An automatic driving vehicle data credible storage method

Abstract

The application relates to an automatic driving car data credible storage method, comprising a space-time reference calibration step, which uniformly maps the collection time of a laser radar, a camera and an inertial measurement unit to a global nanosecond time axis, and then executes an online external parameter adaptive optimization algorithm based on a re-projection photometric error to construct a cross-modal feature alignment model to obtain dynamic external parameters accurately describing the relative position relationship among sensors at the current time; a risk quantitative evaluation step, which firstly utilizes the dynamic external parameters output by the out-of-control reference calibration step to construct a self-supervised cross-modal projection consistency loss, calculates the gradient of an input data pair, and statistically generates an adversarial attack index by calculating the local entropy value of the gradient; then, dynamic graph network uncertainty reasoning driven by the external parameters is executed, a space-time graph is constructed, and then the comprehensive adversarial threat index and the semantic consistency score are input into a full connection risk prediction network to output a system comprehensive safety probability. The above method improves the accuracy of storage in the absence of prior samples.

Description

Technical Field

[0001] This application belongs to the field of autonomous driving technology, and specifically relates to a method for trusted storage of autonomous vehicle data. Background Technology

[0002] Existing autonomous vehicle data recording and risk assessment systems typically consist of an onboard data acquisition terminal, a storage unit, and a back-end analysis platform. Their operation usually includes the following steps.

[0003] The data acquisition and synchronization process involves the system connecting to the vehicle's CAN / FlexRay bus and sensing sensors (such as cameras, lidar, and millimeter-wave radar). Typically, the absolute time or network time protocol provided by the global navigation satellite system is used to label the data from each sensor to establish a unified timeline.

[0004] The data storage process employs a first-in, first-out (FIFO) cyclic overwrite mechanism. When storage space is insufficient, the system automatically overwrites the oldest recorded data. Simultaneously, specific trigger thresholds are set (such as airbag deployment or sudden braking deceleration exceeding a set value). When such events are detected, the system locks data within a certain period before and after the trigger, preventing it from being overwritten and saving it as accident data. In the risk assessment step, the backend platform reads the stored driving data and performs replay analysis based on preset traffic rules and kinematic indicators (such as time to collision (TTC) and headway (THW)). By comparing the vehicle trajectory with the positional relationships of lane lines and obstacles, it determines whether there are violations or collision risks. Some systems also combine simulation software to recreate the accident scenario to assist in determining liability.

[0005] However, the above methods still have some shortcomings. First, the spatiotemporal alignment accuracy of multi-source data is insufficient: existing technologies mainly rely on independent clocks for each sensor or low-precision network synchronization, lacking a hardware-level nanosecond-level synchronization mechanism. In high-speed driving or high-dynamic scenarios, time deviations of tens of milliseconds can cause significant spatial misalignment (i.e., "ghosting" or positional mismatch) between LiDAR point clouds and visual images, making it impossible for the stored data to accurately recreate the three-dimensional spatial state at the moment of the accident, affecting the accuracy of liability determination. Second, the storage strategy lacks value-sensitive cyclic overlay and simple threshold triggering mechanisms, making it passive and static; this easily leads to the storage of a large amount of meaningless or representative data while losing key data. In addition, existing autonomous vehicle data recording and risk assessment systems cannot perceive the potential risk level of the current scenario in real time (such as adversarial attacks, complex intersection games, and other "high-entropy" scenarios), resulting in a large amount of low-value smooth driving data occupying limited storage space, while "near-miss" data or algorithm failure precursor data that have not triggered hard thresholds but are of great analytical value are often overwritten and lost. Furthermore, existing autonomous vehicle data recording and risk assessment systems suffer from weak cross-modal risk identification capabilities and vulnerability to attacks. Current risk assessments are largely based on single-modal features or simple post-fusion trajectory analysis, lacking verification of deep semantic consistency between multi-source data (visual, point cloud, bus). This means that when adversarial attacks targeting sensors exist in the physical world (such as infrared jamming of LiDAR, sticker deception of cameras) causing single-sensor perception errors, existing systems often struggle to identify such risks through cross-modal feature conflict detection, leading to missed or incorrect assessments. Summary of the Invention

[0006] The present invention is proposed based on the above-mentioned background of the prior art. The technical problem to be solved by the present invention is to provide a reliable data storage method for autonomous vehicles, so as to solve at least one of the above problems.

[0007] To address the above problems, the technical solution provided in this application includes: A reliable data storage method for autonomous vehicles, characterized by comprising: a spatiotemporal reference calibration step, which maps the acquisition times of LiDAR, camera, and inertial measurement unit onto a global nanosecond-level time axis; then running a LiDAR point cloud motion distortion correction algorithm based on inertial recursion to eliminate the non-rigid deformation of the point cloud caused by the vehicle's own motion during the mechanical rotation scanning process of LiDAR; and then executing an online extrinsic parameter adaptive optimization algorithm based on reprojection photometric error to construct a cross-modal feature alignment model, thereby obtaining dynamic extrinsic parameters that accurately describe the relative positional relationship between sensors at the current moment. The risk quantification assessment process first utilizes the dynamic external parameters output from the runaway baseline calibration step. Construct a self-supervised cross-modal projection consistency loss. Calculate the input data pair gradient The local entropy value of this gradient is calculated to generate the adversarial attack index. Then, perform extrinsic-driven dynamic graph network uncertainty inference to construct a spatiotemporal graph. In the construction of this spatiotemporal graph, edge weights The calculation formula explicitly introduces the dynamic extrinsic parameter. To prevent the propagation of erroneous risks by severing unreliable cross-modal connections; and then to integrate the overall counter-threat index semantic consistency score Input a fully connected risk prediction network and output the overall system security probability. This probability, in the form of feedback parameters, serves as the operating parameter for both the empty reference calibration step and the dynamic storage decision step, forming a two-way control closed loop; the two-way control closed loop includes when Exceeding the threshold At that time, the storage module is forcibly triggered to enter lossless high-frequency recording mode to prevent critical attack scene data from being compressed and lost; when Continuously below the threshold If the physical external parameter is determined to be in failure, the runaway baseline calibration step is executed, and the current characteristic deviation is used as a priori constraint for the next external parameter optimization.

[0008] Preferably, in the method, the dynamic storage decision step includes: running a real-time calculation algorithm for accident risk entropy based on multi-dimensional feature fusion to quantify the comprehensive accident risk entropy value of the current driving scenario. ; ,in These are the normalized weighting coefficients for dynamic instability, environmental threat probability density, and similarity of historical accident characteristics, respectively. Vehicle dynamics instability ,in, For a moment, For vehicle lateral acceleration, For longitudinal acceleration, For yaw rate, The real-time state vector is formed. This is a vector of pre-statistically calculated mean dynamic parameters under normal driving conditions. This is the corresponding covariance matrix; Expressed as the inverse sum of the threat levels of all targets. ,in To prevent extremely small constants with a denominator of zero, For the number of potential risk targets, the first The estimated collision time for each target is , This is the distance attenuation factor. Based on the calculated comprehensive accident risk entropy value... The system employs an adaptive storage space allocation algorithm based on deep reinforcement learning. Through continuous online learning and inference, it dynamically adjusts the data storage strategy according to real-time traffic conditions. When entropy values ​​are low, it automatically executes high compression or discarding strategies to free up space; when high-risk entropy values ​​are detected, it immediately switches to a lossless storage mode. The deep reinforcement learning-based adaptive storage space allocation algorithm utilizes an intelligent agent whose state space... Based on the current storage buffer occupancy rate Current frame risk entropy value And the decision-making actions of the previous moment. Composition; Action Space A set of discrete storage strategies These correspond to lossless high-frequency storage, lossy compressed storage, and discarding / not storing, respectively; the agent includes a reward function that combines value incentives and spatial penalties. ,in For indicator functions, This is the buffer pressure penalty coefficient, used to suppress excessively rapid growth of storage space. The overflow penalty constant is The buffer capacity is set to the maximum; the agent uses a deep Q-network to fit the action value function. And iteratively updated using the Bellman equation: In the formula For learning rate, This is the discount factor.

[0009] Preferably, the method further includes a trusted evidence verification step, which includes: running a lightweight data fingerprint generation algorithm based on Merkle trees to generate digital fingerprints of the original data stream to be stored, and dividing the original data stream to be stored into segments according to a fixed time window. A separate data block, denoted as Using the SHA-256 hash function Perform one-way encryption on each data block to generate the corresponding leaf node hash value. A binary tree structure is used to recursively calculate the hash value of non-leaf nodes from bottom to top. For any nth node... parent node of the layer Its hash value is generated by the next level. Two child nodes of the layer and The hash values ​​are concatenated and then hashed again to obtain the result; let the hash values ​​of the child nodes be respectively and ,and This indicates a string concatenation operation, then the parent node's hash value. ;pass After several recursive iterations, a unique root node hash value is finally generated. Serving as the digital fingerprint of all data within the current time window. Constructing a distributed evidence storage and anchoring protocol based on a consortium blockchain, this protocol includes blockchain transaction transactions for evidence storage metadata. The transaction The data structure consists of a root hash value Trusted timestamps generated by the Global Navigation Satellite System Vehicle Unique Identification Code and digital signatures for vehicle terminals Composition. When accident liability is determined or the authenticity of data is questioned, a data integrity verification process based on Merkel's proof is initiated, specifying the steps that need to be verified. Data blocks Whether it has been tampered with, only the set of sibling node hashes required on the path from the leaf node to the root node is extracted to form the Merkel proof path. ,in The height of the tree; verification function in accordance with and Recalculate the root hash value If the calculated result is This proves that the data block If the data is authentic and valid, it has not been tampered with; otherwise, it is determined that the data is corrupted or falsified.

[0010] Preferably, the raw data is stored locally or in an encrypted cloud, with only lightweight transaction data stored. The transaction is broadcast to a consortium blockchain network comprised of nodes from regulatory agencies, insurance companies, and automakers; after the smart contract verifies the terminal signature, it packages the lightweight transaction into a block and returns the transaction hash. As evidence for preservation.

[0011] Preferably, the inertial recursive laser point cloud motion distortion correction algorithm includes: assuming a frame of point cloud data from the lidar... Include There are 1 laser point, denoted as _____. , of which laser points The actual physical acquisition time is The inertial recursive laser point cloud motion distortion correction algorithm selects the starting time of the point cloud in that frame. The angular velocity acquired using a high-frequency inertial measurement unit serves as the reference moment. and linear acceleration Derivation of time using Lie groups and Lie algebraic integrals Relative to the reference time Vehicle relative pose transformation matrix ; to the original distortion points After uniformly correcting to the reference time coordinate system, the distortion-free point coordinates are obtained. ,in This is the installation extrinsic parameter matrix of the lidar coordinate system relative to the vehicle coordinate system.

[0012] Preferably, for time arrive The tiny time interval, vehicle rotation matrix With translation vector The recursive formula is as follows: in, For the instantaneous speed of the vehicle, The gravity vector and These are the zero bias errors of the gyroscope and accelerometer, respectively. For Lie algebra To Liqun Exponential mapping; through the exponential mapping from arrive Integrate the measurements from all inertial measurement units within the time period to obtain the time interval. relative rotation matrix and relative translation vector Then construct the homogeneous transformation matrix. .

[0013] Preferably, the spatiotemporal reference calibration step executes an online extrinsic adaptive optimization algorithm based on reprojection photometric error, and constructs a cross-modal feature alignment model using the distortion-corrected laser point cloud and synchronously acquired image data, including: obtaining pixel coordinates through the extrinsic parameter matrix to be optimized. ,in, Let be the extrinsic parameter matrix to be optimized, which contains rotational components. Translation components The algorithm first divides the three-dimensional laser points Through camera intrinsic parameter matrix and current external references Projecting onto the image plane yields pixel coordinates. :in The depth value of this point in the camera coordinate system is given; the image grayscale gradient or edge feature intensity at the projection point is calculated, and a loss function is constructed to minimize the photometric reprojection error. ,in, It is the Huber kernel function. The number of valid feature points, The image feature values ​​at the projection points are used. The nonlinear least squares problem is iteratively solved using the Gauss-Newton method or the Levenberg-Marquardt algorithm, calculating the Jacobian matrix and updating the state variables until the loss function converges, thus obtaining the dynamic extrinsic parameters. .

[0014] Preferably, the spatiotemporal calibration step includes: constructing a vehicle-wide global time synchronization network based on the IEEE 1588 Precision Time Protocol, sending synchronization messages and follow messages through the master clock node, and calculating the clock deviation of the slave clock nodes. Path delay Hardware timestamps are used to uniformly map the acquisition times of LiDAR, cameras, and inertial measurement units to a global nanosecond-level time axis. This provides a unified time reference for subsequent data processing.

[0015] Preferably, the operation of gradient-based truth-free adversarial perturbation detection includes: assuming the lidar point cloud features aligned by the spatiotemporal reference calibration module are as follows: The visual features projected onto the image plane are Construct a cross-modal consistency loss function ,in For dynamic extrinsic parameters Feature projection operator, The input is the original multimodal data; during the inference phase, the input data is computed through backpropagation. Relative to consistency loss gradient mapping tensor ; through calculation The local information entropy density and Laplace variance are used to quantify the counter-attack threat index. .

[0016] Preferably, the set of edges in the spacetime graph It depends on the real-time state output by the spatiotemporal reference calibration module; for spatially associated edges Its connection weight By node and Euclidean distance in a unified spacetime coordinate system Decide; ,in The confidence factor of the extrinsic parameters is calculated based on the residuals of the dynamic extrinsic parameter optimization in Module 1. For the adaptive mapping function; and construct the spatiotemporal graph to run multi-layer graph attention convolution operation, for the th Layer node features A multi-head attention mechanism is introduced to calculate aggregated features: in For nodes The neighborhood, The attention coefficients are used. Through graph convolution aggregation, the algorithm outputs a cross-modal semantic consistency score for each perceived object. . Detailed Implementation

[0017] Preferred embodiments of the present invention are described in detail below.

[0018] This invention proposes a trusted data storage method for autonomous vehicles, which relies on an in-vehicle intelligent terminal system for execution. The method includes the following steps: Spatiotemporal reference calibration steps This system achieves millisecond-level spatiotemporal alignment and error correction of multi-source sensor data, including LiDAR and cameras, based on a multi-level clock synchronization architecture and dynamic coordinate system transformation model. It primarily addresses the physical spatial alignment issues caused by asynchronous data acquisition times and vehicle movement in autonomous vehicles operating at high speeds or in complex dynamic scenarios.

[0019] The spatiotemporal reference calibration step, based on a multi-level clock synchronization architecture and a dynamic coordinate system transformation model, achieves millisecond-level spatiotemporal alignment and error correction of multi-source sensor data such as LiDAR and cameras. The dynamic storage decision module calculates accident risk entropy values ​​and uses deep reinforcement learning algorithms to dynamically allocate storage weights, prioritizing the retention of high-value data. The trusted evidence verification module performs Merkle tree hashing and blockchain uploading on the selected key data, ensuring the integrity of the data throughout its lifecycle. The risk quantification assessment module, based on a cross-modal feature fusion network, identifies physical adversarial attacks and quantifies the current state of autonomous driving. Specifically, this step includes: First, a vehicle-wide global time synchronization network is constructed based on the IEEE 1588 precise time protocol. Synchronization and follow-up messages are sent by the master clock node, and the clock deviation of the slave clock nodes is calculated. Path delay Hardware timestamp technology is used to uniformly map the acquisition times of LiDAR, cameras, and inertial measurement units to a global nanosecond-level time axis. This provides a unified time reference for subsequent data processing.

[0020] Then, based on the completion of time synchronization, this step runs a laser point cloud motion distortion correction algorithm based on inertial recursion to eliminate the non-rigid deformation of the point cloud caused by the vehicle's own motion during the mechanical rotation scanning process of the lidar.

[0021] Specifically, the inertial recursive laser point cloud motion distortion correction algorithm includes: Suppose a frame of point cloud data from a lidar Include There are 1 laser point, denoted as _____. , of which laser points The actual physical acquisition time is .

[0022] Because the vehicle is in motion, points collected at different times are in different vehicle coordinate systems. In this specific embodiment, the inertial recursive laser point cloud motion distortion correction algorithm selects the starting time of the point cloud frame. The angular velocity acquired using a high-frequency inertial measurement unit serves as the reference moment. and linear acceleration Derivation of time using Lie groups and Lie algebraic integrals Relative to the reference time Vehicle relative pose transformation matrix .

[0023] For time arrive The tiny time interval, vehicle rotation matrix With translation vector The recursive formula is as follows: in, For the instantaneous speed of the vehicle, The gravity vector and These are the zero bias errors of the gyroscope and accelerometer, respectively. For Lie algebra To Liqun The exponential mapping. Through the... arrive Integrate all inertial measurement unit (IMU) measurements within the time period to obtain the time interval. relative rotation matrix and relative translation vector Then construct the homogeneous transformation matrix. Finally, the original distortion points After uniformly correcting to the reference time coordinate system, the distortion-free point coordinates are obtained. : in This is the mounting extrinsic parameter matrix of the lidar coordinate system relative to the vehicle coordinate system. Through this inertial recursive lidar point cloud motion distortion correction algorithm, discrete lidar points on a continuous motion trajectory can be restored to the spatial geometry of the same instant.

[0024] Furthermore, to address the issue of minute drift in multi-sensor extrinsics under vehicle vibration, an online adaptive extrinsic optimization algorithm based on reprojection photometric error is executed in this step. This algorithm utilizes the distortion-corrected laser point cloud and synchronously acquired image data to construct a cross-modal feature alignment model.

[0025] Specifically, it includes: S121 obtains the pixel coordinates using the extrinsic parameter matrix to be optimized. .in, Let be the extrinsic parameter matrix to be optimized, which contains rotational components. Translation components The algorithm first identifies the three-dimensional laser points. Through camera intrinsic parameter matrix and current external references Projecting onto the image plane yields pixel coordinates. :in This is the depth value of the point in the camera coordinate system.

[0026] S122 calculates the image grayscale gradient or edge feature intensity at the projection point and constructs a loss function that minimizes the photometric reprojection error. .in, It is the Huber kernel function. Introducing the Huber kernel function can enhance the algorithm's robustness to noise and mismatches. The number of valid feature points, Let be the image feature value at the projection point. The above nonlinear least squares problem is iteratively solved using the Gauss-Newton method or the Levenberg-Marquardt algorithm, calculating the Jacobian matrix and updating the state variables until the loss function converges, thus obtaining dynamic extrinsic parameters that accurately describe the relative positional relationship between the sensors at the current moment. This process ensures that the LiDAR point cloud recorded by the storage system maintains strict pixel-level alignment with the visual image in space, providing a high-fidelity data foundation for subsequent accident reproduction and risk assessment.

[0027] Risk Quantification Assessment Steps Based on a high-precision spatiotemporal reference and real-time output extrinsic reliability factor, this system enables real-time identification of physical world adversarial attacks and sensor system malfunctions in open road environments lacking prior truth values.

[0028] Its particular strength lies in solving the real-time identification problem of physical world adversarial attacks and perception system malfunctions in open road environments where there is a lack of prior truth values ​​for autonomous driving systems.

[0029] Specifically, it includes the following sub-steps: S21 performs gradient-based, ground-nothing adversarial perturbation detection. It utilizes the dynamic extrinsic parameters output from the runaway baseline calibration step. Construct a self-supervised cross-modal projection consistency loss. Calculate the input data pair gradient The local entropy value is calculated to generate an adversarial attack index. .

[0030] To counter physical attacks such as illumination projection and adversarial patching, this step no longer relies on the classification confidence of a single sensor. Instead, it utilizes the geometric constraints of multiple sensors after calibration to construct a self-supervised gradient field. Let the features of the lidar point cloud aligned by the spatiotemporal reference calibration module be... The visual features projected onto the image plane are Construct a cross-modal consistency loss function : in For dynamic extrinsic parameters Feature projection operator, This is the current input multimodal raw data.

[0031] During the inference phase, the input data is computed through backpropagation. Relative to consistency loss gradient mapping tensor : This gradient mapping tensor represents "the minimum perturbation direction required to disrupt intermodal consistency." Under normal observation conditions, gradient energy is mainly concentrated at the high-frequency edges of the object; however, under adversarial attacks, in order to mislead the model while remaining imperceptible to the human eye, the attacker must create drastic distortions on a high-dimensional manifold in the feature space, leading to... An unusual high-frequency oscillating response was observed in non-semantic regions (such as road surfaces, sky, or patch areas).

[0032] Through calculation The local information entropy density and Laplace variance are used to quantify the counter-attack threat index. This enables zero-sample detection of unknown physical attacks without the need for truth labels.

[0033] S22 performs extrinsic-driven dynamic graph network uncertainty reasoning and constructs a spatiotemporal graph. In the construction of this spatiotemporal graph, edge weights The calculation formula explicitly incorporates the dynamic extrinsic parameters output in step one. This ensures that when the reliability of the calibration in step one decreases, the graph network can automatically disconnect unreliable cross-modal connections, preventing the risk of error propagation.

[0034] To accurately quantify the system's perception reliability under extreme weather conditions or partial sensor failure, this algorithm constructs a spatiotemporal graph with heterogeneous features as nodes. Unlike traditional static graph construction, the edge set in this algorithm... Strictly dependent on the real-time state output by the spatiotemporal reference calibration module: for spatially associated edges Its connection weight By node and Euclidean distance in a unified spacetime coordinate system Decide: in The confidence factor of the extrinsic parameters is calculated based on the residuals of the dynamic extrinsic parameter optimization in Module 1. This is an adaptive mapping function. This formula ensures that when the sensor undergoes physical displacement... During descent, the graph structure automatically eliminates noisy connections caused by calibration drift. Based on this dynamic graph structure, the algorithm performs multi-layer graph attention convolution operations. For the... Layer node features A multi-head attention mechanism is introduced to calculate aggregated features: in For nodes The neighborhood, The attention coefficients are used. Through graph convolution aggregation, the algorithm outputs a cross-modal semantic consistency score for each perceived object. .like A reading below the safety threshold indicates a logical break between visual semantics and point cloud geometry in the spatiotemporal graph, suggesting an extremely high risk of perception failure.

[0035] S23 outputs the overall risk probability.

[0036] Comprehensive Counter-Threat Index semantic consistency score Input a fully connected risk prediction network and output the overall system security probability. The Fully Connected Risk Prediction Network is a prediction model built on a Fully Connected Neural Network (FCNN). It can adopt relevant network structures from existing technologies, which will not be elaborated here.

[0037] This probability, as the core feedback signal, is directly input to the preceding dynamic storage decision module and spatiotemporal reference calibration module, forming a two-way control closed loop: when Exceeding the threshold At that time, the storage module is forcibly triggered to enter lossless high-frequency recording mode to prevent critical attack scene data from being compressed and lost; when Continuously below the threshold If the physical extrinsic parameter is determined to be in failure, then step one is executed, and the current characteristic deviation is used as a priori constraint for the next extrinsic parameter optimization.

[0038] More preferably, this embodiment also includes a dynamic storage decision-making step.

[0039] The purpose of the dynamic decision-making steps is to address the technical challenge of high-value accident data in autonomous driving being easily overwritten or discarded under the constraint of limited on-board edge computing resources.

[0040] Specifically, the dynamic storage decision-making steps include: S31 uses a real-time calculation algorithm for accident risk entropy based on multi-dimensional feature fusion to quantify the data retention value of the current driving scenario.

[0041] First, define the time. Vehicle dynamics instability Its input is the vehicle's lateral acceleration. Longitudinal acceleration and yaw rate The constructed real-time state vector To accurately measure the degree to which the vehicle's current state deviates from the normal driving envelope, the algorithm uses the Maharanobis distance as a metric: in, This is a vector of pre-statistically calculated mean dynamic parameters under normal driving conditions. This is the corresponding covariance matrix. Simultaneously, the algorithm calculates the environmental threat probability density. This metric is based on the collision times of surrounding traffic participants (such as vehicles and pedestrians) as output by the perception system. It assumes that there are collisions within the current field of view. The first potential risk target, the The estimated collision time for each target is Introducing a distance decay factor ,but Expressed as the inverse sum of the threat levels of all targets: in To prevent extremely small constants with a denominator of zero.

[0042] In addition, the algorithm also performs cosine similarity matching between the feature vector of the current scene and a pre-set historical accident feature database to obtain the feature similarity of historical accidents. Finally, the time is calculated by weighted summation. Comprehensive accident risk entropy value : In the formula These are the normalized weighting coefficients for dynamic instability, environmental threat probability density, and similarity of historical accident characteristics, respectively.

[0043] S32 is based on the calculated real-time risk entropy value. Then, an adaptive storage space allocation algorithm based on deep reinforcement learning is further run.

[0044] The algorithm constructs an agent with a state space. Based on the current storage buffer occupancy rate Current frame risk entropy value And the decision-making actions of the previous moment. Composition. Action space Defined as a discrete set of storage strategies These correspond to lossless high-frequency storage, lossy compressed storage, and discarding / not storing, respectively. To guide the agent to maximize the retention of high-value data within limited storage space, a reward function incorporating both value incentives and space penalties was designed. : in For indicator functions, This is the buffer pressure penalty coefficient, used to suppress excessively rapid growth of storage space. The overflow penalty constant is This is the maximum capacity of the buffer.

[0045] The agent uses a deep Q-network (DQN) to fit the action value function. And iteratively updated using the Bellman equation: In the formula For learning rate, This is the discount factor.

[0046] Through continuous online learning and reasoning, S33 dynamically adjusts the data storage strategy based on real-time road conditions. When the vehicle is driving smoothly (low entropy value), it automatically executes a high compression or discard strategy to free up space. When a high-risk entropy value is detected (such as a sharp lane change or close-range game), it immediately switches to lossless storage mode, thereby achieving full and high-fidelity retention of key accident data under the condition of limited hardware resources.

[0047] More preferably, in this embodiment, the method further includes a trusted evidence verification step. The purpose of this step is to ensure that high-value incident data, after dynamic filtering and compression, possesses immutability and legally binding integrity throughout the entire process of storage, transmission, and post-incident retrieval. To this end, this step includes... S41 first runs a lightweight data fingerprint generation algorithm based on Merkle trees to generate digital fingerprints of the raw data stream to be stored.

[0048] To address the limited computing resources in the in-vehicle environment, this algorithm divides the raw data stream to be stored (including LiDAR point cloud frames, visual image frames, and CAN bus data packets) into fixed time windows. A separate data block, denoted as Utilizing the highly collision-resistant SHA-256 hash function. Perform one-way encryption on each data block to generate the corresponding leaf node hash value. : Subsequently, the algorithm uses a binary tree structure to recursively calculate the hash value of non-leaf nodes from bottom to top. For any... parent node of the layer Its hash value is generated by the next level. Two child nodes of the layer and The hash values ​​are concatenated and then hashed again to obtain the result. Let the hash values ​​of the child nodes be respectively... and ,and The formula for calculating the hash value of the parent node is as follows: (This refers to a string concatenation operation.) pass After several recursive iterations, a unique root node hash value is finally generated. This root hash value serves as the digital fingerprint of all data within the current time window, for any given original data block. Even the slightest bit flip can cause a dramatic, avalanche-like change in the root hash value, thus enabling the system to keenly detect any form of data tampering.

[0049] After generating the digital fingerprint, S42 executes a distributed evidence storage and anchoring protocol based on a consortium blockchain. In this step, a blockchain transaction containing evidence storage metadata is constructed for the protocol (or the protocol includes). The data structure of this transaction consists of the root hash value. Trusted timestamps generated by the Global Navigation Satellite System Vehicle Unique Identification Code and digital signatures for vehicle terminals Composition. To ensure on-chain efficiency and privacy security, raw data blocks are stored locally or in an encrypted cloud, with only lightweight transaction data being processed. The transaction is broadcast to a consortium blockchain network comprised of nodes from regulatory agencies, insurance companies, and automakers. After the smart contract verifies the terminal signature, it packages the transaction into a block and returns the transaction hash. As evidence for preservation.

[0050] S43 When an accident liability determination or data authenticity is questioned, a data integrity verification process based on Merkel's proof is initiated. This establishes the regulatory body's requirement to verify the... Data blocks Whether the data has been tampered with, only the set of hash values ​​of the required sibling nodes along the path from the leaf node to the root node is extracted (without traversing the entire dataset), forming the Merkel proof path. ,in The height of the tree. Verification function. in accordance with and Recalculate the root hash value : If the calculated result An immutable anchor value on the blockchain Completely consistent, that is, satisfying Then, mathematically, it is rigorously proven that this data block If the data is tamper-proof and authentic, it is considered genuine; otherwise, it is determined that the data is corrupted or falsified. This mechanism enables efficient and low-cost decentralized self-verification and external verification of massive amounts of autonomous driving data without compromising the privacy of the original data.

[0051] This invention, by constructing a closed-loop feedback system of "spatiotemporal calibration - risk assessment - storage decision - evidence verification," produces the following significant beneficial effects compared to existing technologies: 1. Achieved millisecond-level high-fidelity reconstruction of multi-source data in high-speed dynamic scenes. The invention employs a laser point cloud distortion correction algorithm based on Lie group and Lie algebra interpolation, and an online extrinsic parameter optimization algorithm based on reprojection photometric error, outputting extrinsic parameter reliability factors in real time. Existing technologies typically rely on static calibration or low-frequency updates. When a vehicle is traveling at high speed or experiencing severe vibrations, significant displacement occurs within the LiDAR's scanning cycle, leading to spatial misalignment ("ghosting") between the point cloud and the image. This invention utilizes high-frequency IMU data to derive the vehicle's instantaneous pose in the continuous time domain, eliminating non-rigid deformation during radar scanning. Simultaneously, it uses photometric consistency to correct minute sensor displacements in real time. This effectively solves the problem of "spatiotemporal tearing" of data caused by high-speed movement and road bumps, ensuring pixel-level alignment of evidence data in physical space and providing a high-precision three-dimensional reconstruction basis for accident liability determination.

[0052] 2. This invention overcomes the bottleneck of "zero-sample" detection of unknown physical adversarial attacks under truth-defining conditions by constructing a cross-modal projection consistency loss function based on dynamic extrinsic parameters. And by calculating the gradient mapping tensor of the input data relative to the self-supervised loss. To generate an anti-attack index Existing attack detection methods mostly rely on labeled training data to calculate classification loss gradients, which cannot be applied to attacks of unknown categories on open roads (such as novel adversarial patches). This invention utilizes the geometric constraints of the physical world (the consistency between radar depth and visual texture) as a "natural supervisory signal." This cross-modal consistency is broken only when a physical attack occurs, and the gradient exhibits non-semantic high-frequency oscillations. By eliminating the reliance on manually labeled data, this invention achieves real-time, generalized detection of attacks at unknown physical layers, such as lighting interference and sticker spoofing, significantly improving the safety of autonomous driving perception systems.

[0053] 3. This invention improves the robustness of the perception system in scenarios where sensor hardware degrades. When constructing a spatiotemporal graph network for risk reasoning, the edge connection weights are adjusted. Explicitly introduced extrinsic reliability factors truncation function Traditional multimodal fusion algorithms typically assume perfect sensor calibration. Once a sensor physically shifts due to long-term vibration, the fusion algorithm may associate incorrect features, leading to false alarms. This invention introduces... This allows the graph network to "sense" the underlying physical state; when the calibration accuracy decreases, the algorithm automatically cuts off unreliable cross-modal connections, preventing noise from propagating in the graph network. This avoids the failure of the sensing algorithm due to the deterioration of the hardware physical state, significantly reducing the false alarm rate and false negative rate of the system during long-term operation.

[0054] 4. This invention solves the problem of data loss during "covert attacks" under resource constraints. It enhances the resistance to attack index. The state space for reinforcement learning storage decision-making is incorporated, and a reward function including risk feedback is designed; simultaneously, the generation frequency of the Merle tree is dynamically adjusted according to the storage strategy. Existing EDR or data recording systems are mostly based on G-value (collision) triggering, which often fails to trigger storage for "soft attacks" or "dangerous situations" such as inducing vehicles to slowly deviate from their lanes, resulting in the repeated overwriting of key evidence. The decision engine of this invention will forcibly trigger lossless storage when a high-level attack index is detected, even if the vehicle dynamics have not yet become unstable; combined with the variable-frequency Merle tree, it ensures that the granularity of evidence storage matches the risk level. Under limited on-board storage and computing power resources, it achieves full, high-fidelity retention of high-value attack samples and accident precursor data, while significantly reducing the bandwidth cost of blockchain evidence storage.

[0055] 5. A closed-loop secure data governance system with self-healing capabilities has been constructed. This invention establishes a two-way feedback mechanism, whereby the risk assessment module triggers recalibration in reverse when it detects a continuous decrease in consistency, and also considers risk probability. Triggering storage mode switching. Existing technologies typically involve unidirectional data flow (acquisition -> storage), lacking system-level interaction. This invention utilizes upper-layer perception logic (consistency score) to diagnose lower-layer hardware status (calibration failure) and automatically triggers correction; it directly controls storage behavior using risk status. This achieves long-term stable operation of the system without human intervention, forming an intelligent evidence storage system with self-diagnosis and self-recovery capabilities through "perception-guided calibration" and "risk-driven storage." The above description is merely a preferred embodiment of the invention, but the scope of protection of the invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this invention should be included within the scope of protection of this invention.

Claims

1. A method for trusted storage of data in autonomous vehicles, characterized in that, include: The spatiotemporal reference calibration step maps the acquisition times of the LiDAR, camera, and inertial measurement unit onto a global nanosecond-level time axis. Then, an inertial recursive LiDAR point cloud motion distortion correction algorithm is run to eliminate the non-rigid deformation of the point cloud caused by the vehicle's own motion during the LiDAR's mechanical rotation scanning. Next, an online extrinsic parameter adaptive optimization algorithm based on reprojection photometric error is executed to construct a cross-modal feature alignment model, obtaining dynamic extrinsic parameters that accurately describe the relative positional relationships between the sensors at the current moment. ; The risk quantification assessment step first utilizes the dynamic external parameters output from the runaway baseline calibration step. Construct a self-supervised cross-modal projection consistency loss. Calculate the input data pair gradient The local entropy value of this gradient is calculated to generate the adversarial attack index. Then, perform extrinsic-driven dynamic graph network uncertainty inference to construct a spatiotemporal graph. In the construction of this spatiotemporal graph, edge weights The calculation formula explicitly introduces the dynamic extrinsic parameter. To prevent the propagation of erroneous risks by severing unreliable cross-modal connections; and then to integrate the overall counter-threat index semantic consistency score Input a fully connected risk prediction network and output the overall system security probability. This probability, in the form of feedback parameters, serves as the operating parameter for both the empty reference calibration step and the dynamic storage decision step, forming a two-way control closed loop; the two-way control closed loop includes when Exceeding the threshold At that time, the storage module is forcibly triggered to enter lossless high-frequency recording mode to prevent critical attack scene data from being compressed and lost; when Continuously below the threshold If the physical external parameter is determined to be in failure, the runaway baseline calibration step is executed, and the current characteristic deviation is used as a priori constraint for the next external parameter optimization.

2. The method for trusted storage of autonomous vehicle data according to claim 1, characterized in that, In the method, the dynamic storage decision step includes: running a real-time accident risk entropy value calculation algorithm based on multi-dimensional feature fusion to quantify the comprehensive accident risk entropy value of the current driving scenario. ; ,in These are the normalized weighting coefficients for dynamic instability, environmental threat probability density, and similarity of historical accident characteristics, respectively. Vehicle dynamics instability ,in, For a moment, For vehicle lateral acceleration, For longitudinal acceleration, For yaw rate, The real-time state vector is formed. This is a vector of pre-statistically calculated mean dynamic parameters under normal driving conditions. This is the corresponding covariance matrix; Expressed as the inverse sum of the threat levels of all targets. ,in To prevent extremely small constants with a denominator of zero, For the number of potential risk targets, the first The estimated collision time for each target is , This is the distance attenuation factor; Based on the calculated comprehensive accident risk entropy value It runs an adaptive storage space allocation algorithm based on deep reinforcement learning. Through continuous online learning and inference, it dynamically adjusts the data storage strategy according to real-time traffic conditions. When the entropy value is low, it automatically executes a high compression or discard strategy to release space. When a high-risk entropy value is detected, it immediately switches to lossless storage mode. Among them, the adaptive memory allocation algorithm of deep reinforcement learning employs an agent, whose state space... Based on the current storage buffer occupancy rate Current frame risk entropy value And the decision-making actions of the previous moment. Composition; Action Space A set of discrete storage strategies These correspond to lossless high-frequency storage, lossy compressed storage, and discarding / not storing, respectively; the agent includes a reward function that combines value incentives and spatial penalties. ,in For indicator functions, This is the buffer pressure penalty coefficient, used to suppress excessively rapid growth of storage space. The overflow penalty constant is The buffer capacity is set to the maximum; the agent uses a deep Q-network to fit the action value function. And iteratively updated using the Bellman equation: In the formula For learning rate, This is the discount factor.

3. A trusted data storage method for autonomous vehicles according to claim 1 or 2, characterized in that, The method further includes a trusted evidence verification step, which includes: A lightweight data fingerprint generation algorithm based on Merkle trees is run to generate digital fingerprints of the original data stream to be stored. The original data stream to be stored is then divided into segments according to a fixed time window. A separate data block, denoted as Using the SHA-256 hash function Perform one-way encryption on each data block to generate the corresponding leaf node hash value. A binary tree structure is used to recursively calculate the hash value of non-leaf nodes from bottom to top. For any nth node... parent node of the layer Its hash value is generated by the next level. Two child nodes of the layer and The hash values ​​are concatenated and then hashed again to obtain the result; let the hash values ​​of the child nodes be respectively and ,and This indicates a string concatenation operation, then the parent node's hash value. ;pass After several recursive iterations, a unique root node hash value is finally generated. As a digital fingerprint of all data within the current time window; Construct a distributed evidence storage and anchoring protocol based on a consortium blockchain, which includes blockchain transaction transactions for evidence storage metadata. The transaction The data structure consists of a root hash value Trusted timestamps generated by the Global Navigation Satellite System Vehicle Unique Identification Code and digital signatures for vehicle terminals composition; When accident liability is determined or the authenticity of data is questioned, a data integrity verification process based on Merkel's proof is initiated, specifying the steps that need to be verified. Data blocks Whether it has been tampered with, only the set of sibling node hashes required on the path from the leaf node to the root node is extracted to form the Merkel proof path. ,in The height of the tree; verification function in accordance with and Recalculate the root hash value If the calculated result is This proves that the data block If the data is authentic and valid, it has not been tampered with; otherwise, it is determined that the data is corrupted or falsified.

4. The trusted data storage method for autonomous vehicles according to claim 3, characterized in that, Raw data is stored locally or in an encrypted cloud, with only lightweight transaction data stored. The transaction is broadcast to a consortium blockchain network comprised of nodes from regulatory agencies, insurance companies, and automakers; after the smart contract verifies the terminal signature, it packages the lightweight transaction into a block and returns the transaction hash. As evidence for preservation.

5. A trusted data storage method for autonomous vehicles according to claim 3, characterized in that, The inertial recursive laser point cloud motion distortion correction algorithm includes: assuming a frame of point cloud data from a lidar... Include There are 1 laser point, denoted as _____. , of which laser points The actual physical acquisition time is The inertial recursive laser point cloud motion distortion correction algorithm selects the starting time of the point cloud in that frame. The angular velocity acquired using a high-frequency inertial measurement unit serves as the reference moment. and linear acceleration Derivation of time using Lie groups and Lie algebraic integrals Relative to the reference time Vehicle relative pose transformation matrix ; to the original distortion points After uniformly correcting to the reference time coordinate system, the distortion-free point coordinates are obtained. ,in This is the installation extrinsic parameter matrix of the lidar coordinate system relative to the vehicle coordinate system.

6. A method for trusted storage of autonomous vehicle data according to claim 5, characterized in that, For time arrive The tiny time interval, vehicle rotation matrix With translation vector The recursive formula is as follows: in, For the instantaneous speed of the vehicle, The gravity vector and These are the zero bias errors of the gyroscope and accelerometer, respectively. For Lie algebra To Liqun Exponential mapping; through the exponential mapping from arrive Integrate the measurements from all inertial measurement units within the time period to obtain the time interval. relative rotation matrix And relative translation vectors, and then construct homogeneous transformation matrix. .

7. A trusted data storage method for autonomous vehicles according to claim 1 or 5, characterized in that, The spatiotemporal reference calibration step executes an online extrinsic parameter adaptive optimization algorithm based on reprojection photometric error. Using the distortion-corrected laser point cloud and synchronously acquired image data, a cross-modal feature alignment model is constructed, including: Pixel coordinates are obtained from the extrinsic parameter matrix to be optimized. ,in, Let be the extrinsic parameter matrix to be optimized, which contains rotational components. Translation components The algorithm first divides the three-dimensional laser points Through camera intrinsic parameter matrix and current external references Projecting onto the image plane yields pixel coordinates. :in This is the depth value of the point in the camera coordinate system; Calculate the image grayscale gradient or edge feature intensity at the projection point, and construct a loss function that minimizes the photometric reprojection error. ,in, It is the Huber kernel function. The number of valid feature points, The image feature values ​​at the projection points are used. The nonlinear least squares problem is iteratively solved using the Gauss-Newton method or the Levenberg-Marquardt algorithm, calculating the Jacobian matrix and updating the state variables until the loss function converges, thus obtaining the dynamic extrinsic parameters. .

8. As described in claim 1, characterized in that, The spatiotemporal calibration steps include: constructing a vehicle-wide global time synchronization network based on the IEEE 1588 Precision Time Protocol; sending synchronization and follow messages through the master clock node; and calculating the clock deviation of the slave clock nodes. Path delay Hardware timestamps are used to uniformly map the acquisition times of LiDAR, cameras, and inertial measurement units to a global nanosecond-level time axis. This provides a unified time reference for subsequent data processing.

9. A trusted data storage method for autonomous vehicles according to claim 1, characterized in that, The gradient-based ground truth-free adversarial perturbation detection includes: assuming the lidar point cloud features aligned by the spatiotemporal reference calibration module are... The visual features projected onto the image plane are Construct a cross-modal consistency loss function ,in For dynamic extrinsic parameters Feature projection operator, The input is the original multimodal data; during the inference phase, the input data is computed through backpropagation. Relative to consistency loss gradient mapping tensor ; through calculation The local information entropy density and Laplace variance are used to quantify the counter-attack threat index. .

10. A trusted data storage method for autonomous vehicles according to claim 7, characterized in that, The set of edges in the spacetime graph It depends on the real-time state output by the spatiotemporal reference calibration module; for spatially associated edges Its connection weight By node and Euclidean distance in a unified spacetime coordinate system Decide; ,in The confidence factor of the extrinsic parameters is calculated based on the residuals of the dynamic extrinsic parameter optimization in Module 1. For the adaptive mapping function; and construct the spatiotemporal graph to run multi-layer graph attention convolution operation, for the th Layer node features A multi-head attention mechanism is introduced to calculate aggregated features: in For nodes The neighborhood, The attention coefficients are used. Through graph convolution aggregation, the algorithm outputs a cross-modal semantic consistency score for each perceived object. .

Images