Driving decision interpretation method and device, computer equipment and storage medium

Abstract

The invention relates to the field of automatic driving, and discloses a driving decision interpretation method and device, computer equipment and a storage medium, and the method comprises the steps: obtaining historical scene data collected by a vehicle when it is detected that the vehicle triggers a driving decision; determining at least one historical event influencing the vehicle to trigger a driving decision according to the historical scene data; and generating explanation information of the driving decision based on a causal relationship between each historical event and the driving decision. According to the method, the problem of poor interpretability of the automatic driving system caused by non-transparent decision-making process and difficulty in tracing the long-time sequence causal chain is solved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and more specifically to a method, apparatus, computer device, and storage medium for interpreting driving decisions. Background Technology

[0002] As autonomous driving technology develops, its decision-making process becomes increasingly complex, especially in end-to-end systems where perception, planning, and control are deeply integrated. The decision-making logic can easily be perceived as a black box. This lack of transparency makes it difficult for users and developers to understand why the system makes a certain decision in a specific scenario, severely hindering the credibility and efficient debugging of the technology.

[0003] Existing methods for explaining driving decisions primarily focus on causal analysis of the current instantaneous scene or short-term related variables, lacking the ability to effectively trace and present complex causal chains composed of multiple time-series events over a long period. Therefore, when faced with complex driving decisions driven by both historical environmental changes and vehicle behavior events, existing technologies struggle to generate explanations that clearly demonstrate long-term, multi-step reasoning logic, resulting in insufficient interpretability of the decisions. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method, apparatus, computer device, and storage medium for interpreting driving decisions, in order to solve the problem of poor interpretability of autonomous driving systems caused by opaque decision-making processes and the difficulty in tracing long-term causal chains.

[0005] In a first aspect, embodiments of the present invention provide a method for interpreting driving decisions, the method comprising: When a vehicle triggers a driving decision, the historical scene data collected by the vehicle is acquired. Based on the historical scenario data, at least one historical event that influenced the vehicle's driving decision is identified. Explanatory information for the driving decision is generated based on the causal relationship between each historical event and the driving decision.

[0006] Furthermore, acquiring the historical scene data collected by the vehicle includes: Obtain the trigger time of the driving decision; Historical scene data within a preset time range prior to the triggering time is retrieved, wherein the historical scene data includes data collected from different vehicle components with temporal characteristics.

[0007] Furthermore, determining at least one historical event that influences the vehicle's driving decision based on the historical scene data includes: Extract state change features from the historical scene data, and determine at least one candidate historical event based on the state change features; The candidate historical events are analyzed based on preset event logic rules to obtain supplementary historical events; By combining the candidate historical events with the supplementary historical events, at least one historical event that influences the vehicle's driving decision is obtained.

[0008] Furthermore, the step of generating explanatory information for the driving decision based on the causal relationship between each of the historical events and the driving decision includes: Calculate the attention weight of the driving decision for each of the historical events; At least one key event is selected from the historical events based on the attention weight; Construct a target causal chain from each of the key events to the driving decision, and generate explanatory information for the driving decision based on the target causal chain.

[0009] Furthermore, calculating the attention weight of the driving decision for each of the historical events includes: Obtain the behavioral representation of the driving decision and the state representation of each of the historical events; The state representations of each historical event are correlated with the behavioral representations of the driving decisions to obtain the analysis results. Based on the analysis results, the attention weight corresponding to each of the historical events is determined.

[0010] Furthermore, constructing the target causal chain from each of the key events to the driving decision includes: Obtain the occurrence time of each key event in the set of key events; Based on the time of occurrence, the key events and the driving decisions are concatenated into an initial causal chain; The initial causal chain is deduced to obtain the deduction results, and the potential relationships between the key events are determined based on the deduction results; The initial causal chain is adjusted based on the potential relationship to obtain the target causal chain.

[0011] Furthermore, after generating explanatory information for the driving decision based on the causal relationship between each of the historical events and the driving decision, the method further includes: Construct a counterfactual scenario based on at least one historical event that influenced the vehicle's driving decision; The alternative decision under the counterfactual scenario is deduced, and the difference information between the alternative decision and the driving decision is obtained; Based on the difference information, supplementary explanatory information for the driving decision is generated, and the supplementary explanatory information is fused with the explanatory information to obtain the target explanatory information.

[0012] Secondly, embodiments of the present invention provide a driving decision interpretation device, the device comprising: The acquisition module is used to acquire historical scene data collected by the vehicle when a driving decision is triggered by the vehicle. The determination module is used to identify at least one historical event that affects the vehicle's driving decision based on historical scene data; The generation module is used to generate explanatory information for the driving decision based on the causal relationship between each of the historical events and the driving decision.

[0013] Thirdly, embodiments of the present invention provide a computer device, including: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof.

[0014] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer instructions that cause a computer to perform the method described in the first aspect or any of its corresponding embodiments.

[0015] The method provided in this application has the following beneficial effects: The method provided in this application acquires historical scene data collected by the vehicle when a driving decision is triggered by a vehicle, thus reconstructing the decision-making context. This places isolated decisions within a continuous spatiotemporal context, providing a solid data foundation and ensuring that subsequent analysis is not based on a snapshot but on the dynamic process leading to the decision. By identifying at least one historical event influencing the vehicle's driving decision based on historical scene data, the method transforms and extracts data from low-level data to high-level semantic events. It categorizes heterogeneous scene data into key events with clear semantics, identifies and organizes key driving factors in the decision-making process, and prepares structured and understandable logical components for constructing a meaningful explanatory chain. Based on the causal relationship between various historical events and the driving decision, the method generates explanatory information for the driving decision, elevating it from related events to causal logic. This allows for in-depth analysis of the contribution and causal relationship of each event to the final decision, generating a human-understandable explanatory description. This clearly reveals the reasoning process behind the decision, ultimately transforming incomprehensible machine decisions into causal narratives, improving the explainability and user trust of the autonomous driving system. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the method for interpreting driving decisions according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the attention weight calculation mechanism according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the target causal chain construction process according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating another method for interpreting driving decisions according to an embodiment of the present invention; Figure 5 This is a structural block diagram of a driving decision interpretation device according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] According to embodiments of the present invention, a method, apparatus, computer device, and storage medium for interpreting driving decisions are provided. It should be noted that the steps shown in the flowcharts in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment provides a method for interpreting driving decisions. Figure 1 This is a flowchart of a driving decision reasoning method according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: When a vehicle triggers a driving decision, acquire historical scene data collected by the vehicle.

[0021] In this embodiment, during the operation of the autonomous driving system, the internal state of its decision output module (such as the planning control module) or the control signals of the vehicle's actuators are continuously monitored. When it is determined that the vehicle triggers a driving decision (e.g., an emergency braking, an unexpected lane change, or a path replanning), the driving decision interpretation process is triggered. Detecting driving decisions can be achieved in various ways: for example, by parsing the CAN bus data stream and monitoring the transitions of specific commands such as emergency braking signals (ESP intervention), turn signal activation signals, or automatic emergency braking (AEB) system trigger signals; or by monitoring the action commands output by the planning control module and capturing them when they exceed conventional thresholds (e.g., sudden changes in lateral acceleration or abrupt changes in the curvature of the planned path). When a decision is detected, a data association process is initiated to obtain historical scene data closely related to the current decision in the time dimension. This data consists of raw data or preprocessed feature data from various sensors that have been continuously recorded and cached before the decision trigger moment, collectively constituting the environmental and vehicle state context that led to the current decision.

[0022] Step S102: Based on historical scene data, determine at least one historical event that affects the vehicle's driving decision.

[0023] In this embodiment, a pre-trained multimodal event detection model (such as a video understanding model based on TimeSformer) and a preset set of logical rules are used to perform time-series analysis on historical scene data to identify and extract discrete driving situation segments that influence vehicle-triggered driving decisions, i.e., historical events. These historical events influencing vehicle-triggered driving decisions include at least environmental change events and vehicle behavior events. Environmental change events mainly refer to events triggered by external environmental factors, such as "a vehicle ahead cuts into this lane," "traffic light turns red," or "road construction area detected," identified by video models; and "a new stationary obstacle appears ahead," identified by LiDAR point cloud analysis. Vehicle behavior events mainly refer to actions or state changes performed by the vehicle or other traffic participants, such as "the vehicle activates the turn signal" determined by parsing CAN bus data and applying rules (e.g., "turn signal changes from 0 to 1"), or "the vehicle enters the ramp" determined by fusing GPS and IMU data (e.g., "a continuous steering angular velocity is detected and the trajectory coincides with the ramp area"). By using model inference and rule-based judgment, we can identify historical events with clear semantics that occurred before the decision was made from a continuous stream of historical scenario data.

[0024] Step S103: Generate explanatory information for driving decisions based on the causal relationship between various historical events and driving decisions.

[0025] In this embodiment, the attention weight of driving decisions to each historical event is first calculated to quantify the strength of the causal relationship between them, and key events with weights exceeding a threshold are selected. Then, these key events are concatenated with the final driving decision in chronological order, and a multimodal model with reasoning capabilities is used to analyze the deep logical relationships between the events, constructing a coherent target causal chain. Finally, natural language generation technology (e.g., calling a large language model interface) is applied, using this structured target causal chain as input, to generate natural language text, which serves as the explanatory information. This information elucidates the step-by-step reasoning process from the initial state to the final decision.

[0026] As an example, for emergency braking decisions, the key events selected and their attention weights are: E1 (vehicle cutting in front, weight 0.5) and E2 (vehicle begins to decelerate, weight 0.3). The constructed objective causal chain is "vehicle cutting in front (E1) → vehicle begins to decelerate (E2) → emergency braking (D)". The final generated explanation is: "The vehicle performed emergency braking. The reasons are as follows: First, the system detected a vehicle suddenly cutting into the lane ahead (E1); subsequently, the vehicle initiated initial deceleration (E2) to increase the safe distance; based on the above continuous emergency situations, the emergency braking decision (D) was ultimately triggered."

[0027] In this embodiment of the application, obtaining historical scene data collected by the vehicle includes: Step A1: Obtain the trigger time for the driving decision.

[0028] Specifically, when an autonomous driving system is running, its decision output modules (such as planning or control modules) are monitored in real time to detect driving decisions that require interpretation (such as an abnormal emergency lane change). The trigger moment typically corresponds to the point in time when the decision is actually executed or output, for example, by capturing the timestamp of the decision event from the internal state machine or CAN bus signals. Decision trigger events can be identified by parsing specific signals in the CAN bus data stream (such as a turn signal changing from 0 to 1 indicating lane change intention activation, or a steering wheel angle exceeding a threshold indicating a steering decision), and the precise time of occurrence of this event is recorded as the trigger moment. Furthermore, at the software level, a decision event log is maintained, recording the current time (usually based on a high-precision clock synchronized with sensor data) when the planning algorithm generates commands such as lane change, braking, or acceleration. For example, when a "left lane change" decision is detected, the trigger moment might be the moment the left turn signal is read from the CAN bus, or the moment the lane change command is output from the planning module.

[0029] Step A2: Retrieve historical scene data within a preset time range prior to the triggering time. The historical scene data includes data collected from different vehicle components with temporal characteristics.

[0030] Specifically, during vehicle operation, sensor data from different vehicle components is continuously and cyclically stored in a fixed-length time-series database, with timestamps aligned. The database's duration is defined by a preset time range (e.g., 90 seconds). Upon receiving the trigger moment for a driving decision, the system traces back the preset time range (from 90 seconds before the trigger moment to the trigger moment itself) and retrieves all historical scene data with temporal characteristics from the time-series database within this time period. This data is multimodal, multi-source, and heterogeneous data with clear timestamps. It can include image and video sequences captured by 11 cameras (e.g., 7 front-view pinhole cameras and 4 fisheye cameras) at a specific frame rate (e.g., 20 FPS), point cloud frames generated by LiDAR at a specific scanning frequency (e.g., 10 Hz), time-stamped pose and acceleration data provided by the GPS / IMU, and vehicle status signals such as vehicle speed, turn signal status, and accelerator / brake pedal opening periodically output by the Controller Area Network (CAN) bus. All data is strictly aligned and correlated using timestamps to form a complete, coherent historical scene data sequence that includes contextual information.

[0031] By acquiring the trigger moment, the decision-making time point is accurately captured, providing an accurate time benchmark for subsequent data analysis. By retrieving historical scene data within a preset time range, a complete correlation between the decision and the historical context is established. The data collected from different vehicle components with time-series characteristics ensures the continuity and comprehensiveness of multimodal data, providing a sufficient data foundation for constructing long-range causal chains, thereby effectively supporting the retrospective analysis of complex driving decisions.

[0032] In this embodiment of the application, at least one historical event affecting the vehicle's driving decision is determined based on historical scene data, including: Step B1: Extract state change features from historical scene data, and determine at least one candidate historical event based on the state change features.

[0033] Specifically, pre-trained multimodal event detection models (such as TimeSformer-based video understanding models) and pre-defined signal processing logic are used to perform time-series analysis on retrieved historical scene data to extract state change features, i.e., patterns or indicators that can characterize significant changes in the state of the environment, oneself, or other traffic participants. For example, for image and video data, the model extracts time-series changes in visual features such as the brake lights of the vehicle ahead illuminating, pedestrians entering the lane, and traffic light color changes; for LiDAR point clouds, it analyzes features such as a sharp decrease in the distance to obstacles ahead; for CAN bus data, it monitors abrupt changes in discrete or continuous signals such as the vehicle's speed continuously decreasing below a threshold or the turn signal changing from 0 to 1. The extracted state change features are input into an event classifier, which maps feature combinations that conform to specific patterns to predefined candidate historical events (e.g., "the vehicle ahead brakes suddenly," "the turn signal is activated," "the vehicle ahead slows down," etc.). Essentially, candidate historical events are driving situation segments that occur within a historical time period and may influence the final decision, initially identified by the automated model.

[0034] Step B2: Analyze the candidate historical events based on preset event logic rules to obtain supplementary historical events.

[0035] Specifically, a parallel-running rule inference engine is used. This engine has built-in preset event logic rules, which are conditional judgment statements defined by domain knowledge. These rules are specifically designed to handle scenario changes that may not be fully covered by the multimodal event detection model but are crucial to driving logic. For example, a rule for CAN bus data could be: if the vehicle's speed (read from CAN) is below the threshold of 40 km / h for 3 consecutive seconds and the distance to the vehicle in front (read from radar or CAN) is less than 10 meters, then a slow-moving event for the vehicle in front is triggered. A rule for GPS / IMU data could be: if the GPS trajectory coordinates coincide with the polygon of the ramp area in the high-precision map for 5 consecutive seconds and the IMU detects a continuous steering angular velocity greater than the threshold, then an entry into the highway ramp event is triggered. The candidate historical event list and the original historical scenario data (especially structured data such as CAN, GPS / IMU, etc.) are used as inputs to this rule engine. The engine traverses all preset rules, performs logical judgments on the data, and generates corresponding supplementary historical events when the conditions are met.

[0036] Step B3: Merge candidate historical events and supplementary historical events to obtain at least one historical event that influences the vehicle's driving decision.

[0037] Specifically, the set of candidate historical events identified by the multimodal event detection model is merged with the set of supplementary historical events generated by the rule engine to form a preliminary joint event list. Subsequently, a fusion process is performed, the core of which is to deduplicate and merge items from different sources that point to the same real-world event. A spatiotemporal signature (e.g., based on event type, timestamp of occurrence, and location information) is calculated for each event in the list, and a time window (e.g., ±1 second) is set. If two events have highly similar spatiotemporal signatures and their occurrence times fall within this time window, they are considered duplicate reports of the same event, and one is retained (e.g., the rule-based supplementary historical event is preferentially retained because it usually has higher determinism, or a new fused event description is created).

[0038] By extracting state change features to identify candidate historical events, the system achieves automated conversion from raw data to semantic events, improving the efficiency of event recognition. Supplementary historical events are obtained through analysis based on preset event logic rules, which compensates for potential blind spots in model recognition and enhances the completeness and accuracy of event recognition. Finally, by fusing candidate and supplementary historical events, a set of historical events is formed, providing high-quality event input for subsequent causal relationship analysis.

[0039] In this embodiment of the application, explanatory information for driving decisions is generated based on the causal relationship between various historical events and driving decisions, including: Step C1: Calculate the attention weight of driving decisions for each historical event.

[0040] In this embodiment of the application, the calculation of the attention weight of driving decisions for each historical event includes: Step C101: Obtain the behavioral representation of driving decisions and the state representation of each historical event.

[0041] Specifically, large multimodal models (such as LLaVA) fine-tuned with domain data are used as encoders to generate these representations. For the behavioral representation of driving decisions, the final decision (e.g., "change lanes to the left") is obtained by inputting the final decision along with the contextual multimodal data (such as camera images, LiDAR point clouds, and CAN signals at the moment of the decision) into the multimodal model. Specific network layers within the model (e.g., the output vector corresponding to the [CLS] marker in the Transformer architecture or the final hidden layer state after pooling) output a fixed-dimensional numerical vector that comprehensively represents the semantics and context of the decision. This vector is the behavioral representation and acts as a query vector in the attention mechanism. The state representation of each historical event is obtained by inputting data fragments (such as video frames, point cloud frames, and related vehicle signals at the time of the event's occurrence) corresponding to each historical event (e.g., "the vehicle in front is moving slowly") into the same multimodal model. Using the same method described above, a fixed-dimensional numerical vector is extracted for each event; this vector is the state representation of that event and acts as the key vector in the attention mechanism. These representation vectors map both decisions and events to the same high-dimensional semantic space, laying the foundation for subsequent calculations of their correlations.

[0042] Step C102 involves performing a correlation analysis between the state representations of each historical event and the behavioral representations of driving decisions to obtain the analysis results.

[0043] Specifically, the core similarity calculation function in the attention mechanism is used to obtain driving decision data. Behavioral representation (as query vector) ) and each historical event State representation (as key vector) The comparison is performed. For each historical event, the scaled dot product between its state representation and behavior representation is calculated, i.e., the vector dot product operation is performed. Then divide by the scaling factor (in (This refers to the dimension of the key vector). This calculation process is called association analysis, and the calculation formula is as follows:

[0044] in, For the correlation strength score, To calculate the query vector and key vector A function of similarity between them For driving decisions behavioral representations For historical events The state representation.

[0045] This operation quantitatively measures the degree of directional matching and association strength between the decision and each historical event in the semantic space learned by the model. Ultimately, for all N historical events, it outputs a set of N unnormalized association strength scores. This set of scores is the analysis result, reflecting the relative association strength between each historical event and the driving decision.

[0046] Step C103: Determine the attention weight for each historical event based on the analysis results.

[0047] Specifically, the analysis results (i.e., the set of raw, unnormalized correlation strength scores between all historical events and driving decisions) will be used to analyze the results. As input, the Softmax function is applied for normalization. For each original score in the set... Calculate its exponential function value Then divide that value by the sum of the exponential function values ​​of all the original fractions in the set. (Where j ranges from 1 to n). This calculation process generates a new value between 0 and 1 for each historical event, which is the attention weight. The calculation formula is as follows:

[0048] in, For historical events Attention weights For normalization algorithms, This is the set of association strength scores.

[0049] The sum of the attention weights for all historical events equals 1. The normalization process transforms the original association strength scores into a probability distribution. For example... Figure 2 As shown, each historical event ( For the current decision (as) ) Calculate the corresponding attention weights ( The higher the weight, the stronger the causal relationship of the event in constructing the causal chain, and the higher its importance in subsequent screening of key events and construction of explanations.

[0050] By acquiring behavioral and state representations, decisions and events are unified into the same semantic space, laying the foundation for association analysis. The analysis results obtained through association analysis initially reveal the intrinsic connection between decisions and events. Finally, by determining attention weights based on the analysis results, the association strength is transformed into a standardized probability distribution, which makes the importance of each event clear and quantitative, ensuring the objectivity and interpretability of causal attribution.

[0051] Step C2: Select at least one key event from historical events based on attention weights.

[0052] Specifically, each historical event calculated Corresponding attention weights ( ) and preset attention threshold ( The threshold (e.g., 0.1) is compared. The attention weights of all historical events are iterated through, and only those weights greater than or equal to the threshold (i.e., satisfying the condition) are considered. Historical events are selected and compiled into a key event set. Attention weights quantify the contribution of each event to the decision; events below a threshold are considered too weak to constitute a valid cause for the decision. For example, suppose there are five historical events with attention weights of... When the preset threshold At this point, the first event with a weight of 0.05 will be filtered out, while the remaining four events will be retained to form the set of key events used to construct the causal chain. Ultimately, the events used to explain the decision are the core factors that have been quantified and have significant causal correlation.

[0053] Step C3: Construct the target causal chain from each key event to the driving decision, and generate explanatory information for the driving decision based on the target causal chain.

[0054] By calculating attention weights, the contribution of each historical event to the decision was quantified, providing an objective basis for causal screening. By screening key events based on attention weights, irrelevant or secondary events were effectively filtered out, focusing on core causal factors. Finally, by constructing a target causal chain and generating explanatory information, the quantified causal relationship was transformed into a natural language narrative that humans can understand. This systematically solved the problem of overly simplistic decision explanations and a lack of reasoning processes, significantly improving the transparency of decision-making.

[0055] In this embodiment of the application, a target causal chain from each key event to driving decisions is constructed, including: Step C301: Obtain the occurrence time of each key event in the key event set.

[0056] Specifically, the process involves obtaining a set of critical events after filtering. Each critical event in this set is assigned a timestamp at the time of its identification and creation. The determination of this timestamp depends on the data source from which the event was identified: for events identified from video data by a multimodal event detection model (such as TimeSformer) (e.g., "pedestrian crossing"), the occurrence time is set to the specific time point in the video stream as determined by the model; for events triggered from CAN bus or GPS / IMU data by preset event logic rules (e.g., "activating turn signals" or "entering a highway ramp"), the occurrence time is recorded as the timestamp corresponding to the data point that meets the rule's triggering condition. The critical event set is traversed, and the occurrence time attribute stored within each event object is read one by one, thereby obtaining a time identifier for each critical event in the set, providing the necessary temporal information for the next step of constructing a causal chain in chronological order.

[0057] Step C302: Connect the key events and driving decisions into an initial causal chain according to the time of occurrence.

[0058] Specifically, first, all critical events in the critical event set are sorted in ascending order according to their occurrence time, forming a chronologically ordered event list. Then, the event list is logically connected to the driving decision (which occurs latest) in chronological order: the sorted events are connected sequentially using arrows indicating causal or temporal relationships, and the driving decision is added to the end of the sequence. For example, the sorted event sequence is... If decision D (t=60s), then the initial causal chain formed after the splicing operation is " This initial causal chain is a linear preliminary explanatory structure based on pure temporal relations, which provides a basic structural framework for subsequent steps to analyze the deep causal logic between events.

[0059] Step C303: The initial causal chain is deduced to obtain the deduction results, and the potential relationships between key events are determined based on the deduction results.

[0060] Specifically, the initial causal chain (i.e., the sequence of key events arranged chronologically) is input into a multimodal model (such as a finely tuned LLaVA) with long-process reasoning capabilities for logical deduction. This model analyzes the potential causal, conditional, or motivational relationships between adjacent or non-adjacent events in the chain, going beyond mere temporal sequence. For example, the model might infer whether the "vehicle ahead slowing down" event (E2) constitutes the primary motivation for the "finding and confirming the gap on the left" event (E3), or whether the "entering the ramp" event (E1) provides context for the entire decision sequence. The deduction process relies on the model's knowledge learned from driving causal datasets, and its results identify these deep logical connections between events that transcend mere temporal sequence, thus revealing the causal logical structure inherent in the initial causal chain.

[0061] Step C304: Adjust the initial causal chain according to the potential relationships to obtain the target causal chain.

[0062] Specifically, based on the inference results (i.e., the deep underlying relationships between identified key events), the initial causal chain, arranged only in chronological order, undergoes logical reconstruction and semantic enhancement. This adjustment, while maintaining the temporal order, injects explicit causal semantics into the event chain based on the causal logic inferred by the model. This target causal chain is a coherent causal narrative that clearly expresses the "cause-effect" logical relationship between events, providing a structured and logically rich information foundation for generating the final, easily understandable natural language explanation.

[0063] As an example, such as Figure 3 As shown, the model deduces that "left-side gap (E3)" is the direct cause of "turning on the turn signal (E4)," while "entering the ramp (E1)" is the background condition for the entire sequence. Based on this, semantic annotation or structural fine-tuning of the chain is performed to generate a target causal chain such as "due to the background of [entering the ramp (E1)], because [the vehicle in front is moving slowly (E2)], [left-side gap (E3)] is found, and then [turning on the turn signal (E4)] is executed, which ultimately leads to [lane change decision (D)]".

[0064] By obtaining the time of occurrence, accurate temporal information is provided for constructing the causal chain; by piecing together the initial causal chain, a preliminary explanatory framework based on time sequence is established; by deducing and determining potential relationships, the deep causal logic between events that transcends time sequence is revealed; finally, by adjusting to obtain the target causal chain, the simple sequence of events is elevated into a logically rigorous causal narrative, making the explanation of the decision clear and logically persuasive.

[0065] In the embodiments of this application, such as Figure 4 As shown, after generating explanatory information for driving decisions based on the causal relationship between various historical events and driving decisions, the method further includes: Step S201: Construct a counterfactual scenario based on at least one historical event that affects the vehicle's driving decision.

[0066] In this embodiment, at least one key event with a high attention weight is first selected from the historical event set as the intervention target. Then, a counterfactual scenario is constructed by modifying the original historical multimodal data sequence; that is, the data segments related to the target event are logically altered at the data processing level. For example, if the target is a "slow-moving vehicle in front" event, the relative speed field of the vehicle in front in the CAN bus data is modified to a normal value (e.g., 100 km / h) during the time period of the event, or the visual features of the slow-moving vehicle in front are erased in the corresponding video data using a generative model (e.g., GAN), thereby erasing the occurrence of the event at the data level. Simultaneously, all other time-related historical scene data are kept unchanged to ensure that the only difference between the constructed counterfactual scenario and the real historical scenario is the presence or absence of the target event. This data sequence, formed after artificial intervention and assuming the target event did not occur, is the counterfactual scenario that can be used for subsequent deduction and comparison.

[0067] Step S202: Infer alternative decisions under counterfactual scenarios and obtain information on the differences between alternative decisions and driving decisions.

[0068] In this embodiment, counterfactual scenarios (i.e., modified historical multimodal data sequences) are input into a multimodal model with long-process reasoning capabilities, which operates on the same main inference flow. This model, based on the hypothetical scenario data where the "event did not occur," re-performs decision reasoning, and its output is the alternative decision. For example, in a counterfactual scenario where the "vehicle ahead is slowing down" event is removed, the model infers "maintaining the current lane" as the alternative decision, rather than the actual "change lanes left" decision. Subsequently, the alternative decision is compared with the actual driving decision; the difference information is the quantitative or qualitative description of the difference between the two. In its simplest form, the difference information can be a binary label (e.g., "the decision has changed") or a detailed description (e.g., "the decision changed from 'change lanes left' to 'stay in the lane'"). By comparing, the necessity of the historical event on the decision is clarified, i.e., "if the event had not occurred, the decision would have been different."

[0069] Step S203: Generate supplementary explanatory information for driving decisions based on the difference information, and fuse the supplementary explanatory information with the explanatory information to obtain the target explanatory information.

[0070] In this embodiment, natural language generation technology is first used to transform the discrepancy information (i.e., the comparison between the actual decision and the alternative decision) into easily understandable textual statements; this text is the supplementary explanatory information. For example, if the discrepancy information indicates that the decision changes from "change lanes to the left" to "keep lane" after the "vehicle ahead is slowing down" event is removed, the generated supplementary explanatory information could be: "Note that if the vehicle ahead is not slowing down, the system is unlikely to perform this lane change operation." Subsequently, the supplementary explanatory information is added or inserted and merged with the explanatory information (main explanation) generated based on the target causal chain describing the positive logic of the event into a coherent narrative. During fusion, the supplementary explanatory information is usually connected to the main explanation at the end using transitional or supplementary conjunctions. The final output target explanatory information simultaneously includes the positive causal chain of "how the decision occurred" and the counterfactual argument of "why this decision was necessary," enhancing the depth and persuasiveness of the decision explanation from both positive and negative perspectives.

[0071] By constructing counterfactual scenarios, we can explore the necessary impact of key events on decision-making; by extrapolating alternative decisions and obtaining discrepancy information, we can intuitively demonstrate the key role of specific historical events in the final decision; finally, by generating and integrating supplementary explanatory information, we provide counterfactual arguments to support the main explanation, enhancing the depth, completeness, and credibility of decision explanation from both positive and negative dimensions.

[0072] This embodiment also provides a driving decision interpretation device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0073] This embodiment provides a device for interpreting driving decisions, such as... Figure 5 As shown, it includes: The acquisition module 51 is used to acquire historical scene data collected by the vehicle when a driving decision is triggered by the vehicle. The determination module 52 is used to determine at least one historical event that affects the vehicle's driving decision based on historical scene data; The generation module 53 is used to generate explanatory information for driving decisions based on the causal relationship between various historical events and driving decisions.

[0074] In this embodiment of the application, the acquisition module 51 is specifically used to acquire the trigger time of the driving decision; retrieve historical scene data within a preset time range before the trigger time, wherein the historical scene data includes data collected from different vehicle components with temporal characteristics.

[0075] In this embodiment of the application, the determining module 52 is specifically used to extract state change features from historical scene data, determine at least one candidate historical event based on the state change features, analyze the candidate historical events based on preset event logic rules to obtain supplementary historical events, and fuse the candidate historical events and supplementary historical events to obtain at least one historical event that affects the vehicle's driving decision.

[0076] In this embodiment of the application, the generation module 53 includes: The computation submodule is used to calculate the attention weight of driving decisions for each historical event; The filtering submodule is used to filter out at least one key event from historical events based on attention weight; The construction submodule is used to build the target causal chain from each key event to the driving decision, and to generate explanatory information for the driving decision based on the target causal chain.

[0077] In this embodiment, the calculation submodule is specifically used to obtain the behavioral representation of driving decisions and the state representation of each historical event; to perform correlation analysis between the state representation of each historical event and the behavioral representation of driving decisions to obtain analysis results; and to determine the attention weight corresponding to each historical event based on the analysis results.

[0078] In this embodiment of the application, a submodule is constructed to obtain the occurrence time of each key event in the key event set; to concatenate the key events with driving decisions according to the occurrence time to form an initial causal chain; to deduce the initial causal chain to obtain the deduction result, and to determine the potential relationship between key events based on the deduction result; and to adjust the initial causal chain according to the potential relationship to obtain the target causal chain.

[0079] In this embodiment of the application, the device further includes: a deduction module, configured to construct a counterfactual scenario based on at least one historical event that affects the vehicle's driving decision; deduce alternative decisions under the counterfactual scenario and obtain difference information between the alternative decisions and the driving decisions; generate supplementary explanatory information for the driving decisions based on the difference information, and fuse the supplementary explanatory information with the explanatory information to obtain target explanatory information.

[0080] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 6As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system).

[0081] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0082] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0083] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device as shown by a landing page for an app. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0084] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0085] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0086] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0087] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for interpreting driving decisions, characterized in that, The method includes: When a vehicle triggers a driving decision, the historical scene data collected by the vehicle is acquired. Based on the historical scenario data, at least one historical event that influenced the vehicle's driving decision is identified. Based on the causal relationship between each of the historical events and the driving decision, explanatory information for the driving decision is generated.

2. The method according to claim 1, characterized in that, The acquisition of historical scene data collected by the vehicle includes: Obtain the trigger time of the driving decision; Historical scene data within a preset time range prior to the triggering time is retrieved, wherein the historical scene data includes data collected from different vehicle components with temporal characteristics.

3. The method according to claim 1, characterized in that, The step of determining at least one historical event that influences the vehicle's driving decision based on the historical scene data includes: Extract state change features from the historical scene data, and determine at least one candidate historical event based on the state change features; The candidate historical events are analyzed based on preset event logic rules to obtain supplementary historical events; By combining the candidate historical events with the supplementary historical events, at least one historical event that influences the vehicle's driving decision is obtained.

4. The method according to claim 1, characterized in that, The step of generating explanatory information for the driving decision based on the causal relationship between each of the historical events and the driving decision includes: Calculate the attention weight of the driving decision for each of the historical events; At least one key event is selected from the historical events based on the attention weight; Construct a target causal chain from each of the key events to the driving decision, and generate explanatory information for the driving decision based on the target causal chain.

5. The method according to claim 4, characterized in that, The calculation of the attention weight of the driving decision for each of the historical events includes: Obtain the behavioral representation of the driving decision and the state representation of each of the historical events; The state representations of each historical event are correlated with the behavioral representations of the driving decisions to obtain the analysis results. Based on the analysis results, the attention weight corresponding to each of the historical events is determined.

6. The method according to claim 4, characterized in that, The construction of the target causal chain from each of the key events to the driving decision includes: Obtain the occurrence time of each key event in the set of key events; Based on the time of occurrence, the key events and the driving decisions are concatenated into an initial causal chain; The initial causal chain is deduced to obtain the deduction results, and the potential relationships between the key events are determined based on the deduction results; The initial causal chain is adjusted based on the potential relationship to obtain the target causal chain.

7. The method according to claim 1, characterized in that, After generating explanatory information for the driving decision based on the causal relationship between each of the historical events and the driving decision, the method further includes: Construct a counterfactual scenario based on at least one historical event that influenced the vehicle's driving decision; The alternative decision under the counterfactual scenario is deduced, and the difference information between the alternative decision and the driving decision is obtained; Based on the difference information, supplementary explanatory information for the driving decision is generated, and the supplementary explanatory information is fused with the explanatory information to obtain the target explanatory information.

8. A device for interpreting driving decisions, characterized in that, The device includes: The acquisition module is used to acquire historical scene data collected by the vehicle when a driving decision is triggered by the vehicle. The determination module is used to determine at least one historical event that affects the vehicle's driving decision based on the historical scene data; The generation module is used to generate explanatory information for the driving decision based on the causal relationship between each of the historical events and the driving decision.

9. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1 to 7.

Images