Intelligent driving data closed-loop optimization method and system based on vehicle cloud cooperation and active learning

Abstract

The invention discloses an intelligent driving data closed-loop optimization method and system based on vehicle cloud collaboration and active learning, and the method comprises the steps: collecting a plurality of groups of single-mode data, processing each group of single-mode data based on a lightweight data model, and obtaining cognitive uncertainty and random uncertainty, the lightweight data model is obtained by training a data set, the data set comprises a plurality of groups of single-mode data used for training, weighted summation is carried out on cognitive uncertainty and random uncertainty to obtain confidence, and if the confidence is smaller than a preset threshold, a multi-mode sensor is triggered to synchronously collect complete scene data and upload the complete scene data to a cloud end. According to the method, the lightweight data model is deployed at the vehicle end, the model independently runs at the vehicle end, the uncertainty index of the data acquired by the vehicle end is measured, and then a complex and difficult scene with low perception or decision confidence is automatically recognized.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, specifically to an intelligent driving data closed-loop optimization method and system based on vehicle-cloud collaboration and active learning. Background Technology

[0002] The current development of intelligent driving technology heavily relies on data-driven algorithm iteration. The industry's commonly adopted data closed-loop process typically includes: vehicle data collection, data transmission to the cloud, manual screening and labeling, model training, simulation testing, and OTA updates. Existing technical solutions largely depend on large-scale fleets collecting massive amounts of data and retrieving data through random sampling or simple rules (such as system-triggered takeover events).

[0003] Therefore, the existing technology has the following problems: Low data value density: In the data recovered by traditional methods, a large amount of data consists of simple and repetitive scenarios, which have limited value for model optimization. Meanwhile, the "long tail scenarios" that are truly difficult to handle account for a very small percentage of the data (corner cases), resulting in low efficiency of algorithm iteration.

[0004] Insufficient mining accuracy: Existing data mining methods based on simple rules (such as relying solely on collision and takeover events) have low levels of intelligence and are unable to accurately and automatically identify complex and difficult scenarios that lead to low confidence in model perception or decision-making. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method and system for optimizing closed-loop intelligent driving data based on vehicle-cloud collaboration and active learning. A lightweight data model is deployed on the vehicle side, which runs independently on the vehicle side to measure the uncertainty index of the data acquired by the vehicle side, thereby automatically identifying complex and difficult scenarios with low perception or decision confidence.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0007] According to the first aspect of this application, a closed-loop optimization method for intelligent driving data based on vehicle-cloud collaboration and active learning is provided, applied to the vehicle side, including: Collect several sets of single-modal data; The lightweight data model is used to process each set of single-modal data to obtain cognitive uncertainty and stochastic uncertainty. The lightweight data model is trained using a dataset, which includes several sets of single-modal data for training. The confidence level is obtained by weighting and summing cognitive uncertainty and random uncertainty. If the confidence level is less than the preset threshold, the multimodal sensors will be triggered to synchronously collect complete scene data and upload it to the cloud.

[0008] In some embodiments of this application, based on the foregoing scheme, the single modal data is visual data, radar data, inertial navigation data, environmental data, or vehicle status data.

[0009] In some embodiments of this application, based on the foregoing scheme, the following further methods are also included: Preprocessing the single-modal data includes: Several sets of the single-modal data are aligned to the same time axis using IMU timestamps; If the single-modal data is a point cloud, voxel filtering is used; if the single-modal data is an image, Gaussian filtering or median filtering is used for noise reduction. Map the single-modal data with different dimensions to the same interval; If the single-modal data is optical data or remote sensing data, distortion correction and projection transformation are performed to convert the single-modal data into a reflectance or height map.

[0010] In some embodiments of this application, based on the foregoing scheme, the lightweight data model includes an input layer, a shallow feature extraction module, a deep feature extraction module, an attention module, a feature fusion module, a cognitive uncertainty prediction head, and a random uncertainty prediction head. The input layer is used to input single-modal data; The shallow feature extraction module extracts shallow features of single-modality data through a first depthwise separable convolution. The deep feature extraction module extracts deep features from single-modality data through a second depthwise separable convolution; The attention module is used to enhance key parts of shallow and deep features, resulting in enhanced shallow and enhanced deep features. The feature fusion module is used to fuse enhanced shallow features and enhanced deep features to obtain fused features; The cognitive uncertainty prediction head is used to predict the cognitive uncertainty of the fused features; The random uncertainty prediction head is used to predict the random uncertainty of the fused features.

[0011] In some embodiments of this application, based on the aforementioned scheme, the cognitive uncertainty prediction head includes a Dropout layer, which is used to perform multiple forward propagations on the fused features to obtain several output samples, and to obtain a first variance as cognitive uncertainty based on the several output samples.

[0012] In some embodiments of this application, based on the foregoing scheme, the random uncertainty prediction head directly predicts or uses the second variance of the fused features based on the sensor noise model as random uncertainty.

[0013] In some embodiments of this application, based on the foregoing scheme, the lightweight data model further includes: An evolutionary algorithm is used to initialize a population for the network structure, which consists of a first depthwise separable convolution, a second depthwise separable convolution, an attention module, and a feature fusion module. Based on the depth and width of the network structure, the surrogate model is used to obtain the quantity-efficiency ratio, where the depth is the number of convolutional layers in the first depthwise separable convolution, the second depthwise separable convolution, the attention module, and the feature fusion module, and the width is the number of output channels of each convolutional layer; Based on the dose-efficiency ratio score, network structures with scores below the threshold are eliminated, and a new population is generated. Fine-tune the new population, jump to the step of obtaining the quantity-efficiency ratio using a surrogate model based on the depth and width of the network structure, until the optimal network structure is obtained; Channel-level pruning is performed on the optimal network structure to remove unimportant feature maps.

[0014] According to a second aspect of this application, a system is provided, comprising: The acquisition module is used to acquire several sets of single-modal data. The processing module is used to process each set of single-modal data based on a lightweight data model to obtain the cognitive uncertainty and stochastic uncertainty of each set of single-modal data. The lightweight data model is trained using a dataset, which includes several sets of single-modal data for training. The acquisition module is used to obtain the confidence level by weighted summation of cognitive uncertainty and random uncertainty for each set of single-modal data; The transmission module is used to trigger the multimodal sensors to synchronously collect complete scene data and upload it to the cloud if the confidence level is less than a preset threshold.

[0015] According to a third aspect of this application, a computer-readable storage medium is provided that stores a computer program thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0016] According to a fourth aspect of this application, an electronic device is provided, comprising: One or more processors; A memory for storing executable instructions of the processor, which, when executed by the one or more processors, cause the one or more processors to implement the method described above.

[0017] The beneficial effects of this application are as follows: (1) The intelligent driving data closed-loop optimization method and system based on vehicle-cloud collaboration and active learning provided in this application deploys a lightweight data model on the vehicle. The model runs independently on the vehicle. When several sets of single-modal data are collected, each set of single-modal data is processed based on the lightweight data model. The confidence level is obtained by weighted summation of cognitive uncertainty and random uncertainty. If the confidence level is less than the preset threshold, the multimodal sensors are triggered to synchronously collect complete scene data and upload it to the cloud. It is not necessary to transmit a large amount of data to the cloud for processing, which prevents a large amount of simple and repetitive scene data from appearing in the cloud, resulting in low algorithm iteration efficiency.

[0018] (2) The intelligent driving data closed-loop optimization method and system based on vehicle-cloud collaboration and active learning provided in this application uses cognitive uncertainty and random uncertainty to judge the uncertainty of data. By utilizing the complementarity of model perspective and data perspective, the overall uncertainty of data can be comprehensively and accurately quantified, which can offset the extreme impact of single uncertainty and ensure the availability of data uploaded to the cloud.

[0019] (3) The intelligent driving data closed-loop optimization method and system based on vehicle-cloud collaboration and active learning provided in this application adjusts the network structure of each network of the lightweight data model based on evolutionary algorithm to ensure that the lightweight data model is truly lightweight.

[0020] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and are intended to explain the invention, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of an intelligent driving data closed-loop optimization method based on vehicle-cloud collaboration and active learning according to the present invention. Figure 2 This is a schematic diagram of an intelligent driving data closed-loop optimization system based on vehicle-cloud collaboration and active learning according to the present invention. Figure 3 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0022] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0023] It should be understood that the terms "comprising" and other similar expressions in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, or apparatus that includes a series of steps or units and is not limited to the listed steps or units. Furthermore, "first" and "second" are used to distinguish different objects and are not intended to describe a specific order.

[0024] According to the first aspect of this application, Figure 1 As shown, this embodiment provides a closed-loop optimization method for intelligent driving data based on vehicle-cloud collaboration and active learning, applied to the vehicle side, including: Step S101: Collect several sets of single-mode data.

[0025] In some embodiments of this example, the single modal data is visual data, radar data, inertial navigation data, environmental data, or vehicle status data.

[0026] Specifically, visual data includes camera images (RGB) and infrared thermal images; Radar data includes LiDAR point clouds and millimeter-wave radar echo signals; Inertial navigation data includes pose information provided by IMU (accelerometer, gyroscope); Environmental data includes physical layer parameters such as signal-to-noise ratio (SNR), departure angle (AoA), and radial distance; Vehicle status data includes vehicle CAN bus status, device operation logs, driver operation trajectory, voice interaction logs, timestamps, and geographic location information.

[0027] In some embodiments of this example, preprocessing of the single-modal data is also included, including: Spatiotemporal alignment: Using IMU timestamps, several sets of the single-modal data are aligned to the same time axis to solve the asynchronous sampling problem.

[0028] Noise removal: If the single modal data is a point cloud, voxel filtering is used; if the single modal data is an image, Gaussian filtering or median filtering is used for noise removal.

[0029] Standardization: Mapping single-modal data with different dimensions to the same interval, specifically, mapping data with different dimensions to the interval [0, 1] to prevent gradient explosion.

[0030] Geometric correction: If the single-modal data is optical data or remote sensing data, distortion correction and projection transformation are performed to convert the single-modal data into a reflectance or height map.

[0031] Step S102: Process each set of single-modal data based on the lightweight data model to obtain cognitive uncertainty and stochastic uncertainty. The lightweight data model is trained using a dataset, which includes several sets of single-modal data for training.

[0032] Compared to the traditional method of directly collecting and transmitting raw data, a lightweight data model is deployed separately on the vehicle. This lightweight data model runs independently on the vehicle, collecting raw data related to intelligent driving as well as perception or decision outputs from the large-scale intelligent driving model. This lightweight data model calculates and measures the uncertainty index for each set of individual modal data, which is used to measure the reliability and credibility of the predictions for each set of individual modal data.

[0033] In some implementations of this embodiment, the lightweight data model is based on MobileNetV3 (the third generation of the mobile neural network architecture series) and EfficientNet-Lite (a lightweight version of the efficient network architecture), and includes an input layer, a shallow feature extraction module, a deep feature extraction module, an attention module, a feature fusion module, a cognitive uncertainty prediction head, and a random uncertainty prediction head. The input layer is used to input single-modal data; The shallow feature extraction module extracts shallow features of single-modality data through a first depthwise separable convolution. The deep feature extraction module extracts deep features from single-modality data through a second depthwise separable convolution; The attention module is used to enhance key parts of shallow and deep features, resulting in enhanced shallow and enhanced deep features. The feature fusion module is used to fuse enhanced shallow features and enhanced deep features to obtain fused features; The cognitive uncertainty prediction head is used to predict the cognitive uncertainty of the fused features; The random uncertainty prediction head is used to predict the random uncertainty of the fused features.

[0034] In this embodiment, the shallow feature extraction module and the deep feature extraction module respectively use a first depthwise separable convolution and a second depthwise separable convolution to replace the standard convolution, divide the number of input channels into groups, with only one convolution kernel in each group, and finally mix features through pointwise convolution, which greatly reduces FLOPs (floating-point operations).

[0035] In some implementations of this embodiment, the cognitive uncertainty prediction head uses methods such as Monte Carlo to infer and calculate the variance of the model output to estimate the uncertainty caused by insufficient understanding of the model parameters. The cognitive uncertainty prediction head includes a Dropout layer, which is used to perform multiple forward propagations on the fused features to obtain several output samples, and obtain a first variance as the cognitive uncertainty based on the several output samples.

[0036] Specifically, the Monte Carlo Dropout (MCD) method is used to maintain the activation of the Dropout layer during the inference phase, performing multiple forward propagations (e.g., 10-20 times) to calculate the variance of the output probability distribution. The greater the variance, the higher the cognitive uncertainty.

[0037] In some embodiments of this example, the random uncertainty prediction head directly predicts or uses the second variance of the fused features based on the sensor noise model as random uncertainty.

[0038] In some implementations of this embodiment, the method further includes using AutoML (Automated Machine Learning) technology to search for the optimal network structure depth and width for the first depthwise separable convolution, the second depthwise separable convolution, the attention module, and the feature fusion module, and automatically pruning redundant layers to achieve a "quantity-efficiency ratio".

[0039] In this embodiment, the lightweight data model further includes an adjustment module, which includes: An evolutionary algorithm is used to initialize a population for the network structure, which consists of a first depthwise separable convolution, a second depthwise separable convolution, an attention module, and a feature fusion module. Based on the depth and width of the network structure, the surrogate model is used to obtain the quantity-efficiency ratio, where the depth is the number of convolutional layers in the first depthwise separable convolution, the second depthwise separable convolution, the attention module, and the feature fusion module, and the width is the number of output channels of each convolutional layer; Based on the dose-efficiency ratio score, network structures with scores below the threshold are eliminated, and a new population is generated. Fine-tune the new population, jump to the step of obtaining the quantity-efficiency ratio using a surrogate model based on the depth and width of the network structure, until the optimal network structure is obtained; Channel-level pruning is performed on the optimal network structure to remove unimportant feature maps.

[0040] In this embodiment, since training directly on real hardware is too slow, a proxy model is used to predict performance. The input is a descriptor of the candidate network structure (number of convolutional layers, number of output channels per convolutional layer), and the output is the estimated latency and accuracy. This allows for the rapid filtering out of structures that are obviously "too slow" or "too poor," retaining only the promising ones for the next round.

[0041] In this embodiment, a elimination mechanism is used to retain high-scoring structures based on the surrogate model's score, and a mutation operation is employed: the retained structures are fine-tuned (Depth / Width is adjusted), and final pruning is achieved: channel-level pruning is performed on the optimal structure to remove unimportant feature maps.

[0042] In this way, the final determined structure is exported as the final lightweight data model, achieving acceleration, energy efficiency optimization, and quantity-efficiency balance.

[0043] Step S103: The confidence level is obtained by weighted summation of cognitive uncertainty and random uncertainty.

[0044] In some implementations of this embodiment, a linear inverse mapping strategy is employed. Since a higher uncertainty score U represents greater uncertainty in the model (i.e., greater risk), it is necessary to fuse cognitive uncertainty and stochastic uncertainty and map them to a confidence level C of [0, 1], where C=1 represents complete certainty and C=0 represents extreme uncertainty. The calculation formula is as follows:

[0045] Where U is the overall uncertainty score (0≤U≤1) and C is the confidence level of the final output (0≤C≤1).

[0046] Step S104: If the confidence level is less than the preset threshold, the multimodal sensor is triggered to synchronously collect complete scene data and upload it to the cloud.

[0047] Specifically, if the confidence level is less than a preset threshold, the multi-modal sensors (camera, lidar, millimeter-wave radar) are automatically triggered to synchronously collect and cache high-frame-rate complete scene data and transmit it back to the cloud via TBOX.

[0048] In one specific embodiment, vehicle A is labeled "aggressive driving style and low parking frequency". According to the traditional routine, the cloud sends the configuration file to the database. After the vehicle is powered on and connected to the network, a large amount of simple and repetitive scenario data will be collected.

[0049] For users with aggressive driving styles, a separate data collection task is designed. Users tagged as having aggressive driving styles are selected in batches, and data collection parameters are configured. The threshold condition for exiting intelligent driving is increased. The file is then sent to the vehicle. Subsequently, data collection will not be triggered when such users lightly press the brake. Only when they press the brake hard will data collection be triggered. This prevents the collection of a large amount of simple and repetitive scenario data and ensures that "long-tail scenario" data (Corner Cases) is collected.

[0050] Thus, the intelligent driving data closed-loop optimization method based on vehicle-cloud collaboration and active learning provided in this embodiment deploys a lightweight data model on the vehicle side. This model runs independently on the vehicle side. When several sets of single-modal data are collected, each set of single-modal data is processed based on the lightweight data model. Cognitive uncertainty and random uncertainty are used to judge the uncertainty of the data. By utilizing the complementarity of model perspective and data perspective, the overall uncertainty of the data can be comprehensively and accurately quantified. This can offset the extreme impact of single uncertainty, ensure the availability of data uploaded to the cloud, and avoid the problem of low algorithm iteration efficiency caused by a large amount of simple and repetitive scenario data in the cloud.

[0051] According to the second aspect of this application, such as Figure 2 As shown in the figure, this embodiment provides an intelligent driving data closed-loop optimization system based on vehicle-cloud collaboration and active learning, including: The acquisition module is used to acquire several sets of single-modal data. The processing module is used to process each set of single-modal data based on a lightweight data model to obtain the cognitive uncertainty and stochastic uncertainty of each set of single-modal data. The lightweight data model is trained using a dataset, which includes several sets of single-modal data for training. The acquisition module is used to obtain the confidence level by weighted summation of cognitive uncertainty and random uncertainty for each set of single-modal data; The transmission module is used to trigger the multimodal sensors to synchronously collect complete scene data and upload it to the cloud if the confidence level is less than a preset threshold.

[0052] Specifically, this embodiment corresponds one-to-one with the above method embodiments. The functions of each module have been described in detail in the corresponding method embodiments, so they will not be repeated here.

[0053] According to a third aspect of this application, this embodiment provides a computer-readable storage medium having a computer program stored thereon, the computer program including executable instructions that, when executed by a processor, implement the method described above.

[0054] The present invention can implement all or part of the processes in the above methods, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0055] According to the fourth aspect of this application, such as Figure 3 As shown, an electronic device is provided, comprising: One or more processors; Memory is used to store executable instructions for the processor, which, when executed by one or more processors, cause one or more processors to implement the methods described above.

[0056] Electronic devices are manifested in the form of general-purpose computing devices. Components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and a bus connecting different system components (including memory and processor).

[0057] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of a computer system, connecting all parts of the computer system through various interfaces and lines.

[0058] Memory can be used to store computer programs and / or modules. The processor implements various functions of the computer system by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system and at least one application program required for a function (e.g., sound playback, image playback, etc.); the data storage area can store data created based on the use of the mobile phone (e.g., audio data, video data, etc.). Furthermore, memory can include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMedia Cards (SMC), Secure Digital (SD) cards, Flash Cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, servers, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and memory) containing computer-usable program code.

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), servers, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0064] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A closed-loop optimization method for intelligent driving data based on vehicle-cloud collaboration and active learning, applied to the vehicle side, characterized in that, include: Collect several sets of single-modal data; The lightweight data model is used to process each set of single-modal data to obtain cognitive uncertainty and stochastic uncertainty. The lightweight data model is trained using a dataset, which includes several sets of single-modal data for training. The confidence level is obtained by weighting and summing cognitive uncertainty and random uncertainty. If the confidence level is less than the preset threshold, the multimodal sensors will be triggered to synchronously collect complete scene data and upload it to the cloud.

2. The method according to claim 1, characterized in that: The single-modal data can be visual data, radar data, inertial navigation data, environmental data, or vehicle status data.

3. The method according to claim 1, characterized in that, Also includes: Preprocessing the single-modal data includes: Several sets of the single-modal data are aligned to the same time axis using IMU timestamps; If the single-modal data is a point cloud, voxel filtering is used; if the single-modal data is an image, Gaussian filtering or median filtering is used for noise reduction. Map the single-modal data with different dimensions to the same interval; If the single-modal data is optical data or remote sensing data, distortion correction and projection transformation are performed to convert the single-modal data into a reflectance or height map.

4. The method according to claim 1, characterized in that: The lightweight data model includes an input layer, a shallow feature extraction module, a deep feature extraction module, an attention module, a feature fusion module, a cognitive uncertainty prediction head, and a random uncertainty prediction head. The input layer is used to input single-modal data; The shallow feature extraction module extracts shallow features of single-modality data through a first depthwise separable convolution. The deep feature extraction module extracts deep features from single-modality data through a second depthwise separable convolution; The attention module is used to enhance key parts of shallow and deep features, resulting in enhanced shallow and enhanced deep features. The feature fusion module is used to fuse enhanced shallow features and enhanced deep features to obtain fused features; The cognitive uncertainty prediction head is used to predict the cognitive uncertainty of the fused features; The random uncertainty prediction head is used to predict the random uncertainty of the fused features.

5. The method according to claim 1, characterized in that: The cognitive uncertainty prediction head includes a Dropout layer, which is used to perform multiple forward propagations on the fused features to obtain several output samples, and obtain a first variance as the cognitive uncertainty based on the several output samples.

6. The method according to claim 1, characterized in that: The random uncertainty prediction head directly predicts or uses the second variance of the fused features based on the sensor noise model as random uncertainty.

7. The method according to claim 5, characterized in that, The lightweight data model also includes an adjustment module, which includes: An evolutionary algorithm is used to initialize a population for the network structure, which consists of a first depthwise separable convolution, a second depthwise separable convolution, an attention module, and a feature fusion module. Based on the depth and width of the network structure, the surrogate model is used to obtain the quantity-efficiency ratio, where the depth is the number of convolutional layers in the first depthwise separable convolution, the second depthwise separable convolution, the attention module, and the feature fusion module, and the width is the number of output channels of each convolutional layer; Based on the dose-efficiency ratio score, network structures with scores below the threshold are eliminated, and a new population is generated. Fine-tune the new population, jump to the step of obtaining the quantity-efficiency ratio using a surrogate model based on the depth and width of the network structure, until the optimal network structure is obtained; Channel-level pruning is performed on the optimal network structure to remove unimportant feature maps.

8. A closed-loop optimization system for intelligent driving data based on vehicle-cloud collaboration and active learning, characterized in that, include: The acquisition module is used to acquire several sets of single-modal data. The processing module is used to process each set of single-modal data based on a lightweight data model to obtain the cognitive uncertainty and stochastic uncertainty of each set of single-modal data. The lightweight data model is trained using a dataset, which includes several sets of single-modal data for training. The acquisition module is used to obtain the confidence level by weighted summation of cognitive uncertainty and random uncertainty for each set of single-modal data; The transmission module is used to trigger the multimodal sensors to synchronously collect complete scene data and upload it to the cloud if the confidence level is less than a preset threshold.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program includes executable instructions that, when executed by a processor, implement the method of any one of claims 1-7.

10. An electronic device, characterized in that, include: One or more processors; A memory for storing executable instructions of the processor, which, when executed by the one or more processors, cause the one or more processors to perform the method according to any one of claims 1-7.

Images