A human-machine common cognitive semantic decoding method for mixed traffic of autonomous driving

Abstract

This invention relates to a human-machine universal cognitive semantic decoding method for autonomous driving in mixed traffic. It aims to dynamically adjust the temporal boundary of data sampling based on the urgency of the driving task and the required cognitive load, thereby reducing the cognitive and decision-making differences between the driver and the autonomous driving system, and achieving human-machine cognitive semantic decoding. First, a dynamic time window model is constructed based on task analysis. Second, discrete questionnaire data is encoded into interval type-2 fuzzy sets with uncertain coverage. Then, a hierarchical and cascaded cognitive fuzzy inference architecture is constructed to achieve situation identification and risk assessment. Finally, semantic decoding is achieved through fuzzy inference and type reduction algorithms, transforming quantized cognitive vectors into vehicle control parameters. This method builds a human-machine universal cognitive semantic decoding framework based on interval type-2 fuzzy logic, establishing a universal cognitive interaction mechanism that allows mutual understanding between humans and machines, and achieving autonomous driving decision-making and control with both high interpretability and human-like characteristics.

Description

Technical Field

[0001] This invention belongs to the field of cognitive science and human-computer interaction technology, and specifically relates to a human-computer universal cognitive semantic decoding method for autonomous driving and mixed traffic. Background Technology

[0002] In mixed traffic environments where autonomous driving and human driving coexist, the differences in cognition and decision-making between autonomous driving systems and human drivers are a significant cause of traffic accidents and low traffic efficiency. A driver's perception and decision-making process primarily relies on their sensory system and cognitive abilities, combined with their driving experience and skills, to make judgments and determine appropriate driving behaviors, exhibiting significant semantic ambiguity, uncertainty, and irrationality. In contrast, existing autonomous driving perception technologies process and analyze sensor information through complex algorithms, using precise mathematical models to characterize interactive vehicle behavior and generate driving decisions. This dimensional difference between "qualitative cognition" and "quantitative calculation" constitutes a formidable cognitive gap in human-computer interaction.

[0003] Existing solutions primarily focus on sampling human driving behavior data through fixed windows and constructing human-like algorithms for autonomous driving using human-inspired paradigms such as reinforcement learning, imitation learning, and brain-like learning, enabling autonomous vehicles to learn human driving styles. However, these methods are mostly limited to one-way imitation of human driving behavior, neglecting the interpretability and cognitive feedback of this behavior to human drivers. This makes it difficult for drivers to accurately understand the intentions of autonomous vehicles, increasing uncertainty and risk in interactions. Furthermore, due to the complexity and variability of the driving environment, there is still a lack of generalized methods that can decode control parameters from human cognitive semantics to autonomous driving systems. Summary of the Invention

[0004] The purpose of this invention is to provide a universal human-machine cognitive semantic decoding method for autonomous driving and mixed traffic, addressing the problems of existing technologies and reducing cognitive differences between humans and machines. This method constructs a universal human-machine cognitive semantic decoding framework based on interval type-II fuzzy logic: First, a dynamic time window model based on task load is introduced at the data and task processing layer, constructing an adaptive data sampler to dynamically adjust sampling boundaries and synchronize the spatiotemporal phase relationship of multi-source heterogeneous data; second, the discrete cognitive semantic questionnaire data of drivers is encoded into interval type-II fuzzy sets with uncertain coverage, realizing mathematical modeling of the fuzziness of human subjective semantics; then, a hierarchical cascaded fuzzy inference architecture is constructed; finally, semantic decoding and decision generation are achieved through a type reduction algorithm and a cognitive-control mapping interface.

[0005] The present invention proposes a human-machine universal cognitive semantic decoding method for autonomous driving and mixed traffic, the specific steps of which are as follows: Step S1: Decomposition of driving tasks and definition of cognitive attributes; Divide the interaction process into sequential steps. K Each stage is further divided into sub-task sets according to the driver's cognitive processing logic. And assign cognitive attribute labels to each subtask. Establish a dynamic time window model based on task load. .

[0006] Step S2: Construct a vehicle interaction state dataset; Real-time acquisition of vehicle interaction characteristic data reflecting the relative positional relationship, motion status, and traffic conflict risk of vehicles, and preprocessing of the data.

[0007] Step S3: Construct a driver cognitive semantic dataset; The cognitive semantic data of drivers’ understanding of the scene, risk perception and interaction intention at different stages of the task were collected by questionnaire survey, and the data were preprocessed.

[0008] Step S4: Construct a human-machine universal interval type II fuzzy reasoning system; The cognitive semantic data is encoded into a type II fuzzy set with a Footprint of Uncertainty (FOU) region, a membership function is constructed, and the centroid parameter of the rule consequent is preset.

[0009] Step S5: Construct a cognitive fuzzy rule base; Based on the cognitive attribute labels described in step S1, a hierarchical cognitive fuzzy rule base is constructed by combining expert experience. It includes three sub-rule bases: layer A situation identification, layer B risk assessment, and layer C interaction intent generation, so as to realize the layer-by-layer reasoning from input state to interaction intent.

[0010] Step S6: Fuzzy reasoning enables driver cognitive semantic decoding; Real-time vehicle interaction feature data is input into the interval type II fuzzy inference system, hierarchical inference is performed based on the cognitive fuzzy rule base, and the centroid parameters are combined to perform type reduction and defuzzification calculations, and finally output a quantized cognitive vector.

[0011] Step S7: Cognitive decoding and decision generation of the autonomous driving system; The quantized cognitive vector described in step S6 is used as the input variable of the cognitive decoding model of the autonomous driving system. A fuzzy rule base is constructed, and through fuzzification, rule evaluation and defuzzification calculation, the decision control parameters of the autonomous driving system are finally decoded and generated.

[0012] Step S1 is as follows: Step S11: Based on task analysis, analyze the entire human-computer interaction process in chronological order.P Divided into K A sequential driving task phase ; Step S12, decompose cognitive subtasks. At any stage Within the system, based on the driver's cognitive resource usage, a corresponding set of cognitive subtasks is decomposed. ,in, M For the current stage The total number of cognitive subtasks obtained from internal decomposition; Step S13: Establish the logical connection between subtasks and hierarchical cognitive levels. (For the first...) j Sub-tasks Assigning cognitive attribute labels Cognitive attribute tags Where A represents situational awareness, B represents intent prediction, and C represents intent generation; Step S14: Construct a dynamic time window model based on task load.

[0013] Establish a model based on task load to determine data sampling boundaries: ,in, For the current moment K Phase 1 j The time window length for each subtask These are coefficients used to adjust the dimensions and sampling scale. This represents the average task time under undisturbed standard operating conditions, based on statistics from a natural driving dataset. This is a coefficient used to adjust the contraction rate of the time window. To characterize the task load level of subtask interaction complexity, a 7-point Likert scale commonly used in cognitive ergonomics was adopted, and the score was determined by experts.

[0014] Step S2 is as follows: Step S21: Identify the interactive object. The current time window is determined using the dynamic time window model described in step S1. Within the current time window, identify the set of key interactive objects around the vehicle. Specifically, ,in, This refers to the interactive object in front of the vehicle. This refers to the interactive object on the left side of the vehicle. This represents the interactive object on the right side of the vehicle; Step S22: Obtain motion features and construct feature vectors. Iterate through the set of key interaction objects identified in step S21. For each object The motion characteristic data of the vehicle relative to itself are obtained, and a two-dimensional state vector based on the vehicle's body coordinate system is established. ,in, The longitudinal relative distance between the vehicle and the interacting object. The lateral relative distance between the vehicle and the interacting object. The longitudinal relative speed between the vehicle and the interacting object. The lateral relative speed between the vehicle and the interacting object; Step S23: Adaptive sliding truncation and temporal alignment of the feature sequence. Using the current time... t Based on the baseline, construct length The sliding observation window uses a first-in-first-out principle to retain time intervals. The feature data within the data map different heterogeneous data to a unified time axis. By matching timestamps, the temporal deviation caused by asynchronous sampling is eliminated, and a standardized input sample sequence with spatiotemporal alignment is constructed.

[0015] Step S3 is as follows: Step S31, Cognitive Semantic Questionnaire Design. A cognitive semantic questionnaire is designed based on vehicle interaction characteristics. Optionally, the endpoint method is used to design questions, requiring participants to mark the start and end numerical boundaries on the physical domain axis for each semantic label, as they perceive them to conform to the semantic description. Optionally, the cognitive semantic label set for distance is {"near", "medium", "far"}, and the relative speed label set is {"fast", "medium", "slow"}. Step S32: Interval Data Acquisition and Preprocessing. Multiple drivers are invited to mark the cognitive start and end points of the semantic tags in a preset domain space to obtain raw interval samples. The samples are then subjected to validity checks and outlier removal, retaining only valid interval data. ,in, Indicates the first i The starting point for each driver's understanding of the semantic tags within a predefined domain space. Indicates the first i The cognitive endpoint of a driver's perception of the semantic label in a preset domain space; Step S4 is as follows: Step S41, Individual probability distribution mapping. Optionally, a membership function of shape such as Gaussian can be constructed based on a new interval analysis method. This is based on the effective cognitive semantic intervals obtained in step S32. Assuming that a driver's perception of the target semantic label follows a normal distribution within the specified interval, the discrete interval data is mapped to a distribution based on the individual mean. and individual standard deviation A defined Gaussian membership function of type I: ,in, Indicates the firsti The average of individual drivers Indicates the first i The standard deviation of an individual driver It is the inverse function of the error function. The free parameter representing the confidence level, Represented as the first i A Gaussian membership function determined by each driver; Step S42: Determine the FOU parameters for the uncertain coverage region. Optionally, this involves determining the Gaussian membership function of the individual coefficients constructed in step S41. Perform merging operations such as union to construct interval type II fuzzy sets that reflect differences in group cognition. ,in, m The total number of valid cognitive semantic samples obtained in step S32. The membership function in step S41 The only confirmed first i The first type of fuzzy set of each individual is used to form the FOU (Follicular Unit) that reflects the differences in group cognition. Four key feature parameters describing the geometric boundary of the FOU are extracted: mean boundary. Boundary of Standard Deviation ,in, , ,in, Indicates all the first i The mean of each individual The minimum value in, Indicates all the first i The mean of each individual The maximum value in, Indicates all the first i Standard deviation of each individual The minimum value in, Indicates all the first i Standard deviation of each individual The maximum value in; Optionally, in step S43, construct interval type II membership functions of shapes such as Gaussian. Using the feature parameters of the FOU geometric boundaries, construct the Upper Membership Function (UMF) and Lower Membership Function (LMF) to describe the semantic label uncertainty, where the UMF is determined by the maximum variance. and the mean interval Flat-top Gaussian function with defined envelope: LMF is based on minimum variance. The Gaussian function determined at the mean endpoints on the same side: ; Step S44: Determine the centroid parameters of the rule consequents. For the preset semantic labels of consequents in the fuzzy rule base, determine their corresponding constant centroid intervals based on expert experience. ,in, This represents the minimum possible value that experts consider a given semantic tag to be numerically accurate. This represents the maximum possible value that experts believe a certain consequent semantic label can be numerically. This range represents the consensus coverage of the target decision semantic in the numerical domain and serves as the output benchmark for subsequent quantization calculations and the input domain reference for the cognitive decoding model of the autonomous driving system.

[0016] Step S5 is as follows: Optionally, in step S51, construct the A-layer situation identification rule base. For the subtask with cognitive attribute label A, take the state vector of the key interactive object described in step S22 as input, map it to the semantic label set describing the scene attributes according to the IF-THEN rule set by expert experience, and output the semantic attribution tendency of the current traffic scene through reasoning; Optionally, in step S52, a risk assessment rule base of layer B is constructed. For the subtask with cognitive attribute label B, the scene semantic labels identified in layer A and the state vectors of the key interactive objects are used as input, and the corresponding risk perception semantic evaluation is output through reasoning. Optionally, in step S53, construct a rule base for generating interaction intent at layer C. For subtasks with cognitive attribute label C, use the risk assessment output from layer B as input to output the semantics of the interaction intent.

[0017] Step S6 is as follows: Step S61: Multidimensional feature fuzzification and activation interval strength calculation. The upper and lower membership functions constructed in step S4 are called, and the acquired vehicle motion features are input into the interval type-II fuzzy inference system. Optionally, activation intervals for each inference rule can be calculated using operators such as multiplication operators. ,in, , , n This indicates that this is the [number]. n A fuzzy reasoning rule, This represents a set of vehicle motion characteristic data acquired at the current moment. Representing the n The activation range of the rule, p The feature dimension input to this layer, Indicates the first n The lower limit of the activation strength of the rule. Indicates the first n The upper limit of the activation strength of the rule. Representation of features j The input value in the first nThe mapping value on the membership function of the rule. Representation of features j The input value in the first n The mapping value on the membership function of each rule; Step S62, optionally, involves performing a reduction process using an algorithm such as the Enhanced Iterative Algorithm with Stopping Condition (EIASC). This is combined with the expert cognitive centroid interval preset in step S44. Find the left switching index point of the activated rule weights through mathematical iteration. L Switch index point to the right R Transforming vague cognitive logic into numerical ranges with clear left and right boundaries. ,in, This represents the left boundary value of the numerical range after the reduction. ; This represents the right boundary value of the numerical range after the reduction. ; N The total number of rules that have been activated; Step S63: Cognitive score quantification and vector output. Optionally, methods such as the midpoint method can be used to defuzzify the above numerical range to obtain a single quantified cognitive score. The final output is a quantized value derived from scene understanding. Risk perception quantification value and the quantification value of interaction intent The constructed full-dimensional quantitative cognitive vector .

[0018] Step S7 is as follows: Step S71, construct the cognition-control mapping interface. The quantized cognitive vector output from step S6... As input variables for the cognitive decoding model of the autonomous driving system, optionally, a fuzzification method such as single-point fuzzification is used to map the scores of each dimension in the quantized cognitive vector to the preset decoding model input domain space. Alternatively, an interval type II membership function is constructed using a shape such as a Gaussian type isomorphic to step S4 to calculate the membership degree of each input variable to the antecedent of the decoding rule. Step S72: Construct a multi-objective decision rule base. Optionally, establish four parallel fuzzy inference subsystems, each corresponding to one of the four key control output variables of the autonomous vehicle, to process forward / stop decision rules, movement direction rules, speed change rules, and aggregation and coordination rules. Step S73, system decoding and parameter output. Optionally, the inference results of the above four fuzzy subsystems are defuzzified using methods such as interval type II fuzzy inference and EIASC type reduction algorithm, which are isomorphic to step S6, and finally outputs an autonomous driving control parameter vector consisting of stop / forward decision value, movement direction angle, speed change amount and aggregation coordination degree control parameter.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces a driving task analysis method and constructs a dynamic time window model based on task load, realizing a phased quantitative representation of the cognitive and decision-making process. This method is applicable to all driving scenarios and can dynamically adjust the temporal boundary of data sampling according to the urgency of the driving task and the required cognitive load. It uniformly encodes and hierarchically categorizes driving tasks to separate and independently analyze people, vehicles, and environments in different driving scenarios, thus ensuring the method's broad applicability across scenarios. Compared to fixed window sampling, it can more accurately capture effective features related to the current decision, solve the problem of spatiotemporal phase synchronization of multi-source heterogeneous data, and achieve adaptive feature extraction in complex and ever-changing driving environments.

[0020] 2. The human-machine universal cognitive semantic decoding method proposed in this invention adopts interval type II fuzzy logic, which bridges the gap between "qualitative cognition" and "quantitative calculation" between humans and machines; it utilizes FOU and flat-top Gaussian envelope function to accommodate the differences in understanding of the same semantics among different drivers, and realizes the modeling of the uncertainty of individual cognitive differences and semantics among drivers.

[0021] 3. The human-machine universal cognitive semantic decoding method proposed in this invention adopts a hierarchical fuzzy reasoning architecture, establishing an interpretable reasoning process of "situation identification - risk assessment - intent generation" that conforms to human thinking logic. Through EIASC and the cognitive-control mapping interface, a universal decoding process from cognitive semantics to physical control is constructed, transforming abstract human cognitive semantics into physical control parameters for the autonomous driving system. This not only achieves anthropomorphic decision generation for the autonomous driving system but also ensures the interpretability and predictability of the decision logic for the driver, thereby significantly reducing the uncertainty risk in human-machine interaction. Attached Figure Description

[0022] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a schematic diagram representing the human-machine unified decision-making process based on task analysis in this invention; Figure 3 This is a schematic diagram of the fuzzy reasoning process and system calibration process of the present invention; Figure 4 This is an example of the interval type II Gaussian membership function of the present invention. Detailed Implementation

[0023] To better understand the technical solutions in this specification, the embodiments are described in detail below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments in this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this specification.

[0024] This invention discloses a human-machine universal cognitive semantic decoding method for autonomous driving and mixed traffic, comprising seven steps: (e.g.) Figure 1 ) Step S1, Decomposition of Driving Task and Definition of Cognitive Attributes, as shown in the appendix. Figure 2 As shown. The interaction process P Divided into K Each stage is sequentially connected, and within each stage, subtasks are decomposed according to the driver's cognitive processing logic. And assign cognitive attribute labels to each subtask. Based on the load intensity, the task load (including perception, cognition, and movement) is scored from 1 to 7 (the higher the number, the higher the load), and a dynamic time window model based on task load is established. .

[0025] Step S11: Based on task analysis, analyze the entire human-computer interaction process in chronological order. P (Taking the unprotected left turn process at a signalized intersection in an urban area as an example) Divided into K = 3 sequentially connected driving task phases Phase 1 is the approach to the intersection ( Phase 2 is preparing to turn left. Phase 3 is for executing the turn ( ).

[0026] Step S12, decompose cognitive subtasks. At any stage Within the system, based on the driver's cognitive resource usage in areas such as information processing, risk assessment, and control planning, a corresponding set of cognitive subtasks is decomposed. .like Figure 2 As shown, Five corresponding cognitive subtasks were identified. ; Five corresponding cognitive subtasks were identified. ; It is broken down into four corresponding cognitive subtasks. .

[0027] Step S13: Using classic driving task analysis methods from the field of traffic design and driving behavior analysis, establish the logical connection between sub-tasks and hierarchical cognitive levels, providing a basis for the next step. j Sub-tasks Assigning cognitive attribute labels In this embodiment, A represents situational awareness, B represents intent prediction, and C represents intent generation.

[0028] Step S14: Construct a dynamic time window model based on task load. Based on the currently identified driving sub-tasks, according to... Figure 2 Cognitive load score, dynamically calculated data sampling window In this embodiment, using For example, ,because This is a high-risk search task; the preset baseline average time is [not specified]. Set the scaling factor =1.0, shrinkage rate coefficient =0.2, substituting into the formula, we get This means that under high-load conditions, the system automatically shrinks the observation window from the baseline of 2.5 seconds to 0.62 seconds, capturing only the high-frequency feature data of the current instant, simulating the physiological characteristics of humans under stress, where attention is highly focused and short-term memory is dominant.

[0029] Step S2: Construct a vehicle interaction state dataset. Real-time acquisition of vehicle interaction feature data reflecting relative vehicle positions, motion states, and traffic conflict risks is performed, and the data is preprocessed.

[0030] Step S21: Identify the interaction object. In this embodiment, the dataset is an autonomous driving open road test dataset, defined by the window in step S1. As a temporal boundary, identify the set of key interactive objects around the vehicle. .

[0031] Step S22: Obtain motion features and construct feature vectors. In this embodiment, the vehicle-mounted LiDAR and camera acquire motion features of the vehicle and the interactive object. For example, longitudinal relative distance lateral relative distance longitudinal relative velocity Lateral relative velocity Using motion characteristic data, and with the vehicle body as the origin, a two-dimensional state vector is established based on the vehicle's own coordinate system. .

[0032] Step S23: Adaptive sliding truncation and temporal alignment of the feature sequence. Using the current time... tBased on the baseline, construct length The sliding observation window uses a first-in-first-out principle to retain time intervals. The feature data within the data are mapped to a unified time axis. To address the issue of inconsistent sampling frequencies among multiple sensors, timestamp interpolation is used to eliminate phase deviations and construct a standardized input sample sequence that is spatiotemporally aligned.

[0033] Step S3: Construct a driver cognitive semantic dataset. In this embodiment, cognitive semantic data on drivers' understanding of scenarios, risk perception, and interaction intentions at different task stages are collected based on questionnaire surveys, and the data is preprocessed.

[0034] Step S31, design of cognitive semantic questionnaire. In this embodiment, the endpoint method is used to design the questions. A domain space including relative distance and relative speed is set. The subjects are required to mark the start and end numerical boundaries on the domain axis for each semantic label, which they believe to be consistent with the semantic description. For example, the cognitive semantic label set for distance is {"near", "medium", "far"}, and the label set for relative speed is {"fast", "medium", "slow"}. Step S32, Interval Data Acquisition and Preprocessing. Multiple drivers are invited to label the cognitive start and end points of the semantic tags in a preset domain space to obtain raw interval samples. Abnormal samples are cleaned, and those satisfying logical constraints are retained. Valid interval data .

[0035] Step S4: Construct a human-machine universal interval type II fuzzy reasoning system. In this embodiment, a new interval analysis method is used to encode the cognitive semantic data into an interval type II fuzzy set with FOU, and a Gaussian membership function is constructed, with the centroid parameter of the rule consequent preset.

[0036] Step S41, Individual probability distribution mapping. Invoke the effective cognitive semantic interval described in step S32. Based on the new interval analysis method, in this embodiment, it is assumed that the driver's cognitive perception of the target semantic label follows a normal distribution within the interval, thereby mapping the discrete interval data to a distribution based on the individual mean. and individual standard deviation Determined type I Gaussian membership function ,in, It is the inverse function of the error function. is a free parameter representing the confidence level.

[0037] Step S42: Determine the FOU parameters for the uncertain coverage area. In this embodiment, the individual coefficient type-1 Gaussian membership function constructed in step S41 is used. Perform a union operation to construct an interval-type fuzzy set that reflects differences in group cognition. ,in, m The total number of valid cognitive semantic samples obtained in step S32. The membership function in step S41 The only confirmed first i Individual fuzzy sets are used to form a FOU that reflects the differences in group cognition, and four key feature parameters describing the geometric boundary of this FOU are extracted: mean boundary. Boundary of Standard Deviation ,in, , ; Step S43: Construct the Gaussian membership function. (See attached image) Figure 4 As shown, an upper membership function (UMF) and a lower membership function (LMF) describing the upper boundary of the FOU are constructed. Preferably, in this embodiment, the UMF is constructed as a flat-topped Gaussian function, which is located in the core mean interval. The membership degree within each concept is always 1, in order to accommodate the misunderstandings of different drivers regarding the same concept to the greatest extent possible.

[0038] Step S44: Determine the centroid parameters of the rule consequents. Based on expert experience, preset the centroid parameter range for the rule consequents (such as "high risk" or "avoidance"). This interval represents the consensus coverage of the target decision semantics in the numerical domain, and serves as the output benchmark for subsequent quantization calculations and the input domain reference for the cognitive decoding model of the autonomous driving system.

[0039] Step S5: Construct a hierarchical cognitive fuzzy rule base. In this embodiment, a three-layer reasoning architecture consisting of a situation identification layer, a risk assessment layer, and an interaction intent generation layer is established based on expert experience. By combining and passing down the semantics output from the previous layer with real-time motion features, a multi-dimensional cognitive mapping logic is formed.

[0040] Step S51: Construct the Layer A situation identification rule base. Using the state vectors of the key interactive objects described in Step S22 as input, map them to a semantic tag set describing scene attributes (such as "close following"). Output the semantic attribution tendency of the current traffic scene through fuzzy logic IF-THEN rule reasoning. In this embodiment, an example of Layer A situation identification rules is given based on expert experience: IF For "near", and For "near", and For "slow", THEN state momentum This refers to "close following".

[0041] Step S52: Construct the Layer B risk assessment rule base. Using the scene semantic labels identified in Layer A and the real-time longitudinal relative speed as input, the corresponding risk perception semantic evaluation is output through reasoning. In this embodiment, an example of Layer B situation identification rules is given based on expert experience: IF situation is "close following," and The rating is "fast," and the risk is "high."

[0042] Step S53: Construct a C-layer interaction intent generation rule base. Using the risk assessment output from layer B as input, output the avoidance intensity semantics. In this embodiment, based on expert experience, an example of a C-layer situation identification rule is given: IF Risk is "High", THEN Interaction Intent is "Slow Down and Yield".

[0043] Step S6 involves multi-level fuzzy reasoning to achieve semantic decoding of the driver's cognition. For example... Figure 3 As shown in the fuzzy reasoning process section, traffic environment element data (i.e., real-time interaction features between the vehicle and the interactive object) is acquired as input. Multi-dimensional feature fuzzification processing is performed through a fuzzification interface (step S61), transforming objective physical quantities into fuzzy set membership degrees describing cognitive semantics. The fuzzification interface is implemented by the interval type II membership function pre-constructed in step S4. After fuzzification, the data enters the decision unit executing the reasoning, calling the internal knowledge base for hierarchical reasoning. Specifically, the knowledge base includes the hierarchical cognitive fuzzy rule base constructed in step S5 and the membership function database composed of the Gaussian membership functions constructed in step S4. Subsequently, the reasoning result is processed through a defuzzification interface, using preset centroid parameters, to sequentially complete the "reduction" and "defuzzification" of the logic (steps S62 and S63). After the above calculations, the final output is a driver's quantitative cognitive vector, which contains specific quantitative scores for scene understanding, risk perception, and interaction intent.

[0044] Specifically as follows: Step S61: Multidimensional feature fuzzification and activation interval strength calculation. The upper and lower membership functions constructed in step S4 are called, and the real-time acquired vehicle motion features are input into the interval type II fuzzy inference system. In this embodiment, the activation interval of each inference rule is calculated using a multiplication operator. .

[0045] Step S62, Downsizing Processing. In this embodiment, an enhanced iterative algorithm with termination conditions is used, combined with the expert cognitive centroid interval preset in step S44, to find the left switching index point of each rule activation interval calculated in step S61 through mathematical iteration. L Switch index point to the right R Quickly reduce the type II fuzzy set of the interval to a type I interval. .

[0046] Step S63: Cognitive score quantification and vector output. In this embodiment, the midpoint method is used to defuzzify the above numerical range to obtain a single quantified cognitive score. The final output is a quantized value derived from scene understanding. Risk perception quantification value and the quantification value of interaction intent The constructed quantitative cognitive vector .

[0047] Step S7, cognitive decoding and decision generation of the autonomous driving system. For example... Figure 3 As shown in the system calibration process section, the full-dimensional quantized cognitive vector output in step S6 is used as the input parameter of the cognitive decoding model of the autonomous driving system. First, in the fuzzification stage, the mapping of the input parameter on the preset decoding membership function (as described in subsequent step S71) is calculated through "individual calibration" to obtain the membership degree of the input parameter. Then, in the rule evaluation stage, the system performs multi-objective parallel reasoning based on "presumed rules". In this embodiment, a four-dimensional fuzzy rule library including forward / stop decision, movement direction, speed change and aggregation coordination is constructed for evaluation. Finally, in the defuzzification stage, the evaluation reasoning results determine the membership degree of the output parameters of each dimension through "joint calibration" to complete the "decision output". Finally, the system decodes and generates the exact decision control "output parameters" such as the stop / forward decision value, movement direction angle, speed change and aggregation coordination degree of the autonomous driving vehicle.

[0048] Step S71: Construct the cognition-control mapping interface. In this embodiment, the quantized cognition vector output in step S6 is used... As input variables of the cognitive decoding model of the autonomous driving system, a single-point fuzzification method is used to map the scores of each dimension in the quantized cognitive vector to the preset decoding model input domain space. The membership degree of each input variable to the decoding rule antecedent is calculated by using the interval type II Gaussian membership function construction method which is isomorphic to step S4.

[0049] Step S72: Construct a multi-objective decision rule base. In this embodiment, four parallel fuzzy inference subsystems are established: forward / stop decision rule (F1), movement direction rule (F2), speed change rule (F3), and aggregation and coordination rule (F4), which correspond to the four key control output variables of the autonomous vehicle, respectively.

[0050] Example of a forward / stop decision rule: IF (This indicates a high level of risk perception) and (Representing giving way) THEN (A stop / forward decision value of 1 indicates a stop decision); Example of movement direction rule: IF (This represents the scenario where the driver is following another vehicle) THEN (A 0° movement angle indicates maintaining the center of the current lane with no lateral deviation); Example of speed change rule: IF (This represents the scenario where the understanding state is following the car) and (Representing a medium level of risk perception) THEN (change in velocity) (This indicates a gentle deceleration); Example of a clustering and coordination rule: IF (This represents the scenario where the understanding state is following the car) and (Representing giving way) THEN (A clustering coordination degree of 1.5 indicates that the vehicle actively increases its following distance, demonstrating cooperative yielding behavior.)

[0051] Step S73: Decoding and parameter output of the fuzzy inference subsystem. In this embodiment, the inference results of the above four subsystems are defuzzified using the interval type II fuzzy inference and the enhanced iterative algorithm with termination conditions, which are isomorphic to step S6. Finally, the autonomous driving control parameter vector consisting of stop / forward decision value, movement direction angle, speed change, and aggregation coordination degree is output.

[0052] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Those skilled in the art, after reading this specification, can make various modifications or variations to the technical solutions without departing from the concept of the present invention, and all such modifications or variations fall within the protection scope of the present invention.

Claims

1. A human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic, characterized in that, Includes the following steps: Step S1: Decomposition of driving tasks and definition of cognitive attributes; Divide the interaction process into sequential steps. K Each stage is further divided into sub-task sets according to the driver's cognitive processing logic. And assign cognitive attribute labels to each subtask. Establish a dynamic time window model based on task load. ; Step S2: Construct a vehicle interaction state dataset; Real-time acquisition of vehicle interaction characteristic data reflecting the relative positional relationship, motion status, and traffic conflict risk of vehicles, and preprocessing of the data; Step S3: Construct a driver cognitive semantic dataset; The cognitive semantic data of drivers’ understanding of the scene, risk perception and interaction intention at different stages of the task were collected by questionnaire survey, and the data were preprocessed. Step S4: Construct a human-machine universal interval type II fuzzy reasoning system; The cognitive semantic data is encoded into a type II fuzzy set with a Footprint of Uncertainty (FOU) region, a membership function is constructed, and the centroid parameter of the rule consequent is preset. Step S5: Construct a cognitive fuzzy rule base; Based on the cognitive attribute tags described in step S1, a hierarchical cognitive fuzzy rule base is constructed by combining expert experience. It includes three sub-rule bases: layer A situation identification, layer B risk assessment, and layer C interaction intent generation, so as to realize the step-by-step reasoning from input state to interaction intent. Step S6: Fuzzy reasoning enables driver cognitive semantic decoding; Real-time vehicle interaction feature data is input into the interval type II fuzzy inference system, hierarchical inference is performed based on the cognitive fuzzy rule base, and the centroid parameters are combined to perform type reduction and defuzzification calculations, and finally output a quantized cognitive vector. Step S7: Cognitive decoding and decision generation of the autonomous driving system; The quantized cognitive vector described in step S6 is used as the input variable of the cognitive decoding model of the autonomous driving system. A fuzzy rule base is constructed, and through fuzzification, rule evaluation and defuzzification calculation, the decision control parameters of the autonomous driving system are finally decoded and generated.

2. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 1, characterized in that, Step S1 is as follows: Step S11: Based on task analysis, the entire human-computer interaction process P is divided into K sequentially connected driving task stages according to time sequence. ; Step S12, decompose cognitive subtasks; At any stage Within the system, based on the driver's cognitive resource usage, a corresponding set of cognitive subtasks is decomposed. Where M represents the current stage. The total number of cognitive subtasks obtained from internal decomposition; Step S13: Establish the logical connection between subtasks and hierarchical cognitive levels; for the j-th subtask... Assigning cognitive attribute labels Cognitive attribute tags Where A represents situational awareness, B represents intent prediction, and C represents intent generation; Step S14: Construct a dynamic time window model based on task load; Establish a model based on task load to determine data sampling boundaries: ,in, The time window length for the j-th subtask in the k-th stage at the current moment. These are coefficients used to adjust the dimensions and sampling scale. This represents the average task time under undisturbed standard operating conditions, based on statistics from a natural driving dataset. This is a coefficient used to adjust the contraction rate of the time window. To characterize the task load level of the interaction complexity of subtasks, a 7-point Likert scale from cognitive ergonomics was used, and the score was determined by experts.

3. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 1, characterized in that, Step S2 is as follows: Step S21: Identify the interactive object; determine the current time window using the dynamic time window model described in step S1. Within the current time window, identify the set of key interactive objects around the vehicle. ; Step S22: Obtain motion features and construct feature vectors; traverse the set of key interaction objects identified in step S21. For each object The motion characteristic data of the vehicle relative to itself are obtained, and a two-dimensional state vector based on the vehicle's body coordinate system is established. ,in, The longitudinal relative distance between the vehicle and the interacting object. The lateral relative distance between the vehicle and the interacting object. The longitudinal relative speed between the vehicle and the interacting object. The lateral relative speed between the vehicle and the interacting object; Step S23: Adaptive sliding truncation and temporal alignment of the feature sequence; using the current time t as a reference, construct the length... The sliding observation window uses a first-in-first-out principle to retain time intervals. The feature data within the data map different heterogeneous data to a unified time axis. By matching timestamps, the temporal deviation caused by asynchronous sampling is eliminated, and a standardized input sample sequence with spatiotemporal alignment is constructed.

4. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 1, characterized in that, Step S3 is as follows: Step S31, design a cognitive semantic questionnaire. Design a cognitive semantic questionnaire based on vehicle interaction features. Use the endpoint method to design questions, requiring participants to mark the start and end numerical boundaries on the physical domain axis for each semantic label, which they believe conform to the semantic description. Step S32: Interval data acquisition and preprocessing; invite multiple drivers to mark the cognitive start and end points of the semantic tags in a preset domain space, obtain raw interval samples, and perform validity checks and outlier removal on the samples, retaining valid interval data. ,in, This represents the cognitive starting point of the i-th driver regarding the semantic label in the preset domain space. This represents the cognitive endpoint of the semantic label as defined by the i-th driver in the preset domain space.

5. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 4, characterized in that, Step S4 is as follows: Step S41, Individual probability distribution mapping; based on the effective cognitive semantic interval obtained in step S32 Assuming that a driver's perception of the target semantic label follows a normal distribution within the specified interval, the discrete interval data is mapped to a distribution based on the individual mean. and individual standard deviation A defined Gaussian membership function of type I: ,in, This represents the mean of the i-th driver. This represents the standard deviation of the i-th driver. It is the inverse function of the error function. The free parameter representing the confidence level, It is represented as a Gaussian membership function determined by the i-th driver; Step S42, Determine the FOU parameters of the uncertain coverage region; For the individual coefficient type-1 Gaussian membership function constructed in step S41... Perform a merge operation to construct an interval-type fuzzy set that reflects differences in group cognition. Where m is the total number of valid cognitive semantic samples obtained in step S32. The membership function in step S41 The uniquely determined i-th individual is identified as a type-I fuzzy set, which forms the FOU reflecting the differences in group cognition. Four key feature parameters describing the geometric boundary of the FOU are extracted: mean boundary, mean boundary, and so on. Boundary of Standard Deviation ,in, , ,in, This represents the mean of all individuals of the i-th generation. The minimum value in, This represents the mean of all individuals of the i-th generation. The maximum value in, Represents the standard deviation of all i-th individuals The minimum value in, Represents the standard deviation of all i-th individuals The maximum value in; Step S43: Construct interval type II membership functions; utilize the feature parameters of the FOU geometric boundary to construct the upper membership function UMF and lower membership function LMF to describe the semantic label uncertainty, where UMF is composed of the maximum variance and the mean interval Flat-top Gaussian function with defined envelope: ; LMF is based on minimum variance The Gaussian function determined at the mean endpoints on the same side: ; Step S44: Determine the centroid parameters of the rule consequents; for the preset semantic labels of consequents in the fuzzy rule base, calibrate their corresponding constant centroid intervals based on expert experience. ,in, This represents the minimum possible value that experts consider a given semantic tag to be numerically accurate. This represents the maximum possible value that experts believe a certain consequent semantic label can be numerically. This range represents the consensus coverage of the target decision semantic in the numerical domain and serves as the output benchmark for subsequent quantization calculations and the input domain reference for the cognitive decoding model of the autonomous driving system.

6. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 3, characterized in that, Step S5 is as follows: Step S51: Construct the situation identification rule base of layer A; For the subtask with cognitive attribute label A, take the state vector of the key interactive object in step S2 as input, map it to the semantic label set describing the scene attribute according to the IF-THEN rule set by the expert, and output the semantic attribution tendency of the current traffic scene through reasoning. Step S52: Construct a risk assessment rule base for layer B; For subtasks with cognitive attribute label B, take the scene semantic labels identified by layer A and the state vectors of key interactive objects as input, and output the corresponding risk perception semantic evaluation through reasoning. Step S53: Construct a rule base for generating C-layer interactive intents; For the subtask with cognitive attribute label C, the risk assessment output from layer B is used as input, and the semantics of the interaction intent are output.

7. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 1, characterized in that, Step S6 is as follows: Step S61: Multidimensional feature fuzzification and activation interval strength calculation; The upper and lower membership functions constructed in step S4 are called, and the acquired vehicle motion features are input into the interval type II fuzzy inference system. The activation interval of each inference rule is calculated using the multiplication operator. ,in, , , where n indicates that this is the nth fuzzy inference rule. This represents a set of vehicle motion characteristic data acquired at the current moment. Let p represent the activation interval of the nth rule, and p be the feature dimension of the input layer. This represents the lower bound of the activation strength of the nth rule. This represents the upper limit of the activation strength of the nth rule. This represents the mapping value of the input value of feature j onto the membership function of the nth rule. This represents the mapping value of the input value of feature j onto the membership function of the nth rule; Step S62: Perform the reduction processing using an enhanced iterative algorithm with termination conditions; combined with the expert cognitive centroid interval preset in step S4. By using mathematical iteration to find the left switching index point L and the right switching index point R of the activated rule weights, the fuzzy cognitive logic is transformed into a numerical range with clear left and right boundaries. ,in, This represents the left boundary value of the numerical range after the reduction. ; This represents the right boundary value of the numerical range after the reduction. N represents the total number of activated rules. Step S63, Cognitive Score Quantification and Vector Output: The midpoint method is used to defuzzify the above numerical range to obtain a single quantified cognitive score. The final output is a quantized value derived from scene understanding. Risk perception quantification value and the quantification value of interaction intent The constructed full-dimensional quantitative cognitive vector .

8. The human-machine universal cognitive semantic decoding method for autonomous driving mixed traffic according to claim 1, characterized in that, Step S7 is as follows: Step S71, construct the cognition-control mapping interface; convert the quantized cognition vector output in step S6 into a quantized cognitive vector. As input variables for the cognitive decoding model of the autonomous driving system, a fuzzification method is used to map the scores of each dimension in the quantized cognitive vector to the preset input domain space of the decoding model. Step S72: Construct a multi-objective decision rule base; establish four parallel fuzzy inference subsystems, corresponding to the four key control output variables of the autonomous vehicle, to process forward / stop decision rules, movement direction rules, speed change rules, and aggregation and coordination rules. Step S73: System decoding and parameter output; The inference results of the above four fuzzy inference subsystems are defuzzified, and the final output is an autonomous driving control parameter vector consisting of stop / forward decision value, movement direction angle, speed change and aggregation coordination degree control parameter.

Images