An intelligent networked vehicle safe driving decision control system

Abstract

The application relates to the technical field of intelligent driving, in particular to a safe driving decision control system for intelligent networked vehicles, which comprises a data acquisition module, a data inspection module, a data supplement module, a data fusion module, a scene classification module, a database retrieval module and a vehicle real-time control module; the application introduces a dynamic self-adaptive mutation detection mechanism based on combination of physical kinematic model prediction and historical data trend analysis, and further combines multi-sensor data grouping and a collaborative elimination strategy combining Mahalanobis distance and source confidence; the scheme can more intelligently identify unreasonable data caused by instantaneous interference or sensor failure, and when multi-sensor data conflicts, the reliability and consistency of each data point are comprehensively evaluated, so that the quality and reliability of input data are significantly improved, and a more solid data foundation is provided for subsequent decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, and in particular to an intelligent connected vehicle safety driving decision control system. Background Technology

[0002] Intelligent connected vehicles, as an important direction for the future development of the automotive industry, integrate advanced sensors, controllers, actuators, and other devices, and combine them with modern communication and network technologies to achieve intelligent information exchange and sharing among vehicles, roads, people, and the cloud, ultimately achieving the goals of safe, efficient, comfortable, and energy-saving driving. Their technical architecture can generally be divided into three core layers: perception, decision-making, and control. The decision-making system, based on environmental information provided by the perception system, makes judgments and formulates corresponding control strategies, replacing the human driver in making driving decisions; it is the core manifestation of the intelligence level of intelligent connected vehicles.

[0003] However, existing intelligent driving decision-making technologies, especially the mainstream modular architecture, still face significant challenges in practical applications. On the one hand, the modular architecture decomposes driving tasks into independent modules such as perception, decision-making, and control. These modules may experience information transmission losses and inconsistent optimization goals during operation, leading to error propagation. For example, recognition errors in the perception module due to environmental interference (such as severe weather or sensor noise) can directly affect the accuracy of the decision-making module. On the other hand, current technologies are insufficient in handling rare or insufficiently covered "long-tail scenarios" in the training data, and the decision-making system may fail due to a lack of prior knowledge. Furthermore, deep learning-based systems, especially end-to-end architectures, suffer from a "black box effect" in their decision-making process, with opaque logic that increases the difficulty of behavior prediction and accident liability determination. At the vehicle control level, ensuring that the system consistently makes optimal decisions that meet multiple objectives such as safety, comfort, and efficiency in complex dynamic environments, and can effectively intervene in the event of sudden danger, is also a major challenge. The limitations of single-vehicle intelligence, such as insufficient foresight of potential risks outside the sensor's field of view (blind spots) and the information silo phenomenon caused by insufficient collaboration between vehicles, roads, and the cloud, also restrict the improvement of the overall system performance.

[0004] This invention provides an intelligent connected vehicle safety driving decision-making and control system, aiming to offer a systematic solution to the aforementioned problems. The system introduces a two-layer progressive verification mechanism through a data verification module, combining physical models and statistical laws to improve the reliability of input data; utilizes a data supplementation module to predict potential risks outside sensor blind spots and anticipate risks based on logical reasoning through behavioral anomaly analysis and causal graph reasoning; employs a confidence-driven data fusion method combined with spatial consistency verification to improve the accuracy and robustness of the environmental model; performs deep scene understanding and classification based on a multi-dimensional hazard factor quantification model and attention mechanism; and quickly associates prior knowledge through a database retrieval module; finally, the vehicle real-time control module performs multi-objective optimization decisions and real-time safety monitoring under stringent safety rules, ensuring the safe and efficient execution of control commands. This solution aims to improve the perception robustness, decision-making foresight, and control safety of intelligent driving systems in complex real-world environments. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, the present invention proposes an intelligent connected vehicle safe driving decision control system.

[0006] The technical solution of this invention is: an intelligent connected vehicle safe driving decision control system, comprising:

[0007] The data acquisition module is used to collect multimodal data of the vehicle's surrounding environment through LiDAR, cameras, and millimeter-wave radar.

[0008] The data validation module is used to validate the collected multimodal data;

[0009] The data supplementation module is used to supplement and infer the verified data. The supplementation methods include: inferring the environmental state based on the behavior of surrounding vehicles; and constructing a cause-and-effect graph and inferring potential hazards based on the historical and current changes of the surrounding environment.

[0010] The data fusion module is used to fuse the verified and supplemented multimodal data based on the confidence levels of each data source to generate a unified environmental model.

[0011] The scene classification module is used to calculate the hazard factor of each element in the environment based on the fused environment model, classify elements with hazard factors below a set threshold as obstacles, and determine the current driving scene type based on elements with hazard factors above a set threshold.

[0012] The database retrieval module is used to retrieve potential hazard information and preset operation strategies associated with the current driving scenario from the preset scenario database based on the determined current driving scenario type.

[0013] The vehicle real-time control module is used to receive preset operating strategies and, under the constraints of preset stringent safety rules, generate vehicle control commands and execute real-time control.

[0014] As a preferred option, the data verification module employs the following verification methods when verifying the collected multimodal data:

[0015] A11: Perform data mutation detection and remove or fill in any unreasonable mutation data identified;

[0016] A12: Group the data according to the difference between the data points, and within each group, remove the data points that are judged to be the least reliable, based on the confidence level of each data point and the distribution of data within the group.

[0017] As a preferred embodiment, the data verification module, when performing data mutation detection and removing and filling in identified unreasonable mutation data, specifically includes:

[0018] S11: Data mutation detection. For a data point at the current moment from any sensor data stream, calculate the weighted residual between its predicted value based on the physical kinematics model and the actual observed value, and compare the weighted residual with the dynamic adaptive threshold. If the weighted residual is greater than the dynamic adaptive threshold, it is judged as unreasonable mutation data.

[0019] S12: Remove unreasonable data. Remove data points that are determined to be unreasonable mutations from the valid data sequence and mark them as missing values.

[0020] S13: Missing data imputation. For missing values ​​resulting from the removal of data, the best estimate based on the historical valid data sequence and sensor fusion prediction is used for imputation.

[0021] Preferably, when the data validation module groups the data based on the difference between data points, and within each group, removes the data points deemed least reliable based on the confidence level of each data point and the data distribution within the group, the specific steps include:

[0022] S21: Data grouping, based on spatiotemporal correlation, divide the observation data of different sensors on the same physical entity into the same data group;

[0023] S22: Intragroup data distribution and confidence assessment. For each data point within the same data group, calculate its consistency measure with other data points and obtain its own source confidence.

[0024] S23: Calculate the overall distrust score. Calculate the overall distrust score for each data point within the same data set. This score is a function of its data inconsistency measure and its source confidence.

[0025] S24: Outlier removal. The data points in the group are sorted according to the overall distrust score. The data points with the highest scores that exceed the set threshold are identified as outliers and removed.

[0026] As a preferred embodiment, when the data supplementation module supplements and infers the verified data by inferring the environmental state based on the behavior of surrounding vehicles, it specifically includes:

[0027] S31: Behavioral feature extraction, continuously track vehicles around the target and extract behavioral feature vectors, including longitudinal acceleration, longitudinal deceleration, lateral displacement, turn signal status and trajectory curvature change rate;

[0028] S32: Behavioral anomaly quantification: Based on the current scene context, calculate the anomaly score of the target's surrounding vehicle behavior feature vectors relative to their expected normal behavior feature vectors.

[0029] S33: Potential risk area derivation. Based on the spatial location and behavioral characteristics of surrounding vehicles whose behavior anomaly scores exceed a preset threshold, and combined with the road topology, the potential risk areas outside the field of view of the vehicle's sensors and their risk types are deduced.

[0030] S34: Virtual observation data generation. Based on the derived risk areas and their risk types, corresponding virtual observation data is generated and assigned a synthetic confidence level based on behavioral anomaly and derivation logic credibility. This virtual observation data is then added to the environmental model.

[0031] Preferably, when the data supplementation module supplements the tested data by constructing a causal graph and inferring potential hazards based on historical and current changes in the surrounding environment, it specifically includes:

[0032] S41: Cause-effect graph construction. Construct a cause-effect graph model containing multi-level nodes. The node set of the cause-effect graph model includes environmental state nodes, intermediate event nodes, and potential danger nodes. The set of directed edges of the cause-effect graph model represents the causal relationship between nodes.

[0033] S42: Evidence injection and node activation. The verified current environmental observation data is used as evidence and injected into the corresponding environmental state node in the causal graph. The downstream intermediate event nodes and potential danger nodes are activated according to the causal logic.

[0034] S43: Risk probability propagation and calculation: Through a causal network, the probability of evidence is propagated along the directed edges to calculate the posterior probability of each potentially dangerous node being triggered.

[0035] S44: Risk Decision Making and Virtual Data Generation: Potential dangers with posterior probabilities exceeding a preset alert threshold are identified as valid risks, and corresponding virtual threat data and their confidence levels are generated and added to the environment model.

[0036] Preferably, when the data fusion module fuses the verified and supplemented multimodal data based on the confidence levels of each data source to generate a unified environmental model, it specifically includes:

[0037] S51: Confidence normalization and calibration. Receives outputs from various data sources, including observations and their corresponding raw confidence levels, from the data verification module and the data supplementation module. Performs cross-modal normalization on the raw confidence levels to obtain comparable standard confidence levels.

[0038] S52: Calculate the fusion weights. Based on the obtained standard confidence scores, assign fusion weights to each data source and calculate the spatial consistency factor among the data sources.

[0039] S53: Weighted fusion estimation, which uses fusion weights to weight and fuse the observations from each data source to obtain the optimal estimate of the target state;

[0040] S54: Confidence assessment of the fusion result. Based on the consistency of the fusion weights, the distribution of confidence in each data source, and the fusion residuals, the overall confidence of the fused result is calculated.

[0041] Preferably, the scene classification module, based on the fused environment model, calculates the hazard factor of each element in the environment, classifies elements with hazard factors below a set threshold as obstacles, and determines the current driving scene type based on elements with hazard factors above the set threshold. Specifically, this includes:

[0042] S61: Environmental element hazard factor calculation, based on the unified environmental model generated by the data fusion module, calculates the comprehensive hazard factor for each dynamic and static element in the model;

[0043] S62: Element classification based on hazard factors. The hazard factors of each element are compared with preset static obstacle thresholds and dynamic risk thresholds to classify environmental elements.

[0044] S63: Scene feature vector construction: Based on the classification results, extract the macroscopic features of the current environment and construct the scene feature vector;

[0045] S64: Driving scenario classification decision, inputting the scenario feature vector into the scenario classifier for calculation, and outputting the current specific driving scenario type.

[0046] Preferably, the database retrieval module includes the following steps when it is in operation:

[0047] S71: Scene type reception, receiving the current driving scene type identifier output from the scene classification module;

[0048] S72: Multi-level scene matching, based on the current driving scene type identifier, searches in the preset scene database, including two levels: exact matching and fuzzy matching;

[0049] S73: Acquisition of related information: Retrieve potential hazard information and preset operation strategy library associated with the successfully matched scenario from the scenario database;

[0050] S74: Strategy Priority Ranking. The applicability of the retrieved preset operation strategies is evaluated based on the current environmental parameters, and then ranked according to priority.

[0051] Preferably, the vehicle real-time control module, when in operation, specifically includes:

[0052] S81: Policy security verification. Receives a preset operation policy sequence from the database retrieval module, verifies each policy against a preset set of stringent security rules, and filters out a subset of feasible policies that do not violate any security rules.

[0053] S82: Multi-objective optimization decision-making, for each strategy in the subset of feasible strategies, calculates its comprehensive utility value based on the multi-objective optimization function;

[0054] S83: Optimal instruction generation, selects the strategy with the highest comprehensive utility value as the final execution strategy, and decomposes it into specific vehicle-level control instructions;

[0055] S84: Real-time safety monitoring and intervention. During the execution of instructions, the vehicle status and environmental changes are continuously monitored. Once a situation that is about to violate safety rules is detected, the safety intervention mechanism is immediately triggered.

[0056] The beneficial effects of this invention are:

[0057] 1. Compared to existing technologies that typically use simple threshold detection or fixed rules to remove abnormal data, this method is difficult to adapt to complex and ever-changing real road conditions and is prone to mistakenly deleting valid data or retaining invalid data due to rigid rules. This invention introduces a dynamic adaptive mutation detection mechanism based on a combination of physical kinematic model prediction and historical data trend analysis, and further combines multi-sensor data grouping and a collaborative removal strategy that integrates Mahalanobis distance and source confidence. This solution can more intelligently identify unreasonable data caused by transient interference or sensor failure, and comprehensively evaluate the reliability and consistency of each data point when there is conflict in multi-sensor data, thereby significantly improving the quality and reliability of input data and providing a more solid data foundation for subsequent decision-making.

[0058] 2. Compared to existing technologies that mainly rely on direct sensor observation, which lacks the ability to predict potential risks within sensor blind spots and long-term threats based on logical reasoning, this invention innovatively sets up two supplementary reasoning paths: first, by analyzing the abnormality of the behavioral characteristics of surrounding vehicles to infer the risk areas outside their field of vision and generate virtual observation data; second, by constructing a multi-level causal graph model that includes environmental state, intermediate events, and potential hazards, and using Bayesian networks to calculate risk probabilities. This enables the system to predict risks from indirect clues, achieving a leap from perceiving the current situation to predicting the future, and significantly enhancing the system's forward-looking safety decision-making capabilities.

[0059] 3. Compared with existing technologies that often focus on the confidence level of each data source or simple spatial weighting when fusing multi-source data, failing to fully consider the mutual verification of spatial consistency between different sources of data, this invention introduces a spatial consistency factor to correct the weights based on the initial fusion weights calculated from the normalized standard confidence level. Finally, it evaluates the comprehensive confidence level of the fusion result from three dimensions: weighted average confidence level, fusion consistency, and weight concentration. This mechanism ensures that even if a sensor has a high confidence level, its influence will be suppressed if its observations differ significantly from most other sensors, thereby effectively improving the robustness and accuracy of the fusion result.

[0060] 4. Compared to existing technologies that rely heavily on preset rules or simple identification of target types for classifying driving scenarios, lacking in-depth quantitative assessment of the overall risk level and essential characteristics of the scenario, this invention calculates a comprehensive risk factor for each environmental element by integrating multiple dimensions such as collision time, relative distance, target type, behavioral characteristics, and trajectory uncertainty (quantified by information entropy), and then performs fine classification of the elements based on thresholds. Furthermore, after constructing the scenario feature vector, an attention mechanism model incorporating risk urgency factors is used for scenario classification. This method enables the system not only to identify objects in the environment but also to understand the potential threat level and interactive relationships within the scenario, achieving a shift towards proactive cognition.

[0061] 5. Compared to existing technologies where vehicle control modules may prioritize safety as one of their optimization objectives or employ relatively singular intervention strategies in emergencies, posing a risk of breaching safety baselines, this invention innovatively places a stringent set of safety rules as insurmountable hard constraints at the forefront of decision-making, pre-verifying and screening all candidate strategies. Subsequently, multi-objective utility optimization is performed on feasible strategies, and real-time safety monitoring is conducted in conjunction with model predictive control during execution. A layered emergency avoidance mechanism is also designed. This scheme ensures that the system prioritizes safety at all times, thereby providing a higher level of safety assurance in complex and dynamic environments. Attached Figure Description

[0062] Figure 1 The diagram shown is a structural schematic of the intelligent connected vehicle safe driving decision control system of the present invention.

[0063] Figure 2 The diagram illustrates the workflow of the scene classification module in the intelligent connected vehicle safe driving decision control system of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0065] Please see Figures 1-2 This invention provides an embodiment: an intelligent connected vehicle safe driving decision control system, comprising:

[0066] I. Data Acquisition Module

[0067] It is used to collect multimodal data of the vehicle's surrounding environment through LiDAR, cameras and millimeter-wave radar.

[0068] II. Data Validation Module

[0069] Used to verify the collected multimodal data.

[0070] In this embodiment, the data verification module uses the following verification methods when verifying the collected multimodal data:

[0071] A11: Perform data mutation detection and remove or fill in any unreasonable mutation data identified;

[0072] A12: Group the data according to the difference between the data points, and within each group, remove the data points that are judged to be the least reliable, based on the confidence level of each data point and the distribution of data within the group.

[0073] On the one hand, when the data verification module performs data mutation detection and removes and fills in unreasonable mutation data, it specifically includes:

[0074] S11: Data mutation detection. For a data point at the current time from any sensor data stream, calculate the weighted residual between its predicted value based on the physical kinematics model and the actual observed value. Compare the weighted residual with a dynamic adaptive threshold. If the weighted residual is greater than the dynamic adaptive threshold, it is determined to be unreasonable mutation data. The formula for calculating the weighted residual is:

[0075] ;

[0076] in, For weighted residuals, For the current data point, This represents the predicted value for the current moment, obtained using a physical kinematic model based on valid data from the previous moment. This is a moving average of a recent period of valid historical data. and These are the weighting coefficients;

[0077] The formula for calculating the dynamic adaptive threshold is as follows:

[0078] ;

[0079] in, For dynamic adaptive threshold, It is an adjustable scaling factor. This represents the standard deviation of a recent period of valid historical data. Environmental interference coefficient;

[0080] S12: Remove unreasonable data. Remove data points that are determined to be unreasonable mutations from the valid data sequence and mark them as missing values.

[0081] S13: Missing data imputation. For missing values ​​resulting from data removal, imputation is performed using the best estimate based on historical valid data sequences and sensor fusion predictions. The best estimate is calculated as follows:

[0082] ;

[0083] in, This is the best estimate. To integrate weights, This represents the predicted value for the current moment, obtained using a physical kinematic model based on valid data from the previous moment. It is a weighted fusion value of verified valid observations of the same target from other cooperating sensors at time t.

[0084] On the other hand, when the data validation module groups the data based on the difference between data points, and within each group, combines the confidence level of each data point with the data distribution within the group to remove the data points deemed least reliable, it specifically includes:

[0085] S21: Data grouping, based on spatiotemporal correlation, divide the observation data of different sensors on the same physical entity into the same data group;

[0086] S22: Intragroup data distribution and confidence assessment. For each data point within the same data group, calculate its consistency measure with other data points and obtain its own source confidence.

[0087] S23: Calculation of the overall distrust score. A comprehensive distrust score is calculated for each data point within the same dataset. This score is a function of its data inconsistency measure and its source confidence level. The data inconsistency measure is represented by the Mahalanobis distance between the data point and the data centers within the same dataset. The calculation formula is as follows:

[0088] ;

[0089] in, As a measure of data inconsistency, For the current data point, For the same data group, This is a multidimensional mean vector of all data points in the same dataset. Represents the matrix transpose operation. The covariance matrix of the same data group is used to characterize the distribution shape and correlation of the data within the group;

[0090] The formula for calculating the overall distrust score is as follows:

[0091] ;

[0092] in, To incorporate the distrust score, The source confidence level of the current data point is determined by the reliability of the sensor itself, the current environmental conditions, and the confidence level of the target recognition algorithm.

[0093] S24: Outlier removal. The data points in the group are sorted according to the overall distrust score. The data points with the highest scores that exceed the set threshold are identified as outliers and removed.

[0094] Specifically, the following criteria are used when grouping data:

[0095] Spatial proximity criterion: the distance between the target represented by the data point and the spatial coordinates is less than a set threshold;

[0096] Time alignment criteria: Data points' timestamps must be within the same sampling period;

[0097] Target type similarity criterion: Data points are identified as having the same or compatible target types.

[0098] Specifically, the criteria for outlier removal are as follows:

[0099] If a data point satisfies If it is an outlier, it is identified as an outlier and removed. A sensitivity coefficient greater than 1 This is the average of the combined distrust scores for all data points within the same data set.

[0100] In this embodiment, firstly, a data mutation detection step is implemented. This step compares the weighted residuals of predicted values ​​based on a physical kinematic model and historical data trends with a dynamically adaptive threshold to accurately identify and eliminate unreasonable mutation data. Subsequently, intelligent filling based on sensor fusion ensures the continuity and rationality of the data stream. Secondly, through data grouping and outlier removal steps, multi-sensor data is grouped according to spatiotemporal correlation. An innovative approach combines Mahalanobis distance and source confidence to calculate a comprehensive distrust score for each data point, thereby intelligently identifying and removing the most unreliable data points within each group. This scheme, through these two progressive verification mechanisms, significantly improves the reliability and consistency of input data, effectively avoiding the "garbage in, garbage out" problem. This lays a high-quality data foundation for subsequent data fusion, scene understanding, and vehicle control, thereby enhancing the perception robustness and decision-making safety of the entire intelligent driving system in complex real-world environments.

[0101] III. Data Supplement Module

[0102] This is used to supplement the reasoning of the tested data. The supplementary methods include: inferring the environmental state based on the behavior of surrounding vehicles; and constructing a causal graph and inferring potential dangers based on the historical and current changes of the surrounding environment.

[0103] On the one hand, when the data supplementation module supplements and infers the verified data by inferring the environmental state based on the behavior of surrounding vehicles, it specifically includes:

[0104] S31: Behavioral feature extraction, continuously track vehicles around the target and extract behavioral feature vectors, including longitudinal acceleration, longitudinal deceleration, lateral displacement, turn signal status and trajectory curvature change rate;

[0105] S32: Behavioral Anomaly Measurement. Based on the current scene context, calculate the anomaly score of the target's surrounding vehicle behavior feature vectors relative to their expected normal behavior feature vectors. The formula for calculating the anomaly score in behavioral anomaly measurement is as follows:

[0106] ;

[0107] in, The abnormality score, For the extracted behavioral feature vector, This is the expected normal behavior feature vector obtained by the behavior prediction model based on the current scene context. This represents the historical standard deviation of the fluctuations in traffic behavior characteristics within the current specific scenario context. It is a very small positive value;

[0108] S33: Potential risk area derivation. Based on the spatial location and behavioral characteristics of surrounding vehicles whose behavior anomaly scores exceed a preset threshold, and combined with the road topology, the potential risk areas outside the field of view of the vehicle's sensors and their risk types are deduced.

[0109] S34: Virtual observation data generation. Based on the derived risk areas and their risk types, corresponding virtual observation data is generated and assigned a composite confidence score based on behavioral anomaly and logical credibility. This virtual observation data is then added to the environmental model. The formula for calculating the composite confidence score is as follows:

[0110] ;

[0111] in, The composite confidence level of virtual observation data. The abnormality score is the normalized score. The confidence level of the applied derivation rule itself. Let be the time decay function, representing the derived characteristic that the risk decreases over time.

[0112] Specifically, the expected normal behavior feature vector is determined in the following way:

[0113] For following other vehicles, predictions are mainly based on the historical speed of the vehicle in front, the speed of the vehicle itself, and a safe following distance model.

[0114] For intersection scenarios, predictions are mainly based on traffic rules (such as right-of-way rules) and typical traffic trajectories;

[0115] For lane keeping scenarios, predictions are mainly based on the lane centerline and the current vehicle speed.

[0116] Specifically, in the derivation of potential risk regions, the derivation follows predefined geometric and causal mapping rules, including:

[0117] If a vehicle in the same lane ahead brakes unexpectedly, it can be inferred that there may be a static obstacle, a construction area, or an animal crossing in the area directly in front of it.

[0118] If a vehicle on the side suddenly deviates from its lane away from the vehicle without a clear reason, it can be inferred that there may be temporary obstacles or road damage in the area adjacent to its original lane.

[0119] If a vehicle on the side accelerates in violation of the expected yield rule at an intersection without traffic lights, it can be inferred that there is a risk of collision at the intersection's conflict point.

[0120] On the other hand, when the data supplementation module supplements the examined data by constructing causal graphs and inferring potential dangers based on historical and current changes in the surrounding environment, it specifically includes:

[0121] S41: Cause-effect graph construction. Construct a cause-effect graph model containing multi-level nodes. The node set of the cause-effect graph model includes environmental state nodes, intermediate event nodes, and potential danger nodes. The set of directed edges of the cause-effect graph model represents the causal relationship between nodes.

[0122] S42: Evidence injection and node activation. The verified current environmental observation data is used as evidence and injected into the corresponding environmental state node in the causal graph. The downstream intermediate event nodes and potential danger nodes are activated according to the causal logic.

[0123] S43: Risk probability propagation and calculation: Through a causal network, the probability of evidence is propagated along the directed edges to calculate the posterior probability of each potentially dangerous node being triggered.

[0124] S44: Risk Decision Making and Virtual Data Generation: Potential dangers with posterior probabilities exceeding a preset alert threshold are identified as valid risks, and corresponding virtual threat data and their confidence levels are generated and added to the environment model.

[0125] Specifically, typical causal chains included in causal graph models include:

[0126] Observing an intersection ahead, it is inferred that there are vehicles / pedestrians waiting to cross from the side, and further, it is inferred that there is a potential danger: collision at the intersection.

[0127] Observing that items fell from the vehicle ahead, it was deduced that the vehicle door was not closed properly or the cargo was not securely fixed. Further, it was deduced that there was a potential danger: the vehicle ahead might brake suddenly or there was an obstacle on the road.

[0128] The presence of toys such as balls on the road surface suggests that children are playing nearby, and further suggests a potential danger: children may suddenly run onto the road.

[0129] In this embodiment, the risk probability propagation calculation adopts an update rule based on Bayesian networks, and the formula for calculating the posterior probability is:

[0130] ;

[0131] For complex multi-cause-one-effect relationships, the calculation formula is extended as follows:

[0132] ;

[0133] in, This represents a potentially dangerous assumption. As evidence variables, For multiple evidence variables, For posterior probability, The marginal probability of the evidence, i.e., the probability of observing the evidence. The probability, Represents a set of observed evidence, Risk assumption The probability of occurrence Let f(x) be the likelihood function, representing the likelihood under the risk assumption. Evidence observed in cases where it occurs The probability, For multiple evidence posterior probabilities, For joint likelihood probability, This is the summation of all probabilities, i.e., the overall probability of the evidence.

[0134] In this embodiment, a "noise-OR" model is used to process the contribution of multiple independent pieces of evidence to the same danger. Therefore, the posterior probability formula for calculating the non-occurrence of the danger is:

[0135] ;

[0136] in, It is derived from prior probability and likelihood. The joint posterior probability represents the probability when all evidence is observed simultaneously. Under these conditions, danger The probability that it will not happen.

[0137] In this embodiment, the risk decision-making rules in risk decision-making and virtual data generation are as follows:

[0138] If a potentially dangerous node The posterior probability satisfies and If so, the danger is deemed valid;

[0139] in, A sensitivity coefficient greater than 1 This is the absolute probability threshold.

[0140] In this embodiment, the formula for calculating the confidence level of the virtual threat data generated for valid risks is as follows:

[0141] ;

[0142] in, The confidence level of virtual threat data. This represents the confidence level of the causal graph model itself.

[0143] In this embodiment, the data supplementation module enhances the intelligence level of environmental perception through two innovative approaches: First, it infers the environmental state based on the behavior of surrounding vehicles, extracting behavioral features, quantifying anomalies, deducing potential risk areas, and generating virtual observation data to predict risks within sensor blind spots. Second, it constructs a causal graph model, using a Bayesian network to transform environmental evidence into posterior probabilities of potential hazards, and processes multi-evidence associations through a "noise-OR" model to achieve risk prediction based on logical reasoning. This solution, through the deep integration of behavioral analysis and causal reasoning, enables the intelligent driving system to "see the big picture from small details," predicting potential hazards from indirect clues, significantly improving the system's forward-looking safety decision-making capabilities and proactive safety level in complex scenarios.

[0144] IV. Data Fusion Module

[0145] It is used to fuse verified and supplemented multimodal data based on the confidence levels of each data source to generate a unified environmental model.

[0146] In this embodiment, the data fusion module, based on the confidence levels of each data source, fuses the verified and supplemented multimodal data to generate a unified environmental model, specifically including:

[0147] S51: Confidence Normalization and Calibration. This function receives outputs from various data sources, including observations and their corresponding raw confidence levels, from the data verification and data supplementation modules. It then performs cross-modal normalization on the raw confidence levels to obtain comparable standard confidence levels. The normalization process employs a linear transformation method based on sensor prior performance. The principle formula is as follows:

[0148] ;

[0149] in, The standard confidence level is obtained after normalization of the i-th data source. Let be the raw confidence level output by the sensor itself from the i-th data source. For the calibration coefficient of the sensor type, These are calibration coefficients relevant to the current environment.

[0150] S52: Fusion weight calculation. Based on the obtained standard confidence score, a fusion weight is assigned to each data source, and the spatial consistency factor among the data sources is calculated. The fusion weight calculation adopts a soft competition allocation mechanism based on the standard confidence score. The principle formula is as follows:

[0151] ;

[0152] in, Let i be the fusion weight of the i-th data source. As a competing factor, It is an exponential function. The standard confidence level for the j-th data source is... This is an index variable used to iterate through all data sources;

[0153] S53: Weighted fusion estimation, which uses fusion weights to weight and fuse the observations from each data source to obtain the optimal estimate of the target state;

[0154] S54: Confidence assessment of the fusion result. Based on the consistency of the fusion weights, the distribution of confidence in each data source, and the fusion residuals, the overall confidence of the fused result is calculated.

[0155] In this embodiment, a spatial consistency factor is introduced to correct the initial weights during the fusion weight calculation. The corrected final weights are as follows:

[0156] ;

[0157] in, Let be the final weight of the i-th data source. For example, sensitivity coefficient, Let be the spatial distance between the observations of data source i and the observations of data source j. It is a very small positive value. This indicates summing over all data sources except the i-th data source.

[0158] In this embodiment, the formula for calculating the weighted fusion estimate is as follows:

[0159] ;

[0160] in, This is a weighted fusion estimate. Let be the observation value from the i-th data source.

[0161] In this embodiment, the overall confidence level of the fusion result in the confidence assessment of the fusion result is composed of the following three coupled parts:

[0162] ;

[0163] in, For weighted average confidence level, To integrate consistency factors, As the weight concentration factor, Let be the entropy of the weight distribution.

[0164] In this embodiment, the raw confidence scores from different sensors are first calibrated using cross-modal normalization to ensure comparability. Then, a weighting calculation method based on a soft-competition allocation mechanism is used to assign fusion weights according to the standard confidence scores of each data source, and a spatial consistency factor is introduced for correction to improve the rationality of the weight allocation. Next, the optimal estimate of the target state is generated through weighted fusion estimation. Finally, the confidence score of the fusion result is comprehensively evaluated from three dimensions: weighted average confidence score, fusion consistency, and weight concentration. This scheme effectively improves the accuracy and robustness of multi-source data fusion through confidence-driven intelligent weighting and spatial consistency verification. The generated environmental model is more accurate and reliable, providing a high-quality unified perception foundation for subsequent scene classification and decision control, and significantly enhancing the perception capability and decision safety of the intelligent driving system in complex and changing environments.

[0165] V. Scene Classification Module

[0166] Based on the fused environment model, it calculates the hazard factor of each element in the environment, classifies elements with hazard factors below a set threshold as obstacles, and determines the current driving scenario type based on elements with hazard factors above a set threshold.

[0167] In this embodiment, the scene classification module calculates the hazard factor of each element in the environment based on the fused environment model, classifies elements with hazard factors below a set threshold as obstacles, and determines the current driving scene type based on elements with hazard factors above the set threshold. Specifically, this includes:

[0168] S61: Environmental element hazard factor calculation. Based on the unified environmental model generated by the data fusion module, a comprehensive hazard factor is calculated for each dynamic and static element in the model. The formula for calculating the comprehensive hazard factor is as follows:

[0169] ;

[0170] in, For element i, the comprehensive risk factor, Let be the estimated collision time between element i and the vehicle. This is an inverse function of collision time; the smaller the value, the higher the danger. Let i be the relative distance between element i and the vehicle. It is a negative exponential function of distance. For element type, For type-based basic risk weights, For the behavioral characteristics of elements, For behavior-based risk enhancement, It is a measure of the uncertainty of the trajectory of an element. The added risk brought about by uncertainty , , , and These are the weighting coefficients for each factor;

[0171] S62: Element classification based on hazard factors. The hazard factors of each element are compared with preset static obstacle thresholds and dynamic risk thresholds to classify environmental elements.

[0172] S63: Scene feature vector construction: Based on the classification results, extract the macroscopic features of the current environment and construct the scene feature vector;

[0173] S64: Driving scenario classification decision. The scenario feature vector is input into the scenario classifier for calculation, and the current specific driving scenario type is output. The scenario classifier adopts a multilayer perceptron model based on an attention mechanism, and its decision function is:

[0174] ;

[0175] in, The output of the scene classifier, For scene feature vectors, The attention weight vector is calculated using the attention mechanism. This indicates element-wise multiplication. , , and For neural network parameters, Represents the ReLU activation function. The function transforms the output into a probability distribution. This represents selecting the scenario type with the highest probability as the classification result.

[0176] Specifically, the uncertainty measure of element trajectories The entropy of trajectory prediction is used for quantification, and the calculation formula is as follows:

[0177] ;

[0178] in, Let i be the probability that element i will be on a certain predicted path at time t in the future. This indicates that the prediction results are accumulated over the total number of time steps of the prediction.

[0179] In this embodiment, the element classification rule used in the element classification based on hazard factors is as follows:

[0180] if If so, element i is classified as a fixed obstacle, and only basic obstacle avoidance planning is required;

[0181] if If element i is classified as a normal moving obstacle, it needs to be tracked and collaboratively planned.

[0182] if If so, element i will be classified as a high-risk element and will participate in subsequent scenario type judgment;

[0183] in, For static obstacle thresholds, This is a dynamic risk threshold.

[0184] Specifically, when constructing scene feature vectors, the scene feature vectors include features in the following dimensions:

[0185] Spatial distribution characteristics: the orientation (front, back, left, right) and density of high-risk elements relative to the vehicle;

[0186] Road topology features: Current road structure (straight roads, curves, intersections, ramps);

[0187] Interaction relationship characteristics: the relative movement relationship between high-risk elements (same direction, opposite direction, intersection).

[0188] In this embodiment, the calculation of the attention weight vector incorporates a risk urgency factor, and the calculation formula is as follows:

[0189] ;

[0190] ;

[0191] in, The attention weight for the k-th feature. For the k-th scene feature, This is the original score for the k-th feature. The set of scores for all features. As the greatest risk factor, As a risk urgency factor, This means that the three variables are concatenated to form a new feature vector. Here is the attention weight matrix. For bias vectors, This is the transpose of the output weight vector.

[0192] In this embodiment, the hazard factor of each environmental element is first calculated by comprehensively considering collision time, relative distance, target type, behavioral characteristics, and trajectory uncertainty (quantified by information entropy), and then classified according to thresholds. Next, macroscopic features such as the spatial distribution, road topology, and interaction relationships of high-risk elements are extracted to construct a scene vector. Finally, a dynamic weighted key feature model incorporating risk urgency factors is used to output the specific scene type. This scheme transforms abstract driving scenarios into computable risk models, enabling the system not only to identify environmental elements but also to understand the potential threat level and essential characteristics of the scene. This achieves a leap from passive perception to proactive cognition, providing a deeper semantic basis for subsequent safety decisions and significantly improving the situational awareness and decision-making intelligence of the intelligent driving system in complex environments.

[0193] VI. Database Retrieval Module

[0194] It is used to retrieve potential hazard information and preset operation strategies associated with the current driving scenario from the preset scenario database based on the determined current driving scenario type.

[0195] In this embodiment, the database retrieval module includes the following steps when it is in operation:

[0196] S71: Scene type reception, receiving the current driving scene type identifier output from the scene classification module;

[0197] S72: Multi-level scene matching, based on the current driving scene type identifier, searches in the preset scene database, including two levels: exact matching and fuzzy matching;

[0198] S73: Acquisition of related information: Retrieve potential hazard information and preset operation strategy library associated with the successfully matched scenario from the scenario database;

[0199] S74: Strategy Priority Ranking. The applicability of the retrieved preset operation strategies is evaluated based on the current environmental parameters, and then ranked according to priority.

[0200] Specifically, multi-level scene matching includes:

[0201] First, a precise match is performed, comparing the received scene type identifier with the predefined standardized scene identifiers in the scene database. If a completely identical scene identifier is found, all associated information under that scene is directly retrieved.

[0202] If exact matching fails, fuzzy matching is initiated, calculating the similarity between the feature vector of the current scene and the feature vectors of each scene in the scene database, and selecting multiple scenes with similarity exceeding a set threshold as candidate scenes.

[0203] Preferably, scene similarity is calculated based on the following feature dimensions:

[0204] The combination of the number and types of high-risk elements;

[0205] The spatial distribution of high-risk elements relative to the vehicle;

[0206] Types of road topology;

[0207] Traffic rules and constraints.

[0208] Specifically, in the acquisition of related information, the database of potential hazards includes:

[0209] Description of typical hazardous events related to the current scenario type;

[0210] The prior probability of each dangerous event occurring;

[0211] Possible paths for the evolution of dangerous events.

[0212] Specifically, in the acquisition of related information, the preset operation strategy library includes:

[0213] Avoidance strategies for specific dangerous events, including but not limited to one or more combinations of emergency braking, comfort deceleration, and steering avoidance;

[0214] The triggering conditions and constraints for each operational strategy;

[0215] The expected effects and potential associated risks after the implementation of each strategy.

[0216] Specifically, strategy priority ranking includes:

[0217] Adaptive assessment involves matching the retrieved preset operational strategies with the real-time environment model to evaluate the feasibility of each strategy under the current specific parameters.

[0218] Prioritization is performed based on the evaluation results, taking into account the safety margin, comfort impact, and execution efficiency of the strategy. The strategies are then ranked from highest to lowest score to generate a recommended strategy sequence.

[0219] Specifically, in adaptability assessments, the criteria for evaluating the feasibility of strategies include:

[0220] The dynamic distance and speed difference between this vehicle and key surrounding targets;

[0221] Do the current road conditions meet the spatial requirements for strategy execution?

[0222] Does the strategy implementation comply with the mandatory constraints of traffic regulations?

[0223] In this embodiment, a combination of precise matching and fuzzy matching is first used to quickly retrieve a database of potential hazards and preset operating strategies related to the current driving scenario from a pre-defined scenario database. Then, an adaptive evaluation mechanism is used to verify the feasibility of the strategies in conjunction with real-time environmental parameters, and a multi-objective comprehensive score and priority ranking are performed based on safety margin, comfort, and execution efficiency. This scheme effectively transforms abstract scenario types into specific, executable sequences of risk avoidance strategies, ensuring both the standardization of decision-making and giving the system the ability to generalize to unknown scenarios. This significantly improves the response speed and reliability of the intelligent driving system in rapidly invoking prior knowledge and generating optimal decisions in complex environments.

[0224] VII. Vehicle Real-time Control Module

[0225] It is used to receive preset operating strategies and, under the constraints of preset stringent safety rules, generate vehicle control commands and execute real-time control.

[0226] Preferably, the vehicle real-time control module, when in operation, specifically includes:

[0227] S81: Policy security verification. Receives a preset operation policy sequence from the database retrieval module, verifies each policy against a preset set of stringent security rules, and filters out a subset of feasible policies that do not violate any security rules.

[0228] S82: Multi-objective optimization decision-making. For each policy in the subset of feasible policies, its comprehensive utility value is calculated based on a multi-objective optimization function. The multi-objective optimization function is used to calculate the comprehensive utility value of the policy, and the specific formula is as follows:

[0229] ;

[0230] in, Represents a multi-objective optimization function. To score safely, For comfort score, To score for efficiency, Score based on priority. , , and These are the weighting coefficients for each objective;

[0231] S83: Optimal instruction generation, selects the strategy with the highest comprehensive utility value as the final execution strategy, and decomposes it into specific vehicle-level control instructions;

[0232] S84: Real-time safety monitoring and intervention. During the execution of instructions, the vehicle status and environmental changes are continuously monitored. Once a situation that is about to violate safety rules is detected, the safety intervention mechanism is immediately triggered.

[0233] In this embodiment, the stringent security rule set includes inviolable hard constraints, and the verification rules include:

[0234] The minimum distance between the predicted trajectory after strategy execution and any obstacle must be greater than the absolute safe distance;

[0235] The strategy must comply with traffic regulations, including but not limited to: prohibiting crossing solid lines, obeying traffic lights, and yielding rules;

[0236] The vehicle dynamics state generated by the strategy (such as lateral acceleration and yaw rate) must be within the vehicle stability envelope;

[0237] The acceleration impact generated by the strategy must be below the physiological limits that the occupants can withstand.

[0238] In this embodiment, the safety score is calculated using a piecewise function based on the safety margin, and the specific formula is as follows:

[0239] ;

[0240] in, This is the predicted minimum distance between the vehicle and the nearest obstacle after the strategy is implemented. For a comfortable and safe distance, This is an absolutely dangerous distance. This represents the score sensitivity coefficient.

[0241] In this embodiment, the security intervention mechanism in real-time security monitoring and intervention adopts a model predictive control-based security control algorithm, and the intervention conditions are:

[0242] If the deviation between the actual state of the vehicle and the predicted trajectory satisfies:

[0243] ;

[0244] Or the minimum distance between the newly detected obstacle and the vehicle's predicted trajectory satisfies:

[0245] ;

[0246] The current strategy will be immediately terminated, and a preset emergency avoidance strategy will be switched.

[0247] in, The deviation between the actual vehicle state and the predicted trajectory. For safety reasons, The maximum allowable state deviation threshold, The minimum distance between the newly detected obstacle and the vehicle's predicted trajectory. Current vehicle speed This represents the system response time.

[0248] Specifically, the emergency avoidance strategy adopts a tiered decision-making mechanism:

[0249] First priority: Apply maximum braking while maintaining stability;

[0250] Second priority: When the braking distance is insufficient, perform emergency steering within the stability envelope to avoid an obstacle;

[0251] Third priority: When a collision is unavoidable, adopt an impact angle optimization strategy to minimize collision damage.

[0252] The parameters of the stringent safety rule set can be adaptively adjusted according to vehicle status and environmental conditions. The specific formula for how safe distance changes with vehicle speed is as follows:

[0253] ;

[0254] in, For the minimum stationary distance, For the vehicle's speed, To aim at the timing head, This is the maximum deceleration of the vehicle. This represents the maximum deceleration of the vehicle in front.

[0255] In this embodiment, the preset strategy is first rigorously validated against safety rules to eliminate any schemes that might violate collision avoidance, traffic regulations, vehicle stability, and occupant stress limits. Then, a multi-objective optimization function is used to comprehensively evaluate the utility of feasible strategies, taking into account safety, comfort, efficiency, and strategy priority. This generates the optimal vehicle control command. Finally, through real-time predictive monitoring, a layered emergency avoidance mechanism is immediately triggered when a state deviation or sudden danger is detected. This innovative approach places safety rules as insurmountable hard constraints at the decision-making stage and combines them with adaptive safety parameters and utility optimization. This achieves intelligent selection of the optimal control strategy while absolutely ensuring a safety baseline, significantly improving the real-time response capability, risk avoidance capability, and overall driving experience of the intelligent driving system in complex dynamic environments.

[0256] Example 1: Avoiding potential hazards on highway mainlines under adverse weather conditions

[0257] Scenario Description: A connected vehicle (autonomous vehicle) is traveling at a high speed on a highway. The weather is raining heavily, and visibility is low. A large container truck is approaching in the same lane ahead. Due to the truck's size, it completely obstructs the vehicle's sensors from directly perceiving the road conditions further ahead, creating a typical sensor blind spot.

[0258] Implementation process:

[0259] Data Acquisition and Verification: The system's data acquisition module continuously collects environmental data via LiDAR, cameras, and millimeter-wave radar. Heavy rain can introduce noise into sensor data; for example, camera accuracy may decrease due to water splashes, and LiDAR point clouds may fluctuate abnormally due to raindrops. The data verification module immediately activates its two-layer verification mechanism. First, it uses a physical kinematics model (e.g., predicting the relative distance between the vehicle and the truck based on their previous speed and position) to detect unreasonable abrupt changes in the distance data transmitted by the millimeter-wave radar, classifying this as noise and removing it. Simultaneously, it intelligently fills in the gaps using historical data and valid observations from other sensors. Next, the module groups the observation data of the truck ahead from different sensors (e.g., visually recognized contours, radar-detected distance and speed) and calculates a comprehensive distrust score for each data point. Because the source confidence of the camera is automatically lowered in rainy weather, some data points with blurred boundaries are identified as outliers and removed due to inconsistencies with the high-confidence radar data, thus ensuring the accuracy and consistency of the truck's status information.

[0260] Data Supplementation and Inference: The data supplementation module begins its work. Although it cannot directly "see through" the truck, the module closely monitors its behavior. It detects that the truck suddenly braked with an unusually large amount of force, causing its behavior anomaly score to rise sharply. Based on the causal mapping rule of "unexpected emergency braking by a vehicle in the same lane ahead," the module infers that there may be potential risks such as static obstacles or severe congestion in the blind spot area directly in front of the truck. Therefore, the module generates "virtual observation data" (such as a hypothetical obstacle) representing this risk and assigns it a synthetic confidence score based on the truck's behavior anomaly score and the credibility of the inference rule, supplementing the environmental model with this virtual data.

[0261] Data Fusion and Environmental Modeling: The data fusion module combines verified real sensor data (such as the exact location and speed of the truck) with virtual risk data generated by the data supplementation module. While the virtual data has a lower weight than directly observed data when calculating the fusion weights, it is still included in the unified environmental model due to its underlying logic. The generated environmental model not only includes the truck ahead as an entity but also marks a high-probability risk area in front of it.

[0262] Scene Classification: The scene classification module performs calculations based on the fused environment model. The truck ahead is assigned a high-risk factor due to its proximity and high speed, while the additionally inferred risk area is also assigned a high-risk factor due to its uncertainty and potential for high threat. The module integrates features such as the spatial distribution of high-risk elements (directly ahead) and road topology (straight roads), and uses a classification model with an attention mechanism to classify the current scene as a high-risk scenario of "potential sudden obstacles ahead during high-speed following."

[0263] Database Retrieval: Upon receiving the scenario type, the database retrieval module first performs a precise match and successfully retrieves relevant information from the preset scenario database. The database provides typical potential hazards in this scenario (such as rear-end collisions and multi-vehicle pileups) as well as a library of preset operating strategies, including "comfortable deceleration and gradually increasing following distance," "preparing to change lanes," and "emergency braking."

[0264] Real-time vehicle control: The strategy sequence received by the real-time vehicle control module is first validated against a rigorous set of safety rules. For example, if it determines that an immediate emergency lane change is too risky due to a vehicle behind or to the side, this strategy is temporarily shelved. The module then performs multi-objective optimization calculations on the "comfortable deceleration" and "preparing to change lanes" (waiting for an opportunity) strategies, balancing safety, comfort, and traffic efficiency. Ultimately, it selects and executes a gradual deceleration strategy to gradually increase the safe distance from the truck, buying time to observe and respond to real-world situations in the blind spot. During execution, the safety monitoring system continues to operate. Suddenly, the truck veers to one side to avoid an obstacle, and the real obstacle (a piece of fallen cargo) is instantly exposed to the vehicle's sensors. Because the vehicle has already decelerated in advance, gaining sufficient reaction distance and time, the system immediately triggers the intervention mechanism, smoothly executing full braking and successfully stopping safely in front of the obstacle, avoiding an accident.

[0265] Example 2: Interaction between traffic lights-free urban intersections and illegally crossing vehicles

[0266] Scenario Description: A car is about to pass through an uncontrolled intersection and plans to go straight. The car's sensors detect a sedan (a vehicle approaching from the side) approaching the intersection on its right. According to the rules, this car should slow down and yield. However, the car does not show any obvious signs of slowing down.

[0267] The implementation process of the technical solution of this invention:

[0268] Data Acquisition and Verification: The data acquisition module captures information on multiple targets at the intersection, including vehicles approaching from the side, oncoming vehicles, and pedestrians. The data verification module cleans and verifies the multi-sensor data. For example, it discovered that a camera misidentified a vehicle approaching from the side as an SUV due to backlighting, while LiDAR and millimeter-wave radar more consistently showed its outline as a sedan. Through grouping and outlier removal, the camera's misclassification data was corrected, ensuring the accuracy of target recognition.

[0269] Data Supplementation and Inference: The data supplementation module simultaneously activates two inference modes. On one hand, it analyzes the behavior of lateral vehicles: extracting features such as speed, acceleration, and trajectory curvature, and finding that their anomaly score continuously increases relative to the expected behavior of "slowing down and observing at an intersection without traffic lights." On the other hand, the causal graph model is activated: the evidence "intersection ahead" and "lateral vehicle not slowing down" are injected, and the model calculates based on the causal chain (lateral vehicle acceleration, leading to violation of yield rules, then to a surge in the risk of collision at intersecting paths), concluding that the posterior probability of the potential danger node "right-side collision" rapidly exceeds the alert threshold.

[0270] Data Fusion and Environmental Modeling: The fusion module combines real observation data (position and speed of vehicles to the side) with high-probability virtual threat data derived from causal inference (a collision risk from the right). Due to its high posterior probability and clear causal logic, the virtual threat data is prominently identified in the environmental model, alerting the system to the presence of a high-probability, yet unrealized, conflict event.

[0271] Scene Classification: The scene classification module calculates the hazard factors of all elements. Lateral vehicles are considered extremely high-risk elements due to their abnormal behavior and potential intersecting paths. The module combines the road topology features of the "intersection" with the interaction relationship between "lateral vehicles and the vehicle" (path intersection) to accurately classify the scene as a high-risk interaction scenario of "lateral vehicles illegally passing through an intersection without traffic lights".

[0272] Database retrieval: The database retrieval module successfully retrieved a strategy library for similar scenarios through fuzzy matching, including strategies such as "defensive slowing down and yielding", "honking warning", and "accelerating through when it is safe to do so", along with the trigger conditions and potential risk descriptions for each strategy.

[0273] Real-time vehicle control: The control module performs safety verification and multi-objective optimization on the retrieved strategies. It determines that "accelerating through" carries too high a risk, as it cannot guarantee complete avoidance of the lateral vehicle; "honking as a warning" can be used as an auxiliary but not primary measure; therefore, "defensive deceleration and yielding" becomes the optimal choice. The system generates control commands to cause the vehicle to perform a smooth but clear deceleration action, indicating the intention to yield and reserving safety redundancy. While the vehicle decelerates, the lateral vehicle does not decelerate and directly enters the intersection. Because the vehicle has taken evasive action in advance, the two vehicles safely pass each other at the intersection, avoiding a potential collision. Throughout the process, the system monitors the trajectory of the lateral vehicle in real time, ensuring that the vehicle's deceleration is always within a safe and controllable range.

[0274] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. An intelligent connected vehicle safe driving decision control system, characterized in that: The method comprises the following steps: A data acquisition module is used to jointly collect multi-modal data of the surrounding environment of the vehicle through a laser radar, a camera and a millimeter wave radar; A data verification module is used to verify the collected multi-modal data; A data supplement module is used to supplement and reason the verified data, and the supplement method comprises deriving the environmental state according to the behavior of the surrounding vehicle and constructing a causal diagram and deriving potential dangers according to the historical and current changes of the surrounding environment; A data fusion module is used to fuse the multi-modal data after verification and supplement based on the confidence of each data source, and generate a unified environmental model; A scene classification module is used to calculate the danger factor of each element in the environment based on the fused environmental model, classify the elements with a danger factor lower than a set threshold as obstacles, and judge the current driving scene type according to the elements with a danger factor higher than the set threshold; A database retrieval module is used to retrieve potential danger information and preset operation strategies associated with the scene in a preset scene database according to the judged current driving scene type; A vehicle real-time control module is used to receive the preset operation strategy and generate vehicle control instructions and perform real-time control under the constraint of preset strict safety rules; Wherein, when the data verification module verifies the collected multi-modal data, the verification method comprises: A11: data mutation detection is performed, and the identified unreasonable mutation data is removed and filled; A12: group the data according to the difference between the data, and in each group, remove the data points judged to be the most unreliable in combination with the confidence of each data point and the data distribution in the group; Wherein, when the data verification module performs data mutation detection and removes the identified unreasonable mutation data, it specifically comprises: S11: data mutation detection, for the current time data point from any sensor data stream, calculate the weighted residual error of its predicted value based on the physical kinematic model and the actual observation value, and compare the weighted residual error with the dynamic adaptive threshold, if the weighted residual error is greater than the dynamic adaptive threshold, it is judged as unreasonable mutation data; S12: unreasonable data removal, remove the data points judged as unreasonable mutation from the valid data sequence, and mark them as missing values; S13: missing data filling, for the missing values formed after being removed, the best estimated value based on the historical valid data sequence and the sensor fusion predicted value is used for filling; Wherein, when the data verification module groups the data according to the difference between the data, and in each group, removes the data points judged to be the most unreliable in combination with the confidence of each data point and the data distribution in the group, it specifically comprises: S21: data grouping, according to the spatio-temporal correlation, the observation data of the same physical entity by different sensors is divided into the same data group; S22: group data distribution and confidence evaluation, for each data point in the same data group, calculate its consistency measure value with other data points, and obtain its own source confidence; S23: comprehensive distrust score calculation, for each data point in the same data set, a comprehensive distrust score is calculated, which is a function of its data inconsistency measure and its source confidence; S24: outlier elimination, according to the comprehensive distrust score, the data points in the group are sorted, and the data points with the highest score and exceeding the set threshold are determined as outliers and eliminated; Wherein, when the scene classification module calculates the danger factor of each element in the environment based on the fused environment model, classifies the elements with the danger factor below the set threshold as obstacles, and judges the current driving scene type according to the elements with the danger factor above the set threshold, specifically comprising: S61: environment element danger factor calculation, based on the unified environment model generated by the data fusion module, the comprehensive danger factor of each dynamic element and static element in the model is calculated, and the calculation formula of the comprehensive danger factor is: ； wherein, is the overall risk factor for element i, is the predicted time to collision of element i with the ego vehicle, is the inverse function of the time to collision, with smaller values indicating higher risk, is the relative distance of element i to the ego vehicle, is the negative exponential function of the distance, is the element type, is the base risk weight based on the type, is the behavioral characteristic of the element, is the risk premium based on the behavior, is the uncertainty measure of the element's motion trajectory, is the risk premium due to uncertainty, , , , and are the weight coefficients for each factor, respectively. S62: element classification based on danger factor, compare the danger factor of each element with the preset static obstacle threshold and dynamic risk threshold, and classify the environment elements; S63: scene feature vector construction, based on the classification result, the macro features of the current environment are extracted, and the scene feature vector is constructed; S64: driving scene classification decision, input the scene feature vector into the scene classifier for calculation, and output the specific driving scene type, wherein the scene classifier adopts a multi-layer perception machine model based on attention mechanism, and its decision function is: ； wherein, is the output of the scene classifier, is the scene feature vector, is the attention weight vector, computed by an attention mechanism, denotes element-wise multiplication, , , and are neural network parameters, denotes a ReLU activation function, the function converts the output into a probability distribution, denotes selecting the scene type with the highest probability as the classification result. 2.The intelligent networked vehicle safe driving decision control system of claim 1, wherein: When the data supplement module supplements the verified data by inferring the environment state according to the behavior of the surrounding vehicles, specifically comprising: S31: behavior feature extraction, continuously tracking the target surrounding vehicle and extracting the behavior feature vector, the behavior features include longitudinal acceleration, longitudinal deceleration, lateral displacement, turn signal state and trajectory curvature change rate; S32: behavior abnormality quantification, based on the current scene context, the abnormality score of the target surrounding vehicle behavior feature vector relative to its expected normal behavior feature vector is calculated; S33: potential risk area derivation, according to the spatial position and behavior characteristics of the surrounding vehicles whose behavior abnormality score exceeds the preset threshold, combined with the road topological structure, the potential risk area outside the sensor field of view of the vehicle and its risk type are derived; S34: virtual observation data generation, based on the derived risk area and its risk type, corresponding virtual observation data is generated, and a synthetic confidence based on the behavior abnormality degree and the inference logic confidence is given, and the virtual observation data is supplemented to the environment model. 3.The intelligent networked vehicle safe driving decision control system of claim 2, wherein: When the data supplement module supplements the verified data by inferring the potential danger by constructing a causal diagram and deriving the potential danger according to the history and current changes of the surrounding environment, specifically comprising: S41: causal diagram construction, a causal diagram model containing multiple levels of nodes is constructed; S42: evidence injection and node activation, the verified current environment observation data is injected into the corresponding environment state node in the causal diagram as evidence, and the downstream intermediate event nodes and potential danger nodes are activated according to the causal logic; S43: Risk probability propagation and calculation, the probability of evidence is propagated along the directed edges of the causal network, and the posterior probability of each potential dangerous node being triggered is calculated; S44: Risk decision and virtual data generation, potential dangers with posterior probability exceeding the preset alert threshold are determined as effective risks, and corresponding virtual threat data and its confidence are generated to supplement the environment model. 4.The intelligent networked vehicle safe driving decision control system of claim 3, wherein: When the data fusion module fuses the multi-modal data after inspection and supplementation based on the confidence of each data source to generate a unified environment model, it specifically includes: S51: Confidence normalization and calibration, receiving the output of each data source from the data inspection module and the data supplementation module, including the observation value and its corresponding original confidence, and performing cross-modal normalization on the original confidence to obtain comparable standard confidence; S52: Fusion weight calculation, based on the obtained standard confidence, assigning a fusion weight to each data source and calculating the spatial consistency factor between each data source; S53: Weighted fusion estimation, using the fusion weight to perform weighted fusion on the observation values of each data source to obtain the optimal estimation value of the target state; S54: Fusion result confidence evaluation, based on the consistency of the fusion weight, the distribution of the confidence of each data source, and the fusion residual, calculating the comprehensive confidence of the fused result. 5.The intelligent networked vehicle safe driving decision control system of claim 4, wherein: When the database retrieval module is working, it includes the following steps: S71: Scene type receiving, receiving the current driving scene type identifier output from the scene classification module; S72: Multi-level scene matching, based on the current driving scene type identifier, searching in the preset scene database, including two levels of exact matching and fuzzy matching; S73: Associated information acquisition, acquiring the potential danger information base and the preset operation strategy base associated with the matched scene from the scene database; S74: Strategy priority sorting, performing applicability evaluation on the retrieved preset operation strategies according to the current environment parameters, and sorting them according to priority. 6.The intelligent networked vehicle safe driving decision control system of claim 5, wherein: When the vehicle real-time control module is working, it specifically includes: S81: Strategy safety verification, receiving the preset operation strategy sequence from the database retrieval module, verifying each strategy with the preset strict safety rule set one by one, and screening out a feasible strategy subset that does not violate any safety rules; S82: Multi-objective optimization decision, for each strategy in the feasible strategy subset, calculating its comprehensive utility value based on the multi-objective optimization function; S83: Optimal instruction generation, selecting the strategy with the highest comprehensive utility value as the final execution strategy, and decomposing it into specific vehicle bottom-level control instructions; S84: Real-time safety monitoring and intervention, continuously monitoring the vehicle state and environmental changes during instruction execution, and triggering the safety intervention mechanism immediately once a situation that will violate the safety rules is detected.

Images