distance between the traffic participants and the vehicle under the disturbance condition; is the predicted minimum safety distance of the model under the same condition; the minimum value of the ratio reflects the safety keeping ability of the intelligent driving decision model in extreme cases.

[0100] Then, the updated intelligent driving decision model is run in the real car shadow mode to obtain the prediction results, and all perception and decision results are recorded, and the misjudgment rate is calculated by comparing with the preset artificial annotation or high-precision map ; the reliability of the intelligent driving decision model in the real road environment is described by the misjudgment rate; Among them, is the different sample, i.e. error sample, of the prediction results generated by the intelligent driving decision model compared with the preset artificial annotation or high-precision map; is the understanding of the surrounding environment by the intelligent driving decision model at each time, including obstacle recognition, lane line detection, path planning, etc., and the corresponding decision is generated, including steering angle, acceleration, etc.; thus the preset results generated constitute a sample, and the collection of each sample is all samples; When the safety margin is lower than the threshold or the misjudgment rate exceeds the threshold , a feedback signal is generated , indicating that the intelligent driving decision model has safety or perception accuracy defects; When the feedback signal is triggered, the updating process of the parameters of the graph attention network is started to improve the calculation accuracy of the asymmetric factor in the interaction graph; The updating process of the parameters of the graph attention network is realized by gradient descent method, and the updating formula is:

[0101] Among them, is the parameter of the graph attention network before updating; learning rate, used to control the step size of parameter adjustment of the graph attention network to ensure convergence stability; loss function constructed based on the verification feedback signal gradient of the parameter of the graph attention network, loss function comprehensive consideration of safety margin below the threshold or misjudgment rate exceeding the threshold penalty, by optimizing the loss function make the graph attention network more accurately capture the interaction asymmetry between traffic participants, enhance the asymmetric factor discrimination ability; According to the updated parameters of the graph attention network recalculate the asymmetric factor matrix, realize the adaptive adjustment of the traffic participant interaction relationship modeling, form a closed loop optimization mechanism, effectively improve the accuracy of scene similarity calculation, and promote the iterative upgrade of intelligent driving decision-making; Otherwise, generate and output the verification flag , allowing the intelligent driving decision-making model to enter the actual deployment stage.

[0102] As Figure 5 shown, the present application also provides a vehicle. The vehicle comprises: a vehicle sensor network for collecting multi-source vehicle data and historical scene data sets; a data acquisition module in communication with the vehicle sensor network, for acquiring environment perception data in the multi-source vehicle data; a middleware module in communication with the vehicle sensor network, for acquiring middleware running status in the multi-source vehicle data; a vehicle control system in communication with the vehicle sensor network, for acquiring real-time instruction signals in the multi-source vehicle data; a time synchronization module in communication with the vehicle sensor network, for aligning the multi-source vehicle data with a unified clock, and performing resampling and interpolation processing according to the timestamp to obtain an original data stream; a space-time encoder in communication with the time synchronization module, for multi-modal fusion of vehicle operating environment data in the original data stream to obtain an environment feature matrix; a target detection and tracking module in communication with the space-time encoder, for identifying and locating traffic participants based on the environment feature matrix, and extracting state vectors of the traffic participants; a processor in communication with the time synchronization module, the space-time encoder and the target detection and tracking module, the processor being configured to: An edge between two traffic participants is constructed according to the state vector, and an interaction graph of the traffic participants is constructed according to the state vector and the edge, and an asymmetric factor matrix is calculated; A revised similarity is calculated based on the historical scenario dataset, the asymmetric factor matrix and the dynamic feature vector, and a weight coefficient is calculated according to the revised similarity; The weight coefficient and the historical scenario dataset are weighted and fused to obtain an enhanced dataset; The pre-trained intelligent driving decision-making model is updated and trained according to the enhanced dataset, to obtain an updated intelligent driving decision-making model, and an intelligent driving decision is generated according to the updated intelligent driving decision-making model to control the operation of the vehicle.

[0103] A complete multi-source data acquisition system is constructed through the collaborative work of the vehicle-mounted sensor network, the data acquisition module, the middleware module and the vehicle control system; Through the step-by-step processing of the time synchronization module, the space-time encoder and the target detection and tracking module, efficient conversion from raw data to traffic participant state vectors is realized; Finally, the processor completes the whole process of interaction graph construction, asymmetric factor matrix calculation, similarity revision and model optimization, and the innovative algorithm in the method claim is materialized as the specific functional architecture of the vehicle, so that the vehicle itself has the ability of continuous learning and self-optimization, and the engineering landing and productization protection of the intelligent driving decision optimization method are realized.

[0104] In some embodiments, the data acquisition module comprises: a laser radar for acquiring three-dimensional space point cloud data in the environment perception data; a vehicle-mounted camera for acquiring traffic scene continuous image frame data in the environment perception data; a millimeter wave radar for acquiring target detection data in the environment perception data.

[0105] In some embodiments, the processor is further configured to: segment the control instruction data in the raw data stream by a preset length sliding window, and extract a statistical feature vector; record the message delay of the middleware running data in the raw data stream and calculate the central processor occupancy rate, and combine the message delay and the central processor occupancy rate to construct a system load vector; combine the environment feature matrix, the statistical feature vector and the system load vector to obtain a dynamic feature vector.

[0106] In some embodiments, the processor is further configured to: According to the state vector, a first Euclidean distance and a relative speed between any two traffic participants are calculated, and an edge between the two traffic participants is constructed when the first Euclidean distance and the relative speed between the two traffic participants meet a preset judgment condition; According to the state vector of each traffic participant and the edge, an interaction graph of the traffic participants is constructed.

[0107] In some embodiments, the processor is further configured to: The state vector is node-encoded by using a graph attention network, a weight of the edge is calculated according to the node-encoded state vector, and the edge weight is obtained; An asymmetric factor is calculated according to the edge weight corresponding to the two traffic participants in different directions, and an asymmetric factor matrix is generated according to the asymmetric factor.

[0108] In some embodiments, the processor is further configured to: The real-time scene feature sequence of the dynamic feature vector is extracted in chronological order, and a target historical scene in the historical scene data set is selected; A second Euclidean distance between the real-time scene feature sequence and the target historical scene is calculated, a modified dynamic time warping distance is calculated according to the second Euclidean distance and the asymmetric factor matrix by using a dynamic time warping algorithm, and a modified similarity is obtained according to the modified dynamic time warping distance; A weight coefficient is calculated according to the modified similarity by using a decay type mapping function.

[0109] In some embodiments, the processor is further configured to: A knowledge distillation loss is calculated according to the enhanced data set and the preset model parameter by using the pre-trained intelligent driving decision model and the updated intelligent driving decision model; An edge case is generated by using a generative adversarial network, and a decision boundary loss is calculated according to the edge case and the enhanced data set by using the updated intelligent driving decision model; The updated model parameter is obtained according to the knowledge distillation loss and the decision boundary loss.

[0110] In some embodiments, the processor is further configured to: A scene simulation is performed on the updated intelligent driving decision model according to the updated model parameter, and a prediction result is output, and a safety margin and a misjudgment rate are calculated according to the prediction result; The safety margin or the misjudgment rate is compared with a preset threshold, and when the safety margin or the misjudgment rate meets a preset threshold condition, a feedback signal is generated, the parameters of the graph attention network are updated, and the asymmetric factor matrix is recalculated according to the updated parameters of the graph attention network.

[0111] Although the embodiments of the present application have been shown and described above, it is understood that the above-described embodiments are exemplary and are not to be construed as limiting the present application, and that changes, modifications, substitutions and variations can be made by those skilled in the art without departing from the scope of the present application.

Claims

1. A method for intelligent driving control of a vehicle, characterized by, It comprises: acquiring the original data stream of the vehicle sensor network, and performing feature extraction on the original data stream to obtain a dynamic feature vector; extracting a state vector of a traffic participant according to the dynamic feature vector, constructing an edge between any two traffic participants according to the state vector, constructing a traffic participant interaction graph according to the state vector and the edge, and calculating an asymmetric factor matrix; wherein the asymmetric factor matrix is used to represent the asymmetric strength of the interaction between any two traffic participants; acquiring a historical scene data set of the vehicle, calculating a modified similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculating a weight coefficient according to the modified similarity; performing weighted fusion on the weight coefficient and the historical scene data set to obtain an enhanced data set; updating and training a pre-trained intelligent driving decision-making model according to the enhanced data set to obtain an updated intelligent driving decision-making model, generating an intelligent driving decision according to the updated intelligent driving decision-making model, and controlling the vehicle to run.

2. The intelligent driving control method of a vehicle according to claim 1, characterized by, The method for performing feature extraction on the original data stream to obtain a dynamic feature vector comprises: performing multi-modal fusion on vehicle operating environment data in the original data stream through a space-time encoder to obtain an environment feature matrix; segmenting control instruction data in the original data stream through a pre-set length sliding window to extract a statistical feature vector; recording the message delay of middleware running data in the original data stream and calculating the central processor occupancy rate, combining the message delay and the central processor occupancy rate to construct a system load vector; combining the environment feature matrix, the statistical feature vector and the system load vector to obtain the dynamic feature vector.

3. The intelligent driving control method of a vehicle according to claim 2, characterized by, The method for constructing a traffic participant interaction graph according to the state vector and the edge comprises: using a target detection and tracking module to identify and locate a traffic participant from the environment feature matrix, and extracting the position coordinates, current speed, acceleration and heading angle of the traffic participant to form a state vector; calculating the first Euclidean distance and relative speed between any two traffic participants according to the state vector, and constructing an edge between the two traffic participants when the first Euclidean distance and relative speed between the two traffic participants meet a pre-set judgment condition; constructing the traffic participant interaction graph according to the state vector of each traffic participant and the edge.

4. The intelligent driving control method of a vehicle according to claim 1, characterized by, The method for calculating an asymmetric factor matrix comprises: using a graph attention network to perform node coding on the state vector, calculating the weight of the edge according to the node-coded state vector to obtain an edge weight; calculating an asymmetric factor according to the edge weight corresponding to the two traffic participants in different directions, and generating the asymmetric factor matrix according to the asymmetric factor.

5. The intelligent driving control method of a vehicle according to claim 1, characterized by, The method for calculating a modified similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculating a weight coefficient according to the modified similarity comprises: extracting a real-time scene feature sequence of the dynamic feature vector in chronological order, and selecting a target historical scene in the historical scene data set; computing a second Euclidean distance of the real-time scene feature sequence and the target historical scene, computing a modified dynamic time warping distance according to the second Euclidean distance and the asymmetric factor matrix through a dynamic time warping algorithm, and obtaining the modified similarity according to the modified dynamic time warping distance; computing the weight coefficient through a decay type mapping function according to the modified similarity.

6. The intelligent driving control method of a vehicle according to claim 1, characterized by, The method for updating and training the pre-trained intelligent driving decision-making model according to the enhanced data set comprises: computing a knowledge distillation loss according to the enhanced data set and preset model parameters through the pre-trained intelligent driving decision-making model and the updated intelligent driving decision-making model; generating an edge case through a generative adversarial network, and computing a decision boundary loss according to the edge case and the enhanced data set through the updated intelligent driving decision-making model; obtaining the updated model parameters according to the knowledge distillation loss and the decision boundary loss.

7. The intelligent driving control method of a vehicle according to claim 6, characterized by, The intelligent driving control method further comprises: performing scene simulation on the updated intelligent driving decision-making model according to the updated model parameters, outputting a prediction result, and computing a safety margin and a misjudgment rate according to the prediction result; comparing the safety margin or the misjudgment rate with a preset threshold value, generating a feedback signal when the safety margin or the misjudgment rate meets a preset threshold condition, updating the parameters of the graph attention network, and recalculating the asymmetric factor matrix according to the updated parameters of the graph attention network.

8. The intelligent driving control method of a vehicle according to claim 7, characterized by, The method for comparing the safety margin or the misjudgment rate with the preset threshold value further comprises: generating and outputting a pass verification flag when the safety margin or the misjudgment rate does not meet the preset threshold condition. 9.The intelligent driving control method of a vehicle according to any one of claims 1 to 8, characterized in that, The method for obtaining the original data stream of the vehicle sensor network comprises: obtaining multi-source vehicle data of a vehicle sensor network, performing unified clock alignment on the multi-source vehicle data through a time synchronization module, and performing resampling and interpolation processing according to a timestamp to obtain the original data stream.

10. A vehicle characterized by comprising: The method comprises: a vehicle sensor network for collecting multi-source vehicle data and a historical scene data set; a data acquisition module in communication connection with the vehicle sensor network, configured to obtain environmental perception data in the multi-source vehicle data; a middleware module in communication connection with the vehicle sensor network, configured to obtain a middleware running state in the multi-source vehicle data; a vehicle control system in communication connection with the vehicle sensor network, configured to obtain a real-time instruction signal in the multi-source vehicle data; a time synchronization module in communication connection with the vehicle sensor network, configured to perform unified clock alignment on the multi-source vehicle data and perform resampling and interpolation processing according to a timestamp to obtain an original data stream; a space-time encoder in communication connection with the time synchronization module, configured to perform multi-modal fusion on vehicle operating environment data in the original data stream to obtain an environmental feature matrix; a target detection and tracking module in communication connection with the space-time encoder, configured to identify and locate a traffic participant according to the environmental feature matrix, and extract a state vector of the traffic participant; a processor, which is communicatively connected with the time synchronization module, the space-time encoder and the target detection and tracking module, and is configured to: construct edges between the traffic participants according to the state vectors, construct a traffic participant interaction graph according to the state vectors and the edges, and calculate an asymmetric factor matrix; calculate a revised similarity based on the historical scene data set, the asymmetric factor matrix and the dynamic feature vector, and calculate a weight coefficient according to the revised similarity; weight and fuse the weight coefficient and the historical scene data set to obtain an enhanced data set; update and train the pre-trained intelligent driving decision-making model according to the enhanced data set to obtain an updated intelligent driving decision-making model, generate an intelligent driving decision according to the updated intelligent driving decision-making model, and control the vehicle to run.

Images