Vehicle data processing method and device, computer device and storage medium

Abstract

The application relates to a vehicle data processing method and device, computer equipment and a storage medium, and relates to the technical field of intelligent driving. The method comprises the following steps: acquiring a plurality of target image data sent by a vehicle end, processing each target image data, and obtaining a dimension-reduced feature vector corresponding to each target image data; processing each dimension-reduced feature vector according to a kernel density estimation algorithm to obtain a probability density distribution result of the dimension-reduced feature vector; determining a candidate feature vector from the dimension-reduced feature vector according to the probability density distribution result; performing clustering processing on each candidate feature vector to obtain a target feature vector; generating a data collection task according to the target feature vector and sending the data collection task to the vehicle end; and the data collection task is used to instruct the vehicle end to collect target image data corresponding to the target feature vector. The method can reduce resource waste in vehicle data collection and processing.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a vehicle data processing method, apparatus, computer equipment, and storage medium. Background Technology

[0002] With the development of intelligent driving technology, data-driven deep learning models have become the foundation for core aspects of perception, decision-making, and control in intelligent driving processes. The performance of these models is highly dependent on the scale, quality, and diversity of the training data. However, the primary training data acquisition method in related technologies is full-data collection, which requires the vehicle to completely record and upload all sensor data generated during operation to the cloud. This method is simple to implement and ensures no data is missed.

[0003] However, the training data acquisition methods of related technologies collect all data indiscriminately, resulting in high data redundancy and thus wasting resources. Summary of the Invention

[0004] Therefore, it is necessary to provide a vehicle data processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product that can reduce resource waste in response to the above-mentioned technical problems.

[0005] In a first aspect, this application provides a vehicle data processing method, including:

[0006] Multiple target image data sent by the vehicle are acquired, and each target image data is processed to obtain a dimension-reduced feature vector corresponding to each target image data.

[0007] The probability density distribution of the dimensionality-reduced feature vectors is obtained by processing each of the dimensionality-reduced feature vectors according to the kernel density estimation algorithm.

[0008] Based on the probability density distribution results, candidate feature vectors are determined from the dimensionality-reduced feature vectors;

[0009] Clustering is performed on each of the candidate feature vectors to obtain the target feature vector;

[0010] A data acquisition task is generated based on the target feature vector, and the data acquisition task is sent to the vehicle terminal; the data acquisition task is used to instruct the vehicle terminal to acquire target image data corresponding to the target feature vector.

[0011] In one embodiment, the target image data is processed to obtain a dimensionality-reduced feature vector corresponding to each target image data, including:

[0012] Feature extraction is performed on each of the target image data to obtain an initial feature vector corresponding to each of the target image data;

[0013] Principal component analysis (PCA) algorithm is used to reduce the dimensionality of each initial eigenvector, resulting in multiple dimensionality-reduced eigenvectors.

[0014] In one embodiment, the feature vector set composed of the candidate feature vectors is processed according to a kernel density estimation algorithm to obtain the probability density distribution result of the feature vector set, including:

[0015] The density values ​​of each of the dimensionality-reduced eigenvectors are calculated according to the kernel density estimation algorithm.

[0016] Based on the density values ​​of each of the dimensionality-reduced feature vectors, the probability density distribution of the dimensionality-reduced feature vectors is determined.

[0017] In one embodiment, determining candidate feature vectors from the dimensionality-reduced feature vectors based on the probability density distribution results includes:

[0018] The low density threshold is determined based on the probability density distribution results and their corresponding preset quantile values.

[0019] Based on the low density threshold, a target density value that is less than the low density threshold is determined from the density values ​​of each of the dimensionality-reduced feature vectors.

[0020] The dimensionality-reduced feature vector corresponding to the target density value is determined as the candidate feature vector.

[0021] In one embodiment, clustering is performed on each of the candidate feature vectors to obtain a target feature vector, including:

[0022] Clustering is performed on all the candidate feature vectors to obtain multiple feature vector clusters;

[0023] The candidate feature vector corresponding to the cluster center of the feature vector cluster is determined as the target feature vector.

[0024] In one embodiment, sending the data acquisition task to the vehicle includes:

[0025] Based on the target feature vector, determine the target scene corresponding to the target feature vector;

[0026] Based on the collected status data of multiple vehicles, the target vehicle associated with the target scene is determined;

[0027] The data acquisition task is sent to the vehicle end of the target vehicle.

[0028] In one embodiment, the vehicle data processing method further includes:

[0029] Obtain the availability score corresponding to each of the target image data sent by the vehicle terminal;

[0030] Statistical processing is performed on each of the aforementioned availability scores to obtain the availability score threshold;

[0031] The availability score threshold is sent to the vehicle terminal; the availability score threshold is used to instruct the vehicle terminal to process real-time image data with an availability score less than the availability score threshold in order to obtain target image data.

[0032] Secondly, this application also provides a vehicle data processing device, comprising:

[0033] The data acquisition module is used to acquire multiple target image data sent by the vehicle terminal, and process each target image data to obtain the dimension-reduced feature vector corresponding to each target image data.

[0034] The density modeling module is used to process each of the dimensionality-reduced feature vectors according to the kernel density estimation algorithm to obtain the probability density distribution results of the dimensionality-reduced feature vectors;

[0035] The feature determination module is used to determine candidate feature vectors from the dimensionality-reduced feature vectors based on the probability density distribution results.

[0036] The feature processing module is used to perform clustering processing on each of the candidate feature vectors to obtain the target feature vector;

[0037] The task generation module is used to generate a data acquisition task based on the target feature vector and send the data acquisition task to the vehicle terminal; the data acquisition task is used to instruct the vehicle terminal to acquire target image data corresponding to the target feature vector.

[0038] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0039] Multiple target image data sent by the vehicle are acquired, and each target image data is processed to obtain a dimension-reduced feature vector corresponding to each target image data.

[0040] The probability density distribution of the dimensionality-reduced feature vectors is obtained by processing each of the dimensionality-reduced feature vectors according to the kernel density estimation algorithm.

[0041] Based on the probability density distribution results, candidate feature vectors are determined from the dimensionality-reduced feature vectors;

[0042] Clustering is performed on each of the candidate feature vectors to obtain the target feature vector;

[0043] A data acquisition task is generated based on the target feature vector, and the data acquisition task is sent to the vehicle terminal; the data acquisition task is used to instruct the vehicle terminal to acquire target image data corresponding to the target feature vector.

[0044] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0045] Multiple target image data sent by the vehicle are acquired, and each target image data is processed to obtain a dimension-reduced feature vector corresponding to each target image data.

[0046] The probability density distribution of the dimensionality-reduced feature vectors is obtained by processing each of the dimensionality-reduced feature vectors according to the kernel density estimation algorithm.

[0047] Based on the probability density distribution results, candidate feature vectors are determined from the dimensionality-reduced feature vectors;

[0048] Clustering is performed on each of the candidate feature vectors to obtain the target feature vector;

[0049] A data acquisition task is generated based on the target feature vector, and the data acquisition task is sent to the vehicle terminal; the data acquisition task is used to instruct the vehicle terminal to acquire target image data corresponding to the target feature vector.

[0050] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0051] Multiple target image data sent by the vehicle are acquired, and each target image data is processed to obtain a dimension-reduced feature vector corresponding to each target image data.

[0052] The probability density distribution of the dimensionality-reduced feature vectors is obtained by processing each of the dimensionality-reduced feature vectors according to the kernel density estimation algorithm.

[0053] Based on the probability density distribution results, candidate feature vectors are determined from the dimensionality-reduced feature vectors;

[0054] Clustering is performed on each of the candidate feature vectors to obtain the target feature vector;

[0055] A data acquisition task is generated based on the target feature vector, and the data acquisition task is sent to the vehicle terminal; the data acquisition task is used to instruct the vehicle terminal to acquire target image data corresponding to the target feature vector.

[0056] The aforementioned vehicle data processing method, apparatus, computer equipment, computer-readable storage medium, and computer program product acquire multiple target image data sent by the vehicle terminal and process each target image data to obtain a dimensionality-reduced feature vector corresponding to each target image data, thereby improving the efficiency of subsequent statistical and clustering analysis while retaining the key semantic information of the target image data. Further, each dimensionality-reduced feature vector is processed according to a kernel density estimation algorithm to obtain the probability density distribution result of the dimensionality-reduced feature vector, which is then used by the cloud server to understand the density or rarity of the dimensionality-reduced feature vector of each target image data in the feature space of the entire feature vector set. Based on the probability density distribution result, candidate feature vectors are determined from the dimensionality-reduced feature vectors. Each candidate feature vector is clustered to obtain the most representative target feature vector. Further, a data acquisition task is generated based on the target feature vector and sent to the vehicle terminal. The data acquisition task instructs the vehicle terminal to acquire target image data corresponding to the target feature vector. By selecting the most representative candidate feature vectors of the target image data as the target feature vectors, and generating data acquisition tasks based on the target feature vectors, the vehicle-side is instructed to collect real-time image data of the vehicle in a targeted manner according to the target feature vectors. This avoids the indiscriminate collection and processing of real-time image data during vehicle driving, which not only reduces the waste of resources in collecting and processing vehicle data, but also allows for the targeted collection of effective real-time image data that is beneficial to the driving assistance of each vehicle. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating a vehicle data processing method applied to the cloud in one embodiment;

[0059] Figure 2 This is a flowchart illustrating a vehicle data processing method applied to the vehicle end in one embodiment;

[0060] Figure 3 This is a flowchart illustrating a vehicle data processing method applied to the cloud in another embodiment;

[0061] Figure 4 This is a structural block diagram of a vehicle data processing device in one embodiment;

[0062] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] As described in the background section, the vehicle data processing methods of related technologies suffer from high data redundancy, leading to resource waste. The inventors have discovered that this problem arises because, with the development of intelligent driving technology, deep learning models driven by massive amounts of vehicle driving data have become the foundation for core aspects of intelligent driving processes such as perception, decision-making, and control. The performance improvement of these deep learning models is highly dependent on the scale, quality, and diversity of the training data. However, the vehicle data acquisition and processing methods of related technologies still have certain bottlenecks in efficiency and effectiveness, thus restricting the iteration speed of deep learning models and the overall evolution of intelligent driving systems. Specifically, traditional vehicle data acquisition methods mainly adopt a full-data acquisition mode. This mode requires the vehicle to completely collect all sensor data generated during vehicle driving (such as images, LiDAR point clouds, millimeter-wave radar signals, and vehicle CAN (Controller Area Network) bus information) and upload it to the cloud. While the full-data acquisition mode has the advantages of simplicity, ensuring no data omissions, and theoretically covering all data related to vehicle operation scenarios, its limitations are significant. However, the collected data has extremely high redundancy, with over 95% of the data representing simple, repetitive, and routine operating scenarios. The uploading of a large amount of invalid data not only consumes vehicle communication bandwidth but also significantly increases cloud storage and computing resource consumption, raising operating costs and increasing resource waste. Furthermore, the full-data collection model assumes that all vehicle data has equal value for training deep learning models. However, for a high-performance deep learning model, the marginal benefit of information increment from vehicle data generated under common operating scenarios where it performs well is low. For example, vehicle data collected under normal driving conditions has low information increment for deep learning model training, leading to massive amounts of inefficient or even invalid data being uploaded to the cloud. In addition, during the data collection process, the relevant technologies lack understanding of which scenarios the cloud-based deep learning model performs poorly in and which object categories have low recognition rates, resulting in information asymmetry between the vehicle and cloud ends. The vehicle-side system cannot provide targeted and precise training data to supplement the current shortcomings and blind spots of the deep learning model, and can only blindly collect all vehicle data.

[0065] For the reasons mentioned above, this application provides a vehicle data processing method. This method acquires multiple target image data sent from the vehicle and performs a series of processing operations on each target image data to obtain the dimensionality-reduced feature vector and its probability density distribution result corresponding to each target image data. Further, based on the probability density distribution result, a target feature vector is determined from the set of dimensionality-reduced feature vectors, and a data acquisition task is generated based on the target feature vector. This data acquisition task is then sent to the vehicle to instruct the vehicle to collect and process the corresponding target image data as needed according to the target feature vector. This avoids indiscriminate collection and processing of real-time image data during vehicle driving, thereby reducing the waste of resources in collecting and processing vehicle data.

[0066] In one embodiment, such as Figure 1 As shown, a vehicle data processing method is provided. This embodiment illustrates the method by applying it to a cloud server. It is understood that the cloud server is a cloud server that provides cloud computing services and communicates with the vehicle's on-board controller via a network. In this embodiment, the method includes the following steps S102 to S110. Wherein:

[0067] Step S102: Acquire multiple target image data sent by the vehicle terminal, and process each target image data to obtain the dimensionality reduction feature vector corresponding to each target image data.

[0068] The vehicle can be any vehicle that has a communication connection with the cloud server. The relevant vehicle has image data acquisition capabilities, and the relevant vehicle can communicate with the cloud server through its internal vehicle controller.

[0069] The target image data can be a set of image data collected and filtered by the vehicle and corresponding to the training sample data collection requirements of the cloud server. It is used as training sample data for the deep learning model of intelligent driving. Typically, one target image data corresponds to one vehicle driving scenario, and one vehicle driving scenario can correspond to multiple target image data.

[0070] It should be noted that the target image data collected by the vehicle is not limited to the target image data corresponding to the data collection task issued by the cloud. The logic for target image data collection and processing by the vehicle is as follows: for real-time image data during vehicle driving, usability assessment, scene classification, and determination of the similarity between real-time image data and data collection task are performed to obtain usability score and semantic label. If at least one of the usability score, semantic label, and similarity of real-time image data meets the preset data processing triggering condition, the trigger time that meets the preset data processing triggering condition is determined. Based on the trigger time, the real-time image data is cropped and processed to obtain the target image data.

[0071] Optionally, the cloud server acquires multiple target image data sent by the vehicle, and uses a pre-defined deep learning model (such as the large-scale attention visual deep learning model Vision Transformer) as a global feature extractor to perform feature extraction processing on each target image data. The extracted feature vectors are then subjected to dimensionality reduction processing to obtain the dimensionality-reduced feature vectors corresponding to each target image data. The feature extraction processing can involve performing feature extraction operations on the target image data to obtain feature vectors that characterize the image features, typically using convolutional neural networks, handcrafted features, or hybrid methods. The dimensionality-reduced feature vectors can be compact, high-dimensional feature vectors with preserved information after dimensionality reduction processing, improving the efficiency of subsequent statistical analysis and clustering while retaining key semantic information.

[0072] Step S104: Process each dimensionality-reduced eigenvector according to the kernel density estimation algorithm to obtain the probability density distribution of the dimensionality-reduced eigenvector.

[0073] Kernel Density Estimation (KDE) is a non-parametric density estimation method used to estimate the probability density distribution of eigenvectors in the feature space.

[0074] The probability density distribution result can be the set of density distribution values ​​of the dimensionality-reduced eigenvectors in the feature space obtained by the kernel density estimation algorithm, which can reflect the density or rarity of each dimensionality-reduced eigenvector.

[0075] Optionally, the cloud server combines the dimensionality-reduced feature vectors corresponding to each target image data to obtain a feature vector set. Further, the feature vector set is processed according to the kernel density estimation algorithm to obtain the probability density distribution result of the feature vector set. This allows the cloud server to understand the density or rarity of the dimensionality-reduced feature vectors of each target image data in the feature space of the entire feature vector set, laying the groundwork for determining the subsequent acquisition requirements of the target image data.

[0076] Step S106: Based on the probability density distribution results, determine the candidate feature vectors from the dimensionality-reduced feature vectors.

[0077] Among them, the candidate feature vector can be a dimensionality-reduced feature vector that meets the preset conditions and is selected based on the probability density distribution results, and can be used as a candidate object for further clustering and determining the target feature vector.

[0078] Optionally, the cloud server determines candidate feature vectors from the feature vector set based on the density or rarity of each dimensionality-reduced feature vector fed back by the probability density distribution results. It can be understood that the candidate feature vectors can be dimensionality-reduced feature vectors whose density or rarity meets preset conditions, representing that the vehicle driving scene of the target image data corresponding to the candidate feature vector is a relatively rare long-tail scene, which can match the needs of the cloud server to train the deep learning model of intelligent driving.

[0079] Step S108: Cluster the candidate feature vectors to obtain the target feature vector.

[0080] Clustering can be an operation that performs clustering analysis on candidate feature vectors based on the degree of correlation between vehicle driving scenes in target image data. Common clustering algorithms include K-means (a prototype-based, partitioning clustering algorithm), DBSCAN (a density-based clustering algorithm), hierarchical clustering, etc.

[0081] The target feature vector can be a representative feature vector corresponding to the feature vector cluster after clustering of candidate feature vectors, which is used to reflect the data acquisition requirements of the data acquisition task.

[0082] Optionally, the cloud server uses a preset clustering algorithm to cluster the candidate feature vectors, grouping related candidate feature vectors together and identifying the most representative candidate feature vector as the target feature vector. It can be understood that clustering multiple candidate feature vectors is to categorize and summarize the vehicle driving scenes in the target image data corresponding to each candidate feature vector, serving as the basis for subsequent data collection tasks.

[0083] Step S110: Generate a data acquisition task based on the target feature vector and send the data acquisition task to the vehicle terminal.

[0084] The data acquisition task instructs the vehicle-side controller to acquire target image data corresponding to the target feature vector. It should be noted that the vehicle-side controller can maintain a fixed-capacity circular buffer (FIFO, First In First Out) in memory, continuously caching real-time image data, vehicle CAN signal data, and time synchronization information within the most recent preset time period (e.g., 30 seconds). The vehicle-side controller acquires target image data by matching the availability score, semantic label, and similarity of the real-time image data with preset data processing trigger conditions. If at least one of these three factors satisfies the preset data processing trigger conditions, the controller determines the trigger time. Based on this trigger time—for example, using it as the start, end, or center point of a preset time window—the vehicle-side controller extracts real-time image data corresponding to the preset time window from the vehicle-side circular buffer, forming a complete segment of the vehicle driving scene, thus obtaining the target image data.

[0085] Optionally, the cloud server generates a data acquisition task based on the target feature vector and other control parameters (such as task type, execution instructions, etc.) and sends the data acquisition task to the vehicle. Understandably, the cloud server selects the most representative candidate feature vectors of the vehicle driving scenario as the target feature vector, generates a data acquisition task based on the target feature vector, and specifically instructs the vehicle to collect vehicle data, which better meets the training data requirements of deep learning models for intelligent driving.

[0086] In the aforementioned vehicle data processing method, multiple target image data sent by the vehicle are acquired and processed to obtain dimensionality-reduced feature vectors corresponding to each target image data. This improves the efficiency of subsequent statistical and clustering analysis while preserving the key semantic information of the target image data. Further, each dimensionality-reduced feature vector is processed using a kernel density estimation algorithm to obtain its probability density distribution. This allows the cloud server to understand the density or rarity of the dimensionality-reduced feature vectors of each target image data within the feature space of the entire feature vector set. Based on the probability density distribution, candidate feature vectors are determined from the dimensionality-reduced feature vectors. Clustering is then performed on each candidate feature vector to obtain the most representative target feature vector. Finally, a data acquisition task is generated based on the target feature vectors and sent to the vehicle. The data acquisition task instructs the vehicle to acquire target image data corresponding to the target feature vectors. By selecting the most representative candidate feature vectors of the target image data as the target feature vectors, and generating data acquisition tasks based on the target feature vectors, the vehicle-side is instructed to collect real-time image data of the vehicle in a targeted manner according to the target feature vectors. This avoids the indiscriminate collection and processing of real-time image data during vehicle driving, which not only reduces the waste of resources in collecting and processing vehicle data, but also allows for the targeted collection of effective real-time image data that is beneficial to the driving assistance of each vehicle.

[0087] In an exemplary embodiment, step S102 processes each target image data to obtain a dimensionality-reduced feature vector corresponding to each target image data, including:

[0088] Feature extraction is performed on each target image data to obtain the initial feature vector corresponding to each target image data; principal component analysis algorithm is used to reduce the dimensionality of each initial feature vector to obtain multiple dimensionality-reduced feature vectors.

[0089] Principal Component Analysis (PCA) is a data dimensionality reduction method that can transform high-dimensional data into low-dimensional data while preserving as much information as possible from the original data.

[0090] The initial feature vector can be the feature extraction result of the target image data. It is a mathematical vector representing the target image data of an object. Each element (i.e., each dimension) in the initial feature vector represents a certain feature or attribute of the object and serves as the input for subsequent dimensionality reduction operations.

[0091] Optionally, the cloud server uses a preset deep learning model as a global feature extractor, inputs the target image data and historical data from the global feature extractor's basic training library into the global feature extractor, and performs high-dimensional feature extraction through the network of the global feature extractor to obtain the initial feature vector corresponding to each target image data; further, the cloud server uses principal component analysis algorithm to perform dimensionality reduction processing on each initial feature vector to obtain multiple dimensionality-reduced feature vectors.

[0092] In this embodiment, a preset deep learning model is used as a global feature extractor to perform unified, high-dimensional feature extraction on the target image data, generating a representative and widely covered initial feature vector. These initial feature vectors are then subjected to dimensionality reduction processing through principal component analysis to obtain dimensionality-reduced feature vectors. This not only preserves the key semantic information of the target image data but also reduces subsequent storage and computation costs, facilitating subsequent clustering and statistical analysis, thereby improving the efficiency of data processing.

[0093] In an exemplary embodiment, step S104 processes each dimensionality-reduced feature vector according to the kernel density estimation algorithm to obtain the probability density distribution result of the dimensionality-reduced feature vector, including:

[0094] Based on the kernel density estimation algorithm, the density value of each dimensionality-reduced eigenvector is calculated; based on the density value of each dimensionality-reduced eigenvector, the probability density distribution of the dimensionality-reduced eigenvector is determined.

[0095] The density value can be the probability density value calculated by the kernel density estimation algorithm for each dimensionality-reduced feature vector, representing its rarity or local density in the feature space.

[0096] Optionally, the cloud server calculates the density values ​​of each dimensionality-reduced eigenvector based on the kernel density estimation algorithm. Specifically, it uses a Gaussian kernel function and determines the optimal bandwidth h using the Scott rule. For any dimensionality-reduced eigenvector z in the feature space, its density value p(z) is calculated as follows:

[0097] Where N is the sample size. Let z be the displacement vector between point z and the i-th data point. To calculate the Euclidean norm of the standardized displacement vector, according to the definition, the density value of each dimensionality-reduced eigenvector in the feature set can be calculated. Furthermore, the cloud server sorts the density values ​​of each dimensionality-reduced eigenvector in a certain order, such as in ascending order or descending order, to obtain the probability density distribution result of the feature vector set.

[0098] In this embodiment, a kernel density estimation algorithm is used to process the feature vector set, achieving local density estimation of each dimensionality-reduced feature vector in the feature space. This yields the probability density distribution of the feature vector set. Dimensionality-reduced feature vectors with low density values ​​often represent sparse feature vectors with high potential information content, while high-density feature vectors may be redundant or common. The probability density distribution obtained after sorting the density values ​​provides a basis for subsequent candidate feature vector selection, enhances the diversity and information value of vehicle data sampling, reduces data redundancy and transmission costs, and further reduces resource waste in collecting vehicle data.

[0099] In an exemplary embodiment, step S106, based on the probability density distribution results, determines candidate feature vectors from the dimensionality-reduced feature vectors, including:

[0100] Based on the probability density distribution results and their corresponding preset quantile values, a low density threshold is determined; based on the low density threshold, a target density value with a density value less than the low density threshold is determined from the density values ​​of each dimensionality-reduced feature vector; the dimensionality-reduced feature vector corresponding to the target density value is determined as a candidate feature vector.

[0101] Quantiles are a concept in statistics used to describe the value of a set of data at a specific position or proportion. Quantiles divide a dataset into several equal parts after it is arranged in order of size, and the corresponding dividing points are the quantile values. Preset quantile values ​​can be set according to the actual density value filtering needs.

[0102] The low density threshold can be the density value of the dimensionality-reduced feature vector corresponding to a preset quantile value. Dimensionality-reduced feature vectors below this threshold are considered as candidate feature vectors.

[0103] Optionally, the cloud server determines the density value of the dimensionality-reduced feature vector corresponding to the preset quantile value from the probability density distribution results, and sets this density value as the low-density threshold. Further, based on the low-density threshold, the cloud server determines the target density value from the density values ​​of each dimensionality-reduced feature vector, which is less than the low-density threshold, and sets the dimensionality-reduced feature vector corresponding to the target density value as a candidate feature vector. Specifically, the value of the α-th quantile (i.e., the preset quantile value) is taken as the low-density threshold τ. For example, if α is set to 5%, all density values ​​satisfy p(f i The reduced feature vectors of τ are labeled as candidate feature vectors, representing that the corresponding vehicle scenarios are relatively rare. The cloud server needs to collect the target image data corresponding to the candidate feature vectors.

[0104] In this embodiment, a low-density threshold is determined from the probability density distribution results using preset quantile values. Based on the density values ​​of each dimensionality-reduced feature vector and the low-density threshold, a selection process is performed to identify candidate feature vectors that are relatively rare in the feature space, have high information value, and may represent edge vehicle driving scenarios. This allows for targeted feedback of the vehicle data collection needs of the cloud server and also helps the vehicle to understand the vehicle data collection needs, collect vehicle data on demand, avoid the collection of redundant data, and further reduce the waste of resources in collecting and processing vehicle data.

[0105] In an exemplary embodiment, step S108 performs clustering processing on each candidate feature vector to obtain the target feature vector, including:

[0106] Cluster all candidate feature vectors to obtain multiple feature vector clusters; the candidate feature vectors corresponding to the cluster centers of the feature vector clusters are determined as the target feature vectors.

[0107] Among them, the feature vector cluster can be a set of candidate feature vectors formed based on the clustering results. Each cluster contains a set of similar candidate feature vectors, and the candidate feature vectors within the cluster have high similarity.

[0108] The cluster center can be the central eigenvector of each eigenvector cluster, usually the mean or the most representative eigenvector in the cluster, used as the representative feature of the cluster.

[0109] Optionally, the cloud server can select an appropriate clustering algorithm based on the actual clustering scenario to cluster all candidate feature vectors, grouping similar candidate feature vectors into the same feature vector cluster, thus obtaining multiple feature vector clusters. Further, the cloud server determines the candidate feature vector corresponding to the cluster center of the feature vector cluster as the target feature vector. It can be understood that more than one target feature vector can be selected in each feature vector cluster; for example, when determining the cluster center, the mean method is used. In the same feature vector cluster, if there are multiple candidate feature vectors whose mean value is close to that of the cluster's feature vectors, then all of them are taken as target feature vectors.

[0110] In this embodiment, by clustering the candidate feature vectors, multiple feature vector clusters are obtained, and the cluster center of each feature vector cluster is determined as the target feature vector. This can further filter redundant data in the target image data corresponding to the candidate feature vectors, thereby further reducing the waste of resources in vehicle data processing.

[0111] In an exemplary embodiment, step S110 involves sending the data acquisition task to the vehicle, including:

[0112] Based on the target feature vector, determine the target scene corresponding to the target feature vector; based on the collected status data of multiple vehicles, determine the target vehicle associated with the target scene; and send the data collection task to the vehicle end of the target vehicle.

[0113] The target scene can be the type of vehicle driving scenario that corresponds to the target image data corresponding to the target feature vector when it is collected by the vehicle.

[0114] Among them, vehicle status data can be a collection of status information of the vehicle at present or within a specific period of time, which usually includes sensor data, geographical location, environmental parameters, etc.

[0115] The target vehicle can be one of multiple vehicles with image data acquisition capabilities connected to the cloud server, and the vehicle currently in a driving scenario similar to the target scenario; at least two vehicles with image data acquisition capabilities connected to the cloud server constitute a convoy.

[0116] Optionally, the cloud server determines the vehicle driving scenario when the target image data corresponding to the target feature vector is collected, based on the target feature vector, and uses this as the target scenario. Further, the cloud server obtains the status data of multiple vehicles based on real-time fleet status information, and infers the driving scenario of each vehicle based on the status data of each vehicle, identifying the target vehicle whose driving environment is associated with the target scenario, and thus sending the data collection task to the target vehicle's onboard unit. For example, if the target scenario is a mountainous driving scenario in rain or snow, then the vehicle's geographical location (mountainous), sensor data showing low temperature and high humidity, and weather information indicating rain or snow are used to identify the vehicle as the target vehicle, and the data collection task is sent to the target vehicle's onboard unit.

[0117] In this embodiment, by acquiring the target image data corresponding to the target feature vector when it is collected by the vehicle, and based on the fleet status information, the target vehicle whose driving environment is associated with the target scene is determined, and the data collection task is sent to the vehicle terminal of the target vehicle, thereby increasing the probability of the vehicle terminal collecting the target image data and improving the accuracy of the data collection task issuance.

[0118] In one exemplary embodiment, the vehicle data processing method described above further includes:

[0119] Obtain the availability score corresponding to each target image data sent by the vehicle terminal; perform statistical processing on each availability score to obtain the availability score threshold; and send the availability score threshold to the vehicle terminal.

[0120] Among them, the availability score can be a numerical indicator used to measure the availability or value of real-time image data of a vehicle driving scenario in data acquisition or processing tasks.

[0121] The availability score threshold is used to instruct the vehicle to process real-time image data with an availability score lower than the availability score threshold in order to obtain the target image data.

[0122] Statistical processing can be a process of descriptive or inferential analysis of a dataset, including the calculation and summarization of statistics such as density, distribution, mean, variance, and distribution shape.

[0123] Understandably, the vehicle controller acquires vehicle attribute data, including vehicle ID, GPS (Global Positioning System) coordinates, timestamps, etc., and packages, encapsulates, and compresses the vehicle attribute data, target image data, availability score, and semantic tags to obtain the target data packet corresponding to the target image data. If the network conditions meet the upload requirements, the target data packet is sent to the cloud.

[0124] Optionally, the cloud server obtains the availability score corresponding to each target image data sent by the vehicle controller; performs statistical processing on each availability score to obtain an availability score threshold. For example, the cloud server calculates the mean, median, or mode of each availability score and uses it as the availability score threshold. Further, the cloud server sends the availability score threshold to the vehicle.

[0125] It should be noted that the vehicle-side controller uses a preset Gaussian mixture model to process real-time vehicle image data. The calculated log-likelihood value is used as the availability score of the real-time image data, which serves as the basis for evaluating whether the real-time image data needs to be processed, i.e., as an indicator for filtering real-time image data. The vehicle-side controller compares the availability score with a preset score threshold to obtain the score comparison result. If the availability score is less than or equal to the preset score threshold, the real-time image data representing the current vehicle driving scenario has low value and needs to be filtered out; if the availability score is greater than the preset score threshold, the real-time image data representing the current vehicle driving scenario has high value and needs to be acquired and processed to obtain target image data.

[0126] In this embodiment, by statistically analyzing the availability score of the target image data transmitted from the vehicle, an availability score threshold is automatically generated, which improves the scientific validity and rationality of the availability score threshold setting. After the availability score threshold is sent to the vehicle, the vehicle can use it to filter and process the real-time collected data, prioritizing the retention of high-availability real-time image data, discarding low-quality real-time image data, or reducing the transmission of low-quality real-time image data, thereby reducing transmission and storage costs and improving the overall quality and reliability of the training sample data on the cloud server.

[0127] Based on the same inventive concept, in an exemplary embodiment, the vehicle data processing method provided in this application can also be implemented through the interaction between the vehicle-side controller and the cloud, wherein the vehicle-side data acquisition module is deployed on the vehicle-side controller, and the cloud-side task management module is deployed on the cloud server; wherein, the vehicle-side data acquisition module is used to perform real-time value assessment on the sensor data generated in real time, and make an upload decision based on the value score of the data.

[0128] The core of real-time assessment and triggering of vehicle-side data value lies in enabling immediate judgment of data value on resource-constrained vehicle-side devices, encompassing four key stages: feature extraction and processing, score calculation, intelligent decision-making, and data uploading.

[0129] like Figure 2 As shown, this application also provides a flowchart of a vehicle data processing method applied to a vehicle-side controller, which includes steps 11 to 14, wherein:

[0130] Step 11, Feature Extraction Processing: On the vehicle-mounted embedded computing unit, a deeply optimized lightweight convolutional neural network, MobileNetV3, is deployed. Its classification head is removed and transformed into a feature embedding layer. For each frame of image (real-time image data) captured by the vehicle's forward-facing camera, it is scaled to the size required by the model and input into the lightweight model for forward inference. After passing through the model, a fixed-dimensional first feature vector (driving scene feature vector) is obtained. This vector is a dense, low-dimensional semantic representation of the current image scene (vehicle driving scene).

[0131] Step 12, Availability Score Calculation: The cloud periodically sends out parameters for a probability distribution model representing a "normal driving scenario." This model is preferably a Gaussian Mixture Model (GMM), whose parameters include the weights of each Gaussian component, the mean vector, and the covariance matrix. After receiving the model parameters, the vehicle calculates the log-likelihood value of each feature vector under the GMM, which is used as the availability score. The corresponding calculation expression is:

[0132]

[0133] Where S is the log-likelihood value (i.e., the usability score), π j This represents the mixing weight of the j-th Gaussian component, where μ is the mean of the Gaussian component. j The value of the covariance Σj is preset and determined based on the parameters of the probability distribution model representing "conventional vehicle driving scenarios" periodically distributed from the cloud, N(vt|μ j ,Σj) represents the sum of the mean values ​​of the Gaussian components μ j Given the covariance Σj, the driving scene feature vector v corresponding to the observed real-time image data is... t The probability density, N is the Gaussian component pair v t The degree of influence. It is understandable that the higher the usability score S, the more the current vehicle driving scenario conforms to common driving patterns, and the lower the value of its corresponding real-time image data; conversely, the lower the usability score S, the more "unfamiliar" or "abnormal" the current vehicle driving scenario, the higher the value of its corresponding real-time image data, and the higher the potential value for the iteration of deep learning models for intelligent driving, which can promote the improvement of the robustness and generalization ability of deep learning models for intelligent driving in iterative training.

[0134] Step 13: Intelligent triggering decision-making based on multi-factor fusion. The vehicle maintains a decision engine whose input includes at least three aspects. Vehicle driving data is collected when a trigger condition satisfies one or a combination of the following logic: Trigger condition 1: The availability score S is lower than a dynamically adjusted score threshold θ (availability score threshold); Trigger condition 2: A parallel-running lightweight scene classifier detects that the current frame contains a preset key semantic label. This label represents a pre-set special scene, such as heavy rain, dense fog, nighttime backlight, animals, irregularly shaped vehicles, or construction zones. Once detected, uploading is immediately triggered; Trigger condition 3: An active collection task is received from the cloud, containing a target feature vector. The vehicle calculates the cosine similarity between the feature vector and the target feature vector. If the similarity exceeds a matching threshold (preset similarity threshold), uploading is triggered.

[0135] Step 14, Scene Fragment Encapsulation and Upload: The vehicle-side system maintains a fixed-capacity circular buffer in memory, continuously storing the most recent 30 seconds of raw sensor data (such as images, LiDAR (Light Detection and Ranging) point clouds), vehicle CAN signals, and time synchronization information. When a trigger signal is generated, the system immediately extracts a time window of data centered on the trigger moment from the buffer, forming a complete scene fragment (target image data). Subsequently, this scene fragment, trigger cause metadata, availability score, GPS (Global Positioning System) coordinates, timestamp, vehicle ID (Identifier), and other information are encapsulated into a compressed data packet (target data packet), and asynchronously uploaded to the cloud server's data center when network conditions permit.

[0136] The cloud-based global strategic analysis and task management module is responsible for aggregating and analyzing data across the entire domain, identifying blind spots in data distribution, and generating strategic proactive data collection tasks (data collection tasks) to be sent to the vehicles. Its core function is to analyze the completeness of the global data distribution and proactively guide the fleet to "fill in the gaps".

[0137] like Figure 3 As shown, this application also provides a flowchart of a vehicle data processing method applied to the cloud, covering feature extraction processing, coverage analysis and task distribution, etc. The vehicle data processing method includes steps 21 to 23.

[0138] Step 21, Big Data Feature Embedding: The cloud uses a large-scale deep learning model, Vision Transformer (an attention-based image classification model), as the global feature extraction processor. The input consists of keyframes from all "scene clips" uploaded by the vehicles, as well as historical data from the basic training library. For each input image, the network extracts its high-dimensional second feature vector (initial feature vector), and all feature vectors are combined to form a global feature vector set, i.e., the corresponding... Figure 3 The text discusses "integrating full-domain data" and "feature extraction and aggregation to obtain high-dimensional features using large models."

[0139] Step 22, Coverage Quantization and Low-Density Region Identification: First, Principal Component Analysis (PCA) is used to reduce the dimensionality of the high-dimensional feature set, resulting in a low-dimensional feature set corresponding to the reduced feature vectors. Kernel Density Estimation (KDE) is then used to model the probability density of the low-dimensional feature set (probability density distribution results). The density value of each sample point in the low-dimensional feature set can be calculated, all density values ​​are sorted, and the α-th quantile is taken as the low-density threshold τ. In this embodiment, α is set to 5%. All regions satisfying p(f i Sample points with a value less than τ are labeled as blind zone samples. The original high-dimensional feature vectors corresponding to these samples are extracted to form an uncovered feature vector set (candidate feature vector set), which corresponds to... Figure 3 The text discusses "identifying data blind spots, using dimensionality reduction and density estimation to find data blind spots."

[0140] Step 23, Active Data Acquisition Task Generation and Distribution: Cluster analysis is performed on the uncovered feature vector set U to remove redundancy, grouping similar blind spot scenarios into one category. From each cluster (feature vector cluster), one or more core sample feature vectors are selected as "prototype feature vectors" (target feature vectors). Each prototype vector generates an active data acquisition task. Based on real-time fleet status information (such as geographical location and weather), the cloud platform accurately pushes the acquisition task to the target vehicle most likely to encounter the scenario (target scenario), thus distributing the task. This enables the vehicle to accurately identify and collect data on the target scenario according to the acquisition task, i.e., corresponding to... Figure 3 The tasks include "clustering blind spot scenes and generating tasks" and "executing data collection tasks on the vehicle side to accurately identify and collect data from target scenes".

[0141] In this embodiment, by performing value pre-screening on the vehicle side, only data identified as high-value is uploaded, while redundant and repetitive data is filtered out. This significantly reduces data transmission costs and cloud resource consumption, greatly reducing the cost of storing invalid data and preprocessing data in the cloud. For large-scale fleets, the cost-saving effect is extremely significant, reducing the operational costs of data closure. This embodiment makes judgments based on deep learning feature distribution, which can more directly and keenly identify rare, unexpected, and edge scenarios that have not yet been fully learned by the model, such as rare obstacles and complex traffic participant interactions. This makes the collected data more valuable, greatly helping to solve the long-tail problem of autonomous driving and significantly improving the efficiency of constructing high-value data. The method of this embodiment judges based on high-dimensional semantic features learned by large cloud models, avoiding the one-sidedness and high false alarm rate of traditional rule-based methods. The evaluation results are more intelligent, accurate, and robust. At the same time, the vehicle side adopts a lightweight model and simple difference calculation, reducing computational overhead and latency. It can run in real time on resource-constrained automotive-grade chips without affecting the main autonomous driving functions of the vehicle as much as possible, reducing vehicle-side resource consumption while ensuring perception accuracy. This changes the way autonomous driving data is collected, specifically from "blindly uploading all" to "intelligently selectively collecting". This not only saves economic costs, but also accelerates the iteration efficiency of the autonomous driving system's ability to deal with corner cases.

[0142] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0143] Based on the same inventive concept, this application also provides a vehicle data processing apparatus for implementing the vehicle data processing method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more vehicle data processing apparatus embodiments provided below can be found in the limitations of the vehicle data processing method described above, and will not be repeated here.

[0144] In one exemplary embodiment, such as Figure 4As shown, a vehicle data processing device 400 is provided, including: a data acquisition module 401, a density modeling module 402, a feature determination module 403, a feature processing module 404, and a task generation module 405, wherein:

[0145] The data acquisition module 401 is used to acquire multiple target image data sent by the vehicle terminal, and process each target image data to obtain the dimension-reduced feature vector corresponding to each target image data.

[0146] The density modeling module 402 is used to process each dimension-reduced feature vector according to the kernel density estimation algorithm to obtain the probability density distribution result of the dimension-reduced feature vector;

[0147] The feature determination module 403 is used to determine candidate feature vectors from the dimensionality-reduced feature vectors based on the probability density distribution results.

[0148] The feature processing module 404 is used to perform clustering processing on each candidate feature vector to obtain the target feature vector;

[0149] The task generation module 405 is used to generate a data acquisition task based on the target feature vector and send the data acquisition task to the vehicle terminal; the data acquisition task is used to instruct the vehicle terminal to acquire target image data corresponding to the target feature vector.

[0150] Furthermore, in one embodiment, the data acquisition module 401 is also used to extract features from each target image data to obtain an initial feature vector corresponding to each target image data; and to perform dimensionality reduction processing on each initial feature vector using a principal component analysis algorithm to obtain multiple dimensionality-reduced feature vectors.

[0151] Furthermore, in one embodiment, the density modeling module 402 is also used to calculate the density value of each dimension-reduced feature vector according to the kernel density estimation algorithm; and to determine the probability density distribution result of the dimension-reduced feature vector according to the density value of each dimension-reduced feature vector.

[0152] Furthermore, in one embodiment, the feature determination module 403 is also used to determine a low density threshold based on the probability density distribution result and its corresponding preset quantile value; based on the low density threshold, determine a target density value whose density value is less than the low density threshold from the density values ​​of each dimensionality reduction feature vector; and determine the dimensionality reduction feature vector corresponding to the target density value as a candidate feature vector.

[0153] Furthermore, in one embodiment, the feature processing module 404 is also used to perform clustering processing on all candidate feature vectors to obtain multiple feature vector clusters; and to determine the candidate feature vector corresponding to the cluster center of the feature vector cluster as the target feature vector.

[0154] Furthermore, in one embodiment, the task generation module 405 is also used to determine the target scene corresponding to the target feature vector based on the target feature vector; determine the target vehicle associated with the target scene based on the collected state data of multiple vehicles; and send the data collection task to the vehicle end of the target vehicle.

[0155] Furthermore, in one embodiment, the vehicle data processing device 400 further includes a threshold construction module, used to obtain the availability score corresponding to each target image data sent by the vehicle terminal; perform statistical processing on each availability score to obtain an availability score threshold; and send the availability score threshold to the vehicle terminal; the availability score threshold is used to instruct the vehicle terminal to process real-time image data with availability scores less than the availability score threshold in order to obtain target image data.

[0156] Each module in the aforementioned vehicle data processing device 400 can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0157] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores target image data, dimensionality-reduced feature vectors, probability density distribution results, target feature vectors, and data acquisition task data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a vehicle data processing method.

[0158] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0159] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0160] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0161] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0162] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0163] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0164] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A vehicle data processing method, characterized in that, The method includes: Multiple target image data sent by the vehicle are acquired, and each target image data is processed to obtain a dimensionality-reduced feature vector corresponding to each target image data. The target image data is obtained by the vehicle terminal after determining that the availability score of the acquired real-time image data meets a preset data processing trigger condition, and determining the trigger time that meets the preset data processing trigger condition, based on the trigger time of the real-time image data being truncated. The vehicle terminal uses a preset Gaussian mixture model to process the real-time image data of the vehicle, and uses the calculated log-likelihood value as the availability score of the real-time image data. If the availability score of the real-time image data is less than a preset score threshold, it indicates that the number of real-time images in the current vehicle driving scenario needs to be acquired and processed to obtain the value of the target image data. The calculation expression of the availability score is: Where S is the availability score, π j μ represents the mixing weight of the j-th Gaussian component. j Let N(vt|μ) be the Gaussian component mean, Σj be the covariance, and N(vt|μ) be the mean of the components. j ,Σj) represents the sum of the mean values ​​of the Gaussian components μ. j Given the covariance Σj, the driving scene feature vector v corresponding to the observed real-time image data is... t The probability density, N is the Gaussian component pair v t The degree of influence; the mixed weight, the mean of the Gaussian components, and the covariance are periodically sent from the cloud to the vehicle terminal; According to the kernel density estimation algorithm, the density value of each of the dimensionality-reduced feature vectors is calculated; based on the density value of each of the dimensionality-reduced feature vectors, the probability density distribution result of the dimensionality-reduced feature vectors is determined; the density value represents the rarity or local density of each of the dimensionality-reduced feature vectors in the feature space. Based on the low density threshold determined by the probability density distribution results, candidate feature vectors with density values ​​less than the low density threshold are determined from the dimensionality-reduced feature vectors; the vehicle driving scene in the image data corresponding to the candidate features is a long-tail scene, which can match the requirements of training deep learning models for intelligent driving. Clustering is performed on each of the candidate feature vectors to obtain the target feature vector; A data acquisition task is generated based on the target feature vector, and the target scene corresponding to the target feature vector is determined based on the target feature vector. Based on the collected state data of multiple vehicles, the driving environment of each vehicle is determined, and the target vehicle associated with the driving environment of the vehicle and the target scene is identified. The data acquisition task is sent to the vehicle's terminal. The vehicle's state data includes sensor data, geographical location, and environmental parameters. The data acquisition task instructs the vehicle's terminal to acquire target image data corresponding to the target feature vector.

2. The method according to claim 1, characterized in that, The target image data is processed to obtain a dimensionality-reduced feature vector corresponding to each target image data, including: Feature extraction is performed on each of the target image data to obtain an initial feature vector corresponding to each of the target image data; Principal component analysis (PCA) algorithm is used to reduce the dimensionality of each initial eigenvector, resulting in multiple dimensionality-reduced eigenvectors.

3. The method according to claim 1, characterized in that, Based on the low-density threshold determined by the probability density distribution results, candidate feature vectors with density values ​​less than the low-density threshold are determined from the dimensionality-reduced feature vectors, including: The low density threshold is determined based on the probability density distribution results and their corresponding preset quantile values. Based on the low density threshold, a target density value that is less than the low density threshold is determined from the density values ​​of each of the dimensionality-reduced feature vectors. The dimensionality-reduced feature vector corresponding to the target density value is determined as the candidate feature vector.

4. The method according to claim 1, characterized in that, Clustering is performed on each of the candidate feature vectors to obtain the target feature vector, including: Clustering is performed on all the candidate feature vectors to obtain multiple feature vector clusters; The candidate feature vector corresponding to the cluster center of the feature vector cluster is determined as the target feature vector.

5. The method according to claim 1, characterized in that, The method further includes: Obtain the availability score corresponding to each of the target image data sent by the vehicle terminal; Statistical processing is performed on each of the aforementioned availability scores to obtain the availability score threshold; The availability score threshold is sent to the vehicle terminal; the availability score threshold is used to instruct the vehicle terminal to process real-time image data with an availability score less than the availability score threshold in order to obtain target image data.

6. A vehicle data processing device, characterized in that, The device includes: The data acquisition module is used to acquire multiple target image data sent by the vehicle terminal, and process each target image data to obtain a dimensionality-reduced feature vector corresponding to each target image data. The target image data is obtained by the vehicle terminal after determining that the availability score of the acquired real-time image data meets a preset data processing trigger condition, and determining the trigger time that meets the preset data processing trigger condition, based on the trigger time of the real-time image data being truncated. The vehicle terminal is used to process the real-time image data of the vehicle using a preset Gaussian mixture model, and uses the calculated log-likelihood value as the availability score of the real-time image data. If the availability score of the real-time image data is less than a preset score threshold, it indicates that the number of real-time images in the current vehicle driving scenario is in a state that needs to be acquired and processed to obtain the value of the target image data. The calculation expression of the availability score is: Where S is the availability score, π j μ represents the mixing weight of the j-th Gaussian component. j Let N(vt|μ) be the Gaussian component mean, Σj be the covariance, and N(vt|μ) be the mean of the components. j ,Σj) represents the sum of the mean values ​​of the Gaussian components μ. j Given the covariance Σj, the driving scene feature vector v corresponding to the observed real-time image data is... t The probability density, N is the Gaussian component pair v t The degree of influence; the mixed weight, the mean of the Gaussian components, and the covariance are periodically sent from the cloud to the vehicle terminal; The density modeling module calculates the density value of each of the dimensionality-reduced feature vectors according to the kernel density estimation algorithm; and determines the probability density distribution result of the dimensionality-reduced feature vectors based on the density value of each of the dimensionality-reduced feature vectors; the density value represents the rarity or local density of each of the dimensionality-reduced feature vectors in the feature space. The feature determination module is used to determine candidate feature vectors whose density values ​​are less than the low density threshold from the dimensionality-reduced feature vectors based on the low density threshold determined by the probability density distribution results; the vehicle driving scene in the image data corresponding to the candidate features is a long-tail scene, which can match the requirements of training deep learning models for intelligent driving. The feature processing module is used to perform clustering processing on each of the candidate feature vectors to obtain the target feature vector; the vehicle driving scene of the image data corresponding to the candidate features is a long-tail scene, which can match the requirements of training deep learning models for intelligent driving. The task generation module is used to generate a data acquisition task based on the target feature vector, and to determine the target scene corresponding to the target feature vector based on the target feature vector; to determine the driving environment of each vehicle based on the collected status data of multiple vehicles, and to determine the target vehicle associated with the driving environment of the vehicle and the target scene; and to send the data acquisition task to the vehicle end of the target vehicle; the vehicle status data includes sensor data, geographical location, and environmental parameters; the data acquisition task is used to instruct the vehicle end to acquire target image data corresponding to the target feature vector.

7. The apparatus according to claim 6, characterized in that, The data acquisition module is further configured to extract features from each of the target image data to obtain an initial feature vector corresponding to each of the target image data; and to perform dimensionality reduction processing on each of the initial feature vectors using a principal component analysis algorithm to obtain multiple dimensionality-reduced feature vectors.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images