Method and apparatus for providing an explanation for a driving assistance event

Abstract

The application provides a method for providing an explanation for a driving assistance event, comprising: in response to triggering of the driving assistance event, obtaining feature-level perception data associated with the driving assistance event; based on the feature-level perception data, immediately generating a semantic explanation for the driving assistance event, and outputting the semantic explanation in real time in a vehicle; performing anti-tampering processing on at least the feature-level perception data, and associating the anti-tampering processed feature-level perception data with the semantic explanation to form a binding evidence package; and storing the binding evidence package in an encrypted storage area local to the vehicle. The application also provides an apparatus for providing an explanation for a driving assistance event and a computer program product. By capturing feature-level perception data in real time and quickly outputting an explanation for a driver, the driver can understand the reason for the system behavior and improve the acceptance of the system. Through the anti-tampering and evidence storage operation, credible evidence can be provided for subsequent after-sales analysis, responsibility identification or system optimization.

Description

Technical Field

[0001] This application relates to a method for providing interpretation of driving assistance events, and also to an apparatus for providing interpretation of driving assistance events and a computer program product. Background Technology

[0002] When a vehicle's driver assistance system detects a hazard, it will proactively implement interventions such as warnings, steering, or braking to ensure driving safety. However, when these systems suddenly intervene, drivers often cannot immediately understand the specific reasons behind their actions. This can easily cause driver confusion and anxiety, reducing their trust in and willingness to use driver assistance functions. It may also lead drivers to mistakenly believe that the vehicle is malfunctioning, thus increasing unnecessary after-sales consultations and inspection requests. Furthermore, most existing interpretive tools are geared towards technical professionals, and their output consists of technical parameters or raw sensor data, primarily used for post-event data analysis and accident reconstruction, lacking real-time, readable interpretive capabilities for ordinary drivers.

[0003] Therefore, existing solutions for interpreting driver assistance incidents still have significant shortcomings. Summary of the Invention

[0004] The purpose of this application is to provide a method for interpreting driving assistance events, an apparatus for interpreting driving assistance events, and a computer program product, so as to at least solve some of the problems in the prior art.

[0005] According to a first aspect of this application, a method for providing an explanation of driving assistance events is provided, wherein the method includes the following steps: In response to the triggering of a driving assistance event in the vehicle, acquire feature-level perception data associated with the driving assistance event; Based on the feature-level perception data, a semantic interpretation for the driving assistance event is immediately generated and the semantic interpretation is output in real time in the vehicle; At least the feature-level sensing data undergoes tamper-proofing processing, and the tamper-proofed feature-level sensing data is associated with the semantic interpretation to form a bound evidence package; and The binding evidence package is stored in the vehicle's local encrypted storage area.

[0006] This application specifically includes the following technical concept: when a driving assistance event is triggered, the system captures relevant feature-level perception data in real time and can quickly output a driver-friendly, easy-to-understand explanation. Through this user-friendly explanation, the driver can immediately understand the reason for the system's behavior, thereby reducing confusion and enhancing trust and acceptance of the driving assistance function. Furthermore, through tamper-proof processing and the formation of a bound evidence package, verifiable and tamper-proof event records related to each driving assistance event can be created, providing credible evidence for subsequent after-sales analysis, liability determination, or system optimization.

[0007] In one exemplary embodiment, the feature-level perception data includes at least one of the following: a camera image containing a target detection bounding box, a target category label, a range-Doppler slice from a radar sensor, a target-level point cloud from a lidar sensor, clustering results of the lidar point cloud, and / or motion state features extracted from inertial measurement unit data.

[0008] In one exemplary embodiment, the semantic interpretation includes natural language text describing the category of the driving assistance event and its triggering cause; and / or, the semantic interpretation is generated by a language model deployed locally in the vehicle, and the language model outputs a single line of text and / or text within a limited number of characters as the semantic interpretation; and / or, the language model includes a large language model, such as at least one of the Pangu, Doubao, Qwen, and GPT series models.

[0009] In one exemplary embodiment, the method further includes: generating a visualized evidence thumbnail based on the feature-level perception data, and synchronously displaying the evidence thumbnail and the semantic interpretation on the vehicle's human-machine interface; performing at least one of the following processes on the feature-level perception data before storing the feature-level perception data and / or displaying the evidence thumbnail: - Desensitization processing, such as blurring, occluding and / or replacing facial and / or license plate areas; - Compression processing, such as resolution reduction; and / or - Labeling processing, such as visualizing key targets of interest, including pedestrians, traffic signs and / or other vehicles.

[0010] In one exemplary embodiment, the anti-tampering process includes: performing cryptographic hash calculations on the feature-level perceptual data, the semantic interpretation, and their associated metadata to generate corresponding hash values; wherein the binding evidence package includes the feature-level perceptual data, the semantic interpretation, the metadata, and the hash values.

[0011] In one exemplary embodiment, the encrypted storage area is an encrypted ring storage area, which is configured to retain binding evidence package records corresponding to a predetermined number of recently occurring driver assistance events; and / or, access to the encrypted storage area is controlled by the vehicle user's permission permissions.

[0012] In one exemplary embodiment, the semantic interpretation generation and the anti-tampering processing are performed in partitions of different security levels within the in-vehicle computing platform. Exemplarily, the semantic interpretation generation is performed in a first partition, and the anti-tampering processing is performed in a second partition, where the security level of the second partition is higher than that of the first partition. Exemplarily, the first partition includes a quality management partition, and the second partition includes a monitored security partition.

[0013] In one exemplary embodiment, the system monitors whether the time elapsed from the triggering of the driving assistance event to the completion of the semantic interpretation generation exceeds a predetermined delay time; if the predetermined delay time is exceeded, a predefined templated interpretation text is output as a replacement for the semantic interpretation, and a timeout event is recorded.

[0014] According to a second aspect of this application, an apparatus for interpreting driving assistance events is provided, the apparatus including a memory and a processor, the memory storing computer program instructions that, when executed by the processor, enable the processor to perform the method according to a first aspect of this application.

[0015] According to a third aspect of this application, a computer program product is provided, comprising computer program instructions, wherein, when executed by a processor, the computer program instructions enable the processor to perform the method according to a first aspect of this application. Attached Figure Description

[0016] The principles, features, and advantages of this application will be better understood below with reference to the accompanying drawings. The drawings include: Figure 1 A flowchart is shown illustrating a method for providing interpretation of driving assistance events according to an exemplary embodiment of this application; Figure 2 The diagram illustrates the processing procedure in different security level partitions according to an exemplary embodiment of this application, and the final bound evidence package. Figure 3 This illustration shows a semantic explanation and evidence thumbnail displayed on the vehicle's human-machine interface in response to a driver assistance event in an exemplary scenario. Figure 4This illustration shows a semantic explanation and evidence thumbnail displayed on the vehicle's human-machine interface in response to a driver assistance event in another exemplary scenario; and Figure 5 A schematic diagram of an apparatus for providing interpretation of driving assistance events according to an exemplary embodiment of this application is shown. Detailed Implementation

[0017] To make the technical problems to be solved, the technical solutions, and the beneficial technical effects of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and several exemplary embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit the scope of protection of this application.

[0018] Figure 1 A flowchart illustrating a method for providing an interpretation of a driving assistance event according to an exemplary embodiment of this application is shown. The method includes steps S1 to S4.

[0019] In step S1, in response to the triggering of a driving assistance event in the vehicle, feature-level perception data associated with the driving assistance event is acquired.

[0020] In this context, a driver assistance event refers to an intervention control or warning operation actively triggered by a Level 2 or higher level driver assistance system or autonomous driving system. For example, such events include active intervention or warning operations triggered by the following driver assistance functions: automatic emergency braking, forward collision warning, lane keeping assist, emergency lane keeping, automatic emergency avoidance, adaptive cruise control adjustment, lane departure mitigation, and cross-traffic alert.

[0021] In one embodiment, the triggering of a driving assistance event can be identified by monitoring the output signals of the relevant driving assistance function or changes in system state. For example, when the automatic emergency braking function issues a braking command, or the forward collision warning function activates an alarm signal, it is considered that the corresponding driving assistance event has been detected.

[0022] In this context, feature-level perception data refers to lightweight, semantic environmental information derived from raw sensor data processed by perception algorithms. This information contains target-level features that can be directly used for driving decision-making and interpretation. For example, raw sensor data consists of unprocessed signals directly collected by sensors, such as camera pixel streams, radar echo signals, and raw point clouds from lidar. Feature-level perception data, on the other hand, uses algorithms such as target detection to semantically extract information from the raw data, focusing on environmental targets and their attributes.

[0023] In one embodiment, the feature-level sensing data includes at least one of the following: - Camera images containing object detection bounding boxes: After identifying the target in the image through the object detection model, an image containing two-dimensional or three-dimensional bounding boxes is generated, or an image snapshot of the target area is further cropped out; - Target category labels: The detected targets are presented in a list format, including the target identifier (ID), category (such as pedestrian, vehicle, bicycle) and corresponding confidence score; -Radar sensor range-Doppler slice: A two-dimensional spectrum of radar echoes processed jointly in the range-velocity domain, enabling simultaneous resolution and detection of range and radial velocity of multiple targets.

[0024] - Target-level point cloud / clustering results from LiDAR point clouds: After ground filtering and point cloud clustering, only point cloud data associated with key targets of interest are retained; and / or - Motion state features extracted from inertial measurement unit data: For example, high-level semantic features such as rapid acceleration, rapid deceleration, sharp turns, or bumpy road surfaces derived from raw acceleration and angular velocity data.

[0025] In one embodiment, feature-level perception data is acquired in the following manner: when a driving assistance event is detected, within a predetermined time window before and after the event is triggered, feature-level perception data directly related to driving function decisions is acquired from the perception algorithm module of the driving assistance system, which serves as key perception evidence for function intervention.

[0026] In step S2, based on feature-level perception data, a semantic interpretation of the driving assistance event is immediately generated and the semantic interpretation is output in real time in the vehicle.

[0027] In this context, semantic explanation refers to driver-oriented, human-readable prompts that can take the form of text, symbols, and / or graphics. For example, the explanation can be displayed in real time through the vehicle's human-machine interface (such as a head-up display, instrument panel, or center console screen) when a driver assistance event is triggered, and may be accompanied by voice prompts.

[0028] Here, "immediately generating semantic explanations" means generating semantic explanations within a very short time after a driving assistance event is triggered, rather than retrospectively retrieving them from stored event records. This distinguishes this solution from traditional event-triggered data loggers, ensuring real-time generation and output of explanatory information. Drivers can instantly understand the triggering reason after a driving assistance event is triggered, thus improving the user experience.

[0029] In one embodiment, semantic interpretation includes natural language text describing the category of the driving assistance event and its triggering cause. For example, for an automatic emergency braking event, the output might be "Pedestrian crossing ahead, vehicle actively applies emergency braking." For a forward collision warning event, the output might be "Vehicle ahead decelerates suddenly, please be careful." For a lane keeping assist event, the output might be "Vehicle deviates from lane, actively corrected." For an adaptive cruise control adjustment event, the output might be "Vehicle ahead too close, actively decelerated."

[0030] In one embodiment, to ensure real-time performance, the semantic interpretation can be generated by a language model deployed locally on the vehicle. In other cases, it can also be generated using a cloud-based language model. The language model includes large language models, such as at least one of the Pangu, Doubao, Qwen, and GPT series models. This language model is configured, for example, to output single-line text and / or text within a limited word count as the semantic interpretation. Furthermore, the semantic interpretation can support multilingual output, including switching or side-by-side display of at least one local language and alternative languages.

[0031] In one embodiment, the language model can be fine-tuned and trained based on a general large language model using specialized driving scenario data. The input to the language model is, for example, feature-level perceptual data and its associated metadata, as well as the corresponding triggered driving assistance event identifiers; the output is natural language text describing the category of driving assistance event and the triggering reason, thereby being trained to generate interpretable causal descriptions based on perceptual data.

[0032] In one embodiment, a visualized evidence thumbnail can be generated based on feature-level perception data, and the evidence thumbnail and semantic explanation can be displayed synchronously on the vehicle's human-machine interface to help users understand the cause of the event more intuitively and enhance trust.

[0033] The evidence thumbnail can be, for example, a visualization of feature-level perception data. For example, for an automatic emergency braking event, the evidence thumbnail may include a forward-looking camera image containing pedestrian bounding boxes; for a forward collision warning event, the evidence thumbnail may include a radar range-Doppler slice with the target triggering the warning marked; for a lane keeping assist event, the evidence thumbnail may also include an environmental point cloud or image with lane lines and vehicle deviation distance marked.

[0034] In one embodiment, an evidence thumbnail refers to simplified visualizations that do not contain full detail, such as a small area cropped from a full image, to highlight key targets or scenarios that triggered the driver assistance event. This evidence thumbnail is intended to facilitate quick human understanding of the basis for system behavior, rather than for complete scene reconstruction or accident restoration.

[0035] In one embodiment, the evidence thumbnail may also be associated with a confidence indicator. This confidence indicator may be generated, for example, based on detection scores and sensor data consistency calibration. The detection score, for example, represents the internal confidence of a single sensor in its detection results. Sensor consistency refers to the consistency of detection results from multiple sensors for the same scene.

[0036] In one embodiment, at least one of the following processes may be performed on the evidence thumbnail before it is displayed: - Desensitization processing, such as blurring, occluding and / or replacing facial and / or license plate areas; - Compression processing, such as resolution reduction; and / or - Labeling processing, such as visualizing key targets of interest, including pedestrians, traffic signs, and / or other vehicles. In one embodiment, it may also monitor whether the time elapsed from the triggering of a self-driving assistance event to the completion of semantic interpretation generation exceeds a predetermined delay. If the predetermined delay is exceeded, a predefined templated interpretation text is output as a replacement for the semantic interpretation, and a timeout event is recorded.

[0037] For example, the predetermined delay time is, for instance, the end-to-end time from the triggering of a driving assistance event to the completion of semantic interpretation rendering output, such as no more than 250 milliseconds. The templated interpretation text can be a pre-generated fixed-format interpretation for various types of driving assistance events, such as "Automatic emergency braking has been triggered" or "Lane keeping assist has been intervened," etc.

[0038] For example, timing monitoring can be implemented using a watchdog timer: a countdown is started when a driving assistance event is triggered; if the semantic interpretation is generated before the countdown ends, the output is normal; if the timeout occurs, an alternative output process is triggered, templated interpretation text is used as output, and the timeout event is recorded.

[0039] In step S3, at least the feature-level sensing data is subjected to anti-tampering processing, and the anti-tampering feature-level sensing data is associated with semantic interpretation to form a bound evidence package.

[0040] In one embodiment, the tamper-proofing process includes performing a cryptographic hash calculation to generate a corresponding hash value, such as using the SHA-256 algorithm, and optionally combining it with a Merkle chain to enhance data integrity protection. By storing the original data and its hash value together, the data can be subsequently verified as tampered with by recalculating the hash value and comparing it with the stored value. If they match, it proves that the data has not been altered; if they do not match, it indicates that it has been tampered with.

[0041] In one embodiment, the hash value may also be associated with semantic interpretation and evidence abbreviation. Figure 1It is also displayed on the vehicle's human-machine interface for users or service personnel to verify when needed.

[0042] In one embodiment, considering computational overhead and latency constraints, cryptographic hash calculations can be performed only on the feature-level perceptual data itself. In another embodiment, cryptographic hash calculations can also be performed on the feature-level perceptual data, the semantic interpretation, and its associated metadata together. In this case, the binding evidence package contains the feature-level perceptual data, the semantic interpretation, the metadata, and the corresponding hash value. This will be discussed in conjunction with... Figure 2 Further explanation.

[0043] In step S4, the binding evidence package is stored in the encrypted storage area of ​​the vehicle.

[0044] In one embodiment, the encrypted storage area is an encrypted ring-shaped storage area, configured, for example, to retain binding evidence package records corresponding to a predetermined number of most recent driver assistance events, automatically overwriting the oldest event record when the storage space is full. This encrypted ring-shaped storage area may be located in the vehicle's on-board memory, such as a secure storage chip or an encrypted partition. Each binding evidence package may be associated with an event identifier and a timestamp for subsequent retrieval and access by event ID or time range.

[0045] In one embodiment, access to the encrypted storage area is controlled by the vehicle user's permission permissions, and the bound evidence package stored therein can only be accessed after authorization. For example, access control can be implemented through the vehicle's digital key or user authentication mechanism to ensure data privacy and security.

[0046] In one embodiment, similar to the processing of the evidence thumbnail described above in conjunction with step S2, compression and / or desensitization processing may be performed on the feature-level perceptual data before storage, particularly before tamper-proofing. Exemplarily, compression processing may include reducing image resolution or pixel depth, and filtering the radar point cloud to remove non-target point cloud data such as ground and sky. Desensitization processing may include blurring, occluding, or replacing facial and license plate areas appearing in the image to ensure privacy.

[0047] In one embodiment, the semantic interpretation generation and the tamper-proof processing can be performed separately in different security level partitions of the onboard computing platform. This deployment method, for example, conforms to the principle of mixed-criticality automotive computing, isolating tasks of different security levels. This ensures that critical tasks are not interfered with by non-critical tasks, thereby improving the overall reliability and security of the system. This will be discussed in conjunction with the following... Figure 2 Further explanation.

[0048] Figure 2 The diagram illustrates the processing steps in different security level partitions according to an exemplary embodiment of this application, and the resulting bound evidence package.

[0049] exist Figure 2 At the top, feature-level perception data 2, extracted in response to a driving assistance event, is shown. This data serves as the basis for subsequent semantic interpretation generation and evidence preservation processes. Optionally, in addition to feature-level perception data 2 itself, associated metadata 3 is also shown, which may include, for example, the identifier of the source sensor of the feature-level perception data, the data acquisition timestamp, the driving assistance event type, and the target detection confidence score.

[0050] Subsequently, semantic interpretation generation and tamper-proof processing (e.g., cryptographic hash calculation) are performed based on feature-level perceptual data 2 and optionally also based on its metadata 3. Figure 2 As shown, these two types of processing are performed in different security level partitions 201 and 202 of the onboard computing platform.

[0051] Figure 2 The left side is the first partition, 201, with a lower security level. This partition may include, for example, a Quality Management (QM) partition used to run non-safety-critical software components. This partition is typically configured with a certain resource quota to ensure the real-time performance of non-safety-critical tasks such as semantic interpretation generation.

[0052] Figure 2 The right side is a second partition 202 with a higher security level, which may include, for example, a monitored safety partition for running ASIL-level software components. The quality management partition 201 and the safety partition 202 operate without interference, for example, through a memory protection unit (MPU). Optionally, the second partition 202 may also include a hardware security module (HSM), which is a dedicated security hardware built into the vehicle's microcontroller, providing independent computing and storage resources for performing security-critical operations such as encryption, key management, and secure storage.

[0053] The aforementioned partitioning and isolation mechanism can effectively enable uninterrupted operation between tasks of different security levels, ensuring the overall reliability and security of the system.

[0054] The first partition 201 receives feature-level perceptual data as input, and after semantic interpretation generation processing, outputs a semantic interpretation in natural language form 4. The second partition 202 receives feature-level perceptual data and its metadata as input, performs cryptographic hash calculation on it, and outputs the corresponding hash value 5.

[0055] After completing the above processing, the feature-level perceptual data 2, its associated metadata 3, hash value 5, and semantic interpretation 4 are linked and integrated to form a bound evidence package 6. This integration achieves both the joint storage of the hash value and the original evidence data to ensure subsequent data integrity verification, and the association of the semantic interpretation with the evidence data to guarantee the integrity of the evidence chain. This bound evidence package 6 is then transferred to an encrypted storage area for secure storage, for later retrieval and verification.

[0056] It should be noted that Figure 2 In the illustrated embodiment, the generation of semantic interpretation and the tamper-proof processing of evidence data are shown to be executed in parallel. However, in practical applications, the execution order of the two can be flexibly adjusted according to actual needs, such as being processed sequentially in a certain order.

[0057] Figure 3 This illustration shows a semantic explanation and evidence thumbnail displayed on the vehicle's human-machine interface in response to a driver assistance event in an exemplary scenario.

[0058] The human-machine interface 30 is, for example, an electronic instrument panel screen or a vehicle infotainment screen. In the uppermost right area 32 of the interface 30, for example, a semantic explanation of the triggered driving assistance event type "emergency braking triggered" and its cause description "pedestrian 85 meters ahead" is displayed in large, bold font. This semantic explanation is a short statement in natural language, such as presented as a single line of text, ensuring that the driver can quickly and easily obtain key information.

[0059] In the lower right area 33 of interface 30, a more detailed event description and explanation of the cause can also be optionally displayed. For example, the specific driving assistance function involved is "Automatic Emergency Braking (AEB)," the specific intervention operation is "Active Deceleration," and the specific triggering cause is "Radar detected a moving target ahead, and the camera identified a pedestrian crossing the red light." In addition, this embodiment can also optionally display metadata related to the feature-level perception data, such as the confidence level "0.92," the event identifier "AEB-0031," and the hash value "3F2A…9C" corresponding to the feature-level perception data of this event.

[0060] On the left side of interface 30, an evidence thumbnail 31 is also displayed. This thumbnail 31 is, for example, a road environment image captured by the vehicle's forward-facing camera, in which the key target triggering the event, "a pedestrian crossing the road ahead," is marked with a two-dimensional bounding box. To protect privacy, the facial areas appearing in image 31 have been blurred (e.g., pixelated). Through this visual presentation of evidence, the vehicle driver can grasp the basis of the system's decision-making in real time and intuitively, thereby enhancing their understanding and trust in the system's behavior.

[0061] Figure 4 This illustration shows a semantic explanation and evidence thumbnail displayed on the vehicle's human-machine interface in response to a driver assistance event in another exemplary scenario.

[0062] and Figure 3 Unlike other emergency braking scenarios, this exemplary scenario involves an automatic emergency avoidance function. For example... Figure 4 As shown, the top right area 42 of the interface 40 displays the triggered driving assistance event type "Emergency Avoidance Triggered" and its cause description "Vehicle ahead brakes suddenly, swerve to the left to avoid it." This description is a short statement in natural language, allowing the driver to quickly understand why the system is performing an avoidance operation.

[0063] In the lower right area 43 of interface 40, more detailed explanations of the event and its cause can be optionally displayed. For example, the specific driver assistance function involved is "Automatic Emergency Avoidance (AES)," the specific intervention operation is "changing lanes to the left to avoid the obstacle," and the specific triggering reason is "the radar detected that the vehicle in front is decelerating suddenly, the vehicle's braking distance is insufficient, and there is space to avoid the obstacle in the adjacent lane." In addition, metadata related to the feature-level perception data can also be optionally displayed, such as the confidence level "0.94," the event identifier "AES-0027," and the hash value "3F4B…8C."

[0064] On the left side of interface 40, an evidence thumbnail 41 is displayed. This is, for example, an image taken by a forward-facing camera, in which a vehicle braking suddenly ahead is marked with a bounding box 410. To protect privacy, the license plate area in image 41 has been blurred (e.g., mosaicked). Alternatively, in an embodiment not shown, the thumbnail 41 could also be a range-Doppler slice from a radar sensor, presenting the sudden deceleration of the vehicle ahead as a two-dimensional image. Through this visualization, the driver can intuitively confirm that the vehicle ahead did indeed brake suddenly and that the left lane had the conditions to avoid it, thus quickly understanding the basis for the system's decision to perform an emergency avoidance maneuver.

[0065] Figure 5 A schematic diagram of an apparatus for providing interpretation of driving assistance events according to an exemplary embodiment of this application is shown.

[0066] The device 10 includes a processor 11 and a memory 12, wherein the memory 12 may include a computer-readable storage medium such as a hard disk, RAM, or flash memory, and stores computer program instructions. The processor 11 may be a central processing unit (CPU), a microcontroller unit (MCU), a graphics processing unit (GPU), a neural network processing unit (NPU), a digital signal processor (DSP), or other general-purpose or special-purpose processor. When the processor 11 executes the computer program instructions in the memory 12, it is able to perform the method for providing interpretation of driving assistance events as described above.

[0067] It should be noted that the specific process of providing explanations for driving assistance events mentioned in this embodiment is described in a relatively brief manner. The relevant technical details have been described in the foregoing method embodiments, and can be directly referred to in the description of the corresponding parts, and will not be repeated here.

[0068] Although specific embodiments of this application are described in detail herein, they are given for illustrative purposes only and should not be construed as limiting the scope of this application. Various substitutions, modifications, and alterations can be conceived without departing from the spirit and scope of this application.

Claims

1. A method for providing explanations for driver assistance events, wherein, The method includes the following steps: In response to the triggering of a driving assistance event in the vehicle, feature-level perception data associated with the driving assistance event is acquired (2). Based on the feature-level perception data (2), a semantic interpretation (4) for the driving assistance event is immediately generated and the semantic interpretation (4) is output in real time in the vehicle. At least the feature-level sensing data (2) is subjected to anti-tampering processing, and the anti-tampering feature-level sensing data (2) is associated with the semantic interpretation (4) to form a bound evidence package (6); and The binding evidence package (6) is stored in the encrypted storage area of ​​the vehicle.

2. The method according to claim 1, wherein, The feature-level sensing data (2) includes at least one of the following: Includes camera images with target detection bounding boxes, target category labels, range-Doppler slices from radar sensors, target-level point clouds from lidar, clustering results of lidar point clouds, and / or motion state features extracted from inertial measurement unit data.

3. The method according to claim 1 or 2, wherein, The semantic interpretation (4) includes natural language text describing the category of the driving assistance event and its triggering cause; and / or The semantic interpretation (4) is generated by a language model deployed locally on the vehicle, and the language model outputs a single line of text and / or text within a limited number of characters as the semantic interpretation (4).

4. The method according to any one of claims 1 to 3, wherein, The method further includes: Based on the feature-level perception data (2), a visualized evidence thumbnail (31) is generated, and the evidence thumbnail (31) and the semantic interpretation (4) are displayed synchronously on the human-machine interface (30) of the vehicle. Before storing the feature-level perceived data (2) and / or displaying the evidence thumbnail (31), at least one of the following processes is performed on it: - Desensitization processing, such as blurring, occluding and / or replacing the face and / or license plate area; - Compression processing, such as resolution reduction; and / or - Labeling processing, such as visualizing key targets of interest, including pedestrians, traffic signs and / or other vehicles.

5. The method according to any one of claims 1 to 4, wherein, The anti-tampering process includes: Cryptographic hash calculations are performed on the feature-level perceptual data (2), the semantic interpretation (4), and its associated metadata (3) to generate corresponding hash values ​​(5). The binding evidence package (6) includes the feature-level perception data (2), the semantic interpretation (4), the metadata (3), and the hash value (5).

6. The method according to any one of claims 1 to 5, wherein, The encrypted storage area is an encrypted ring-shaped storage area, which is configured to retain the binding evidence package (6) records corresponding to the most recently occurring predetermined number of driving assistance events; and / or Access to the encrypted storage area is controlled by the vehicle user's permission permissions.

7. The method according to any one of claims 1 to 6, wherein, The semantic interpretation (4) generation and the anti-tampering processing are performed in different security level partitions of the vehicle computing platform, respectively; Specifically, the semantic interpretation (4) generation is performed in the first partition (201), and the anti-tampering processing is performed in the second partition (202), wherein the security level of the second partition (202) is higher than that of the first partition (201). Specifically, the first partition (201) includes a quality management partition, and the second partition (202) includes a monitored security partition.

8. The method according to any one of claims 1 to 7, wherein, Monitor whether the time elapsed from the triggering of the driving assistance event to the completion of the semantic interpretation (4) exceeds the predetermined delay time; If the predetermined delay time is exceeded, a predefined templated explanation text is output as a replacement for the semantic explanation (4), and a timeout event is recorded.

9. An apparatus (10) for providing interpretation of driving assistance events, the apparatus comprising a memory (12) and a processor (11), the memory (12) storing computer program instructions which, when executed by the processor (11), enable the processor (11) to perform the method according to any one of claims 1 to 8.

10. A computer program product comprising computer program instructions, wherein, When executed by a processor, the computer program instructions enable the processor to perform the method according to any one of claims 1 to 8.

Images