Search

CN-122009248-A - Storage optimization method, device and system for automatic driving data

CN122009248ACN 122009248 ACN122009248 ACN 122009248ACN-122009248-A

Abstract

The invention provides a storage optimization method, device and system for automatic driving data, which comprises the steps of collecting multi-mode data, scene perception characteristics and event information of a vehicle, evaluating scene value and event triggering probability of the vehicle in each data segment, generating a candidate strategy set according to the scene value, the event triggering probability, compliance constraint and the resource state of the vehicle, screening out an optimal strategy, executing optimization processing and transmission on the optimal strategy, and executing difference synchronization and strategy closed loop updating under a network interrupt and recovery scene, solving at least one of the problems of rough resource allocation, low-value data occupation bandwidth caused key evidence loss, dependence on a hard threshold, incapability of covering potential high-risk segments, multi-mode independent coding, high redundancy, poor compression effect, compliance rule solidification, difficult cross-region adaptation, no priority of weak network transmission, unreliable return, no resource closed loop and unstable system efficiency of the existing automatic driving data.

Inventors

  • Mi Fujia

Assignees

  • 宁波均联智行科技股份有限公司

Dates

Publication Date
20260512
Application Date
20260414

Claims (10)

  1. 1. A storage optimization method for automatic driving data, the storage optimization method comprising: Collecting multi-mode data of a vehicle, and extracting scene perception characteristics and event information of the vehicle; based on the scene perception features and the event information, evaluating the scene value and event triggering probability of the vehicle in each data segment; Loading a preset compliance template to determine a compliance beam condition for data processing; generating a candidate strategy set conforming to the multi-mode data according to the scene value, the event triggering probability, the compliance constraint condition and the resource state of the vehicle; screening an optimal strategy from the candidate strategy set according to the current mode of the vehicle; And executing optimization processing and transmission on the optimal strategy, and executing difference synchronization and strategy closed-loop updating under the network interruption and recovery scene.
  2. 2. The storage optimization method of claim 1, wherein said performing optimization processing and transmission on said optimal policy comprises: Performing adaptive coding and cross-mode redundancy elimination according to the optimal strategy; placing the optimization results of completing the self-adaptive coding and the cross-mode redundancy elimination into different transmission queues according to the priority sequence formulated by the optimal strategy; selecting corresponding data fragments according to the real-time network feedback and the priority order to transmit; In a network congestion or weak network state, the transmission priorities of event fragments and high-value fragments are automatically improved, and low-value fragments are actively discarded or degraded.
  3. 3. The storage optimization method of claim 1, wherein the screening out the optimal policy from the candidate policy set comprises: solving the maximum value integral under the constraint of the resource state, and outputting the optimization result of the multi-mode data; wherein the optimization objective function that maximizes the value integral is: Equation 1: ; The constraint conditions of the resource state are as follows: Equation 2: ; In the formula, The maximum value is indicated and the maximum value, The value of the scene is represented by the value of the scene, Representing the event trigger probability, i representing an i-th data segment in the multimodal data, w representing a set of n of the data segments, M representing a set of the multimodal data, M representing a subset of M, A policy decision vector representing a subset m at the ith of said data segments, a representing coefficients of said scene value, β representing coefficients of said event trigger probability, Represents the resource cost function required to execute the optimal strategy, lambda represents the loss weight of the resource cost, Representing a sum of storage occupancy for the optimal policy, Representing an upstream wideband occupancy summation over the optimal strategy, Indicating the current remaining available storage capacity of the vehicle, Indicating the current network available upstream bandwidth of the vehicle, Representing the set of policies allowed by the aggregate panel.
  4. 4. The method of storage optimization as claimed in claim 1, wherein, The scene value is comprehensively determined according to the scene perception characteristics; the event triggering probability is obtained by predicting a lightweight time sequence model according to the event information and is used for predicting whether an event is triggered in a first time period in the future; The scene perception features comprise at least one of scene complexity, interaction subject density, near-miss-touch index, system boundary proximity, key region coverage rate, collision time threshold value, vehicle transverse and longitudinal acceleration peak value, number of target objects, density of target objects, road structure, illumination condition, weather condition, shielding state and vehicle end safety self-evaluation level; the event information includes at least one of a collision, a sudden deceleration, a driving assistance trigger, and a driver take over.
  5. 5. The method for optimizing storage according to any one of claim 2 to 4, wherein, The compliance constraint condition comprises at least one of a data minimum fidelity rule, a data retention time length rule, a privacy information anonymization level rule and an event evidence obtaining window rule; The data minimum fidelity rule limits the lower limit of video resolution and frame rate and the lower limit of the data density of the point cloud in the event window; the data retention time length rule distinguishes the minimum storage period of the common data fragment and the event data fragment; the privacy information anonymization level rule limits a desensitization processing level standard of the privacy information; The event evidence obtaining window rule limits that data in a preset time length before and after the event occurrence time is taken as a reference to be preserved by adopting the highest fidelity total.
  6. 6. The storage optimization method according to claim 5, wherein the multi-modal data includes vehicle-mounted video data, radar data, point cloud data, CAN signals and vehicle log data; Performing adaptive encoding and cross-modality de-redundancy according to the optimal strategy includes: Executing at least one of adaptive code rate regulation, adaptive GOP setting, ROI QP adaptive regulation and scene switching I frame densification treatment on the vehicle-mounted video data; Performing voxel rasterization ratio adaptive adjustment and/or octree sparsification ratio adaptive adjustment on the point cloud data; performing self-adaptive adjustment of a reflection intensity threshold value and/or sparsification processing of tracking track points on the radar data; And performing differential coding and/or semantic redundancy elimination processing on the CAN signal and the log data.
  7. 7. The method of storage optimization of claim 6, wherein, Performing adaptive encoding and cross-modality de-redundancy according to the optimal strategy includes: Extracting geometric information and motion vectors of the vehicle based on the point cloud data and the radar data, and generating a video ROI mask; And reducing the code rate of the video non-ROI area according to the video ROI mask, and executing redundancy elimination regulation for improving quantization parameters or reducing the frame rate on the video non-ROI area.
  8. 8. The method of storage optimization of claim 5, wherein, Any one candidate strategy in the candidate strategy set comprises at least one of coding quality level of corresponding modal data, sparseness degree of point cloud, data granularity of radar, anonymization level, data retention time length, data uploading priority and privacy desensitization level.
  9. 9. A storage optimizing device of automatic driving data, characterized in that the storage optimizing device is configured to implement the storage optimizing method of automatic driving data according to any one of claims 1 to 8, the storage optimizing device of automatic driving data comprising: The system comprises a data extraction unit (100), wherein the data extraction unit (100) is used for collecting multi-mode data of a vehicle and extracting scene perception characteristics and event information of the vehicle; A data evaluation unit (200), wherein the data evaluation unit (200) is used for evaluating the scene value and event triggering probability of the vehicle in each data segment based on the scene perception feature and the event information; A template loading unit (300), the template loading unit (300) being configured to load a preset compliance template to determine a compliance beam condition for data processing; a policy generation unit (400), where the policy generation unit (400) is configured to generate a candidate policy set according to the scenario value, the event trigger probability, the compliance constraint, and the resource status of the vehicle; A policy screening unit (500), where the policy screening unit (500) is configured to screen an optimal policy from the candidate policy set according to a current modality of the vehicle; And the policy execution unit (600) is used for executing optimization processing and transmission on the optimal policy and executing difference synchronization and policy closed-loop updating under the network interruption and recovery scene.
  10. 10. A storage optimizing system of automatic driving data, characterized in that the storage optimizing system of automatic driving data is applied to the storage optimizing apparatus of automatic driving data according to claim 9 to execute the storage optimizing method of automatic driving data according to any one of claims 1 to 8.

Description

Storage optimization method, device and system for automatic driving data Technical Field The invention relates to the technical field of automatic driving, in particular to a storage optimization method, a storage optimization device and a storage optimization system for automatic driving data. Background Along with the evolution of the automatic driving technology from auxiliary driving to highly automatic driving, a vehicle-mounted system needs to acquire, process and return massive multi-source heterogeneous sensing data in real time, including videos, point clouds, milliradars, CAN post-signaling, vehicle logs and the like, and the data are not only input basis of automatic driving decision control, but also core digital evidence of accident backtracking, risk analysis, algorithm iteration and compliance evidence obtaining. The main stream solutions aiming at automatic driving data storage and uploading in the current industry mostly adopt fixed configuration and event driving modes, for example, a fixed code rate and a fixed sampling rate are commonly adopted for videos, point clouds and radars, acquisition quality is temporarily improved or buffering is started only after hard threshold events such as collision, rapid deceleration and the like are triggered, a data retention strategy mostly uniformly sets retention duration and quality lower limit according to vehicle types or regions, scene complexity and risk level are not distinguished, each sensing mode is independently compressed and encoded, and cross-mode information coordination and the like are not carried out. However, in the prior art, at least one of the following problems of rough resource allocation, critical evidence loss caused by bandwidth occupation of low-value data, dependence on a hard threshold, incapability of covering potential high-risk fragments, multi-mode independent coding, high redundancy, poor compression effect, compliance rule solidification, difficult cross-region adaptation, no priority of weak network transmission, unreliable return, strategy non-resource closed loop and unstable system efficiency exist in the storage and uploading of the automatic driving data. Disclosure of Invention The method solves the technical problems that in the prior art, at least one of the following problems of rough resource allocation, key evidence loss caused by bandwidth occupation of low-value data, dependence on a hard threshold, incapability of covering potential high-risk fragments, multi-mode independent coding, high redundancy, poor compression effect, compliance rule solidification, difficult cross-region adaptation, no priority of weak network transmission, unreliable return, strategy non-resource closed loop and unstable system efficiency exist in the automatic driving data storage and uploading. In order to solve the above problems, in a first aspect, the present invention provides a storage optimization method for automatic driving data, the storage optimization method comprising: collecting multi-mode data of a vehicle, and extracting scene perception characteristics and event information of the vehicle; based on scene perception characteristics and event information, evaluating scene values and event triggering probabilities of vehicles in all data fragments; Loading a preset compliance template to determine a compliance beam condition for data processing; Generating a candidate strategy set conforming to multi-mode data according to scene value, event triggering probability, compliance constraint conditions and resource states of vehicles; screening an optimal strategy from the candidate strategy set according to the current mode of the vehicle; and executing optimization processing and transmission on the optimal strategy, and executing difference synchronization and strategy closed-loop updating under the network interruption and recovery scene. Compared with the prior art, the technical effect achieved by adopting the technical scheme is that scene perception characteristics and event information are extracted from the multi-mode data of the vehicle, scene value and event triggering probability of data fragments are evaluated, the optimal strategy adapting to the current mode is generated and screened out by combining compliance constraint of the compliance template and real-time resource state of the vehicle, targeted optimization processing and transmission are executed, and the difference synchronization and strategy closed-loop updating mechanism after network interrupt recovery are matched, so that the resource refined distribution of automatic driving data storage and uploading is realized, the limitation of hard threshold triggering is broken through, the effective coverage of potential high-risk fragments is realized, the multi-mode independent coding mode is abandoned, the data redundancy is reduced, and the compression effect is improved. In addition, the scheme realizes flexible adaptation of the complian