Driving trajectory data mining method, device and medium

Abstract

Embodiments of the present application provide a driving trajectory data mining method, device and medium. The method comprises: obtaining a to-be-mined data set according to original natural driving log data; the to-be-mined data set comprises a plurality of data segments; each data segment comprises a trajectory position sequence and a plurality of dimension driving state sequences; for each data segment, determining a trajectory feature vector of the data segment according to the plurality of dimension driving state sequences in the data segment; and performing effective segment screening on the to-be-mined data set based on the trajectory feature vector of each data segment and an information theory-based pruning algorithm to obtain a target mining data set. The method is used to mine driving data with high information quantity, so as to meet the demand of automatic driving model training for high-quality trajectory data.

Description

Technical Field

[0001] This application relates to the field of autonomous driving training data processing, and in particular to a method, device and medium for mining driving trajectory data. Background Technology

[0002] End-to-end intelligent driving systems are driven by big data. They input sensor perception results into learning models such as deep neural networks, and automatically plan safe paths based on the output results to guide and control the vehicle to achieve autonomous driving.

[0003] However, natural driving logs collected from real vehicles generally suffer from a contradiction between data scale and data quality: the total amount of raw logs collected from the whole vehicle is huge, but the proportion of high-quality driving segments with exemplary value and high annotation quality is extremely low.

[0004] Against this backdrop, how to efficiently filter out high-quality driving data subsets with high information content from large-scale natural driving logs for use in training autonomous driving models is a key technical challenge that urgently needs to be addressed. Summary of the Invention

[0005] This application provides a driving trajectory data mining method, device, and medium to mine driving data with high information content, so as to meet the needs of autonomous driving model training for high-quality trajectory data.

[0006] In a first aspect, embodiments of this application provide a method for mining driving trajectory data, including:

[0007] Based on the original natural driving log data, a dataset to be mined is obtained; the dataset to be mined includes multiple data segments; each data segment includes a trajectory location sequence and a driving state sequence in multiple dimensions;

[0008] For each data segment, a trajectory feature vector is determined based on the driving state sequence of multiple dimensions in the data segment;

[0009] Based on the trajectory feature vector of each data segment and an information-theoretic pruning algorithm, the dataset to be mined is filtered for effective segments to obtain the target mining dataset.

[0010] In one possible implementation, determining the trajectory feature vector of the data segment based on a multi-dimensional sequence of driving states in the data segment includes:

[0011] For each dimension of the driving state sequence, feature values ​​of the data segment in that dimension are extracted based on the driving state sequence;

[0012] Based on the feature values ​​of the data segment in different dimensions, the trajectory feature vector of the data segment is constructed.

[0013] In one possible implementation, the dataset to be mined includes sub-datasets corresponding to different scene events; the effective fragment selection of the dataset to be mined is performed based on the trajectory feature vector of each data fragment and an information-theoretic pruning algorithm to obtain the target mining dataset, including:

[0014] For each scene event, if the number of data fragments in the subset corresponding to the scene event is greater than the preset quota number corresponding to the scene event, then an information theory-based pruning algorithm is executed on multiple data fragments corresponding to the scene event until the number of data fragments is equal to the preset quota number, thereby obtaining multiple target data fragments corresponding to the scene data.

[0015] If the number of data segments corresponding to the scene event is less than or equal to the preset quota number corresponding to the scene event, then each data segment corresponding to the scene event is determined as a target data segment.

[0016] The target mining dataset is obtained by summarizing the target data fragments corresponding to events in different scenarios.

[0017] In one possible implementation, performing an information-theoretic pruning algorithm on multiple data fragments corresponding to the scene event until the number of data fragments equals the preset quota includes:

[0018] Step a, construct the algorithm iteration set, and initialize the algorithm iteration set to an empty set;

[0019] Step b: In the first iteration, for each data segment in the subset, the distribution deviation between the data segment and the subset is calculated based on the trajectory feature vector of the data segment; the data segment with the smallest distribution deviation in the subset is transferred to the algorithm iteration set and enters the next iteration.

[0020] Step c: For each remaining data segment in the subset, based on the trajectory feature vector of the data segment, calculate the information entropy increment and distribution deviation increment brought about by adding the data segment to the algorithm iteration set, and calculate the first score of the data segment based on the information entropy increment and distribution deviation increment;

[0021] Step d: Transfer the top preset number of data segments with the highest scores in the subset to the algorithm iteration set, and proceed to the next iteration;

[0022] Step e, repeat steps c and d until the number of fragments in the algorithm iteration set equals the preset quota.

[0023] In one possible implementation, the method for calculating the information entropy of any non-empty dataset in the information-theory-based pruning algorithm includes:

[0024] The multiple feature dimensions corresponding to the non-empty dataset are discretized into multiple multi-dimensional joint intervals;

[0025] For each multi-dimensional joint interval, the number of data segments in the non-empty dataset whose trajectory feature vectors fall into the multi-dimensional joint interval is counted, and the empirical frequency of the non-empty dataset in the multi-dimensional joint interval is determined based on the number of segments.

[0026] The information entropy of the non-empty dataset is calculated based on the empirical frequency corresponding to each multi-dimensional joint interval.

[0027] In one possible implementation, before performing effective fragment filtering on the multiple data fragments based on the trajectory feature vector of each data fragment and an information-theoretic pruning algorithm to obtain the target mining dataset, the method further includes:

[0028] For any two data segments in the dataset to be mined, if the cosine similarity between the trajectory feature vectors of the two data segments is greater than a preset threshold, then the two data segments are determined to be a redundant pair.

[0029] Redundant data fragments are removed from the redundant pairs in the dataset to be mined.

[0030] In one possible implementation, the method further includes:

[0031] Based on the preset evaluation indicators for each evaluation dimension, a second score is calculated for each data segment in the dataset to be mined; wherein, the evaluation dimension includes at least one of the comfort dimension, task consistency dimension, and data rationality dimension;

[0032] Accordingly, the redundant data fragment removal process for redundant pairs in the dataset to be mined includes:

[0033] The data segment with the lower second score in the redundant pair is removed from the dataset to be mined.

[0034] In one possible implementation, obtaining the dataset to be mined based on the raw natural driving log data includes:

[0035] For each raw data segment in the natural driving log data, determine the scene event corresponding to the raw data segment, and determine the time window of the raw data segment based on the scene event;

[0036] The original data segment is segmented using the time window to obtain the segmented data segment;

[0037] The segmented data fragments belonging to the same scene event are collectively constructed into a subset of the corresponding scene event, and multiple subsets are aggregated to obtain the dataset to be mined.

[0038] In one possible implementation, the method further includes:

[0039] Remove the original data segments in the natural driving log data that do not meet the preset hard constraints;

[0040] The hard constraints include at least one of the following: hard constraints in the security dimension, hard constraints in the compliance dimension, and hard constraints in the availability dimension of the positioning module or sensor fusion module.

[0041] Secondly, embodiments of this application provide a driving trajectory data mining device, comprising:

[0042] The acquisition module is used to acquire a dataset to be mined based on the original natural driving log data; the dataset to be mined includes multiple data segments; each data segment includes a trajectory location sequence and a driving state sequence in multiple dimensions;

[0043] The determination module is used to determine the trajectory feature vector of each data segment based on the driving state sequence of multiple dimensions in the data segment;

[0044] The processing module is used to filter effective segments of the dataset to be mined based on the trajectory feature vector of each data segment and an information theory-based pruning algorithm, so as to obtain the target mining dataset.

[0045] In one possible implementation, the determining module is specifically used for:

[0046] For each dimension of the driving state sequence, feature values ​​of the data segment in that dimension are extracted based on the driving state sequence;

[0047] Based on the feature values ​​of the data segment in different dimensions, the trajectory feature vector of the data segment is constructed.

[0048] In one possible implementation, the dataset to be mined includes sub-datasets corresponding to different scenario events; the processing module is specifically used for:

[0049] For each scene event, if the number of data fragments in the subset corresponding to the scene event is greater than the preset quota number corresponding to the scene event, then an information theory-based pruning algorithm is executed on multiple data fragments corresponding to the scene event until the number of data fragments is equal to the preset quota number, thereby obtaining multiple target data fragments corresponding to the scene data.

[0050] If the number of data segments corresponding to the scene event is less than or equal to the preset quota number corresponding to the scene event, then each data segment corresponding to the scene event is determined as a target data segment.

[0051] The target mining dataset is obtained by summarizing the target data fragments corresponding to events in different scenarios.

[0052] In one possible implementation, the processing module is specifically used for:

[0053] Step a, construct the algorithm iteration set, and initialize the algorithm iteration set to an empty set;

[0054] Step b: In the first iteration, for each data segment in the subset, the distribution deviation between the data segment and the subset is calculated based on the trajectory feature vector of the data segment; the data segment with the smallest distribution deviation in the subset is transferred to the algorithm iteration set and enters the next iteration.

[0055] Step c: For each remaining data segment in the subset, based on the trajectory feature vector of the data segment, calculate the information entropy increment and distribution deviation increment brought about by adding the data segment to the algorithm iteration set, and calculate the first score of the data segment based on the information entropy increment and distribution deviation increment;

[0056] Step d: Transfer the top preset number of data segments with the highest scores in the subset to the algorithm iteration set, and proceed to the next iteration;

[0057] Step e, repeat steps c and d until the number of fragments in the algorithm iteration set equals the preset quota.

[0058] In one possible implementation, the method for calculating the information entropy of any non-empty dataset in the information-theory-based pruning algorithm includes:

[0059] The multiple feature dimensions corresponding to the non-empty dataset are discretized into multiple multi-dimensional joint intervals;

[0060] For each multi-dimensional joint interval, the number of data segments in the non-empty dataset whose trajectory feature vectors fall into the multi-dimensional joint interval is counted, and the empirical frequency of the non-empty dataset in the multi-dimensional joint interval is determined based on the number of segments.

[0061] The information entropy of the non-empty dataset is calculated based on the empirical frequency corresponding to each multi-dimensional joint interval.

[0062] In one possible implementation, the device further includes:

[0063] The first rejection module includes:

[0064] The determining unit is used to determine that, for any two data segments in the dataset to be mined, if the cosine similarity between the trajectory feature vectors of the two data segments is greater than a preset threshold, the two data segments are redundant pairs.

[0065] The elimination unit is used to perform redundant data fragment elimination processing on redundant pairs in the dataset to be mined.

[0066] In one possible implementation, the device further includes:

[0067] The scoring module is used to calculate a second score for each data segment in the dataset to be mined based on preset evaluation indicators for each evaluation dimension; wherein the evaluation dimension includes at least one of the comfort dimension, task consistency dimension, and data rationality dimension.

[0068] Accordingly, the first rejection module is specifically used for:

[0069] The data segment with the lower second score in the redundant pair is removed from the dataset to be mined.

[0070] In one possible implementation, the acquisition module is specifically used for:

[0071] For each raw data segment in the natural driving log data, determine the scene event corresponding to the raw data segment, and determine the time window of the raw data segment based on the scene event;

[0072] The original data segment is segmented using the time window to obtain the segmented data segment;

[0073] The segmented data fragments belonging to the same scene event are collectively constructed into a subset of the corresponding scene event, and multiple subsets are aggregated to obtain the dataset to be mined.

[0074] In one possible implementation, the device further includes:

[0075] The second elimination module is used to eliminate original data segments in the natural driving log data that do not meet the preset hard constraints;

[0076] The hard constraints include at least one of the following: hard constraints in the security dimension, hard constraints in the compliance dimension, and hard constraints in the availability dimension of the positioning module or sensor fusion module.

[0077] Thirdly, embodiments of this application provide a computer device, including: a memory and a processor;

[0078] The memory stores computer-executed instructions;

[0079] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0080] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0081] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0082] The driving trajectory data mining method, device, and medium provided in this application obtain a dataset containing multiple data segments from the original natural driving log data. For each data segment, a trajectory feature vector is determined by combining the trajectory position sequence and the driving state sequence of multiple dimensions. Based on the trajectory feature vector of each data segment and the information theory-based pruning algorithm, effective segments are selected from the dataset to be mined. This enables the mining of a target dataset with high information content from massive logs, thereby meeting the requirement of high-quality trajectory data for autonomous driving model training. Attached Figure Description

[0083] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0084] Figure 1 A flowchart illustrating a driving trajectory data mining method provided in Embodiment 1 of this application is provided;

[0085] Figure 2 A flowchart illustrating a specific driving trajectory data mining method provided in Embodiment 2 of this application;

[0086] Figure 3 A flowchart illustrating a specific driving trajectory data mining method provided in Embodiment 3 of this application;

[0087] Figure 4 This is a schematic diagram of the structure of a driving trajectory data mining device provided in Embodiment 4 of this application;

[0088] Figure 5 This is a schematic diagram of the structure of a driving trajectory data mining device provided in Embodiment 5 of this application;

[0089] Figure 6 This is a schematic diagram of the structure of a computer device provided in Embodiment Six of this application.

[0090] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0091] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0092] To facilitate understanding of the technical content of this application, the background technology is described in detail below:

[0093] In existing technologies, multiple driving data segments are typically segmented from raw natural driving logs. These segments are then filtered, retained, or deleted based on rules such as collision rate, number of violations, or trajectory distance thresholds. The basic idea behind this approach is to determine whether a segment can be used for training by using preset metrics, or to compress the data size by deduplicating similar trajectories, thereby forming the dataset required for subsequent model training.

[0094] However, using only trajectory distance or a fixed threshold for redundancy removal is insufficient to accurately reflect the differences in driving state evolution between segments, easily leading to the retention of too many redundant segments or the accidental deletion of effective segments. This not only increases the difficulty of controlling the size of the training data but also weakens the dataset's ability to represent complex driving behaviors, affecting the training effect and generalization ability of autonomous driving models.

[0095] Therefore, how to efficiently filter out high-quality driving data subsets with high information content from large-scale natural driving logs for use in training autonomous driving models has become an urgent technical problem to be solved.

[0096] To address the aforementioned issues, this application proposes a driving trajectory data mining method. This method obtains the dataset to be mined from raw natural driving log data. For each data segment, a trajectory feature vector is determined based on a multi-dimensional driving state sequence. The method then combines the trajectory feature vectors of each data segment with an information-theoretic pruning algorithm to filter effective segments, thereby obtaining the target mining dataset. This method effectively ensures that the dataset used for training has a high information content. Furthermore, the pruning algorithm for filtering the raw natural driving log data effectively controls the size of the training data.

[0097] It should be noted that this application can be deployed in a data processing workflow that connects with a vehicle data acquisition module, a data storage system, and a model training platform, providing higher-quality trajectory samples for subsequent training. The method provided in this application can be executed using computer equipment with data processing capabilities, such as a data mining module deployed in a log storage platform, a training data management platform, or an independent data processing server. This application does not impose specific limitations on the specific form of the executing entity.

[0098] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0099] Figure 1 This application provides a flowchart illustrating a driving trajectory data mining method according to Embodiment 1. Figure 1 As shown, the method includes:

[0100] S101: Obtain the dataset to be mined based on the original natural driving log data.

[0101] The dataset to be mined includes multiple data segments, each of which includes a trajectory location sequence and a multi-dimensional driving state sequence.

[0102] In this step, the raw natural driving log data serves as the input data source for the method, carrying the driving process information collected by the vehicle in a real road environment; the dataset to be mined is used to receive the candidate data set sorted from the raw natural driving log data, and serves as the processing object for subsequent trajectory feature vector determination and segment selection; data segments constitute the basic processing unit of the dataset to be mined, referring to independent segments split from the raw natural driving log data; trajectory position sequence is used to characterize the position change information of the vehicle's motion trajectory in the data segment, and multiple-dimensional driving state sequences are used to characterize the driving state change information of the data segment in different dimensions, such as one or more of the following: speed sequence, acceleration sequence, heading angle sequence, relative distance / speed sequence with surrounding vehicles, etc.

[0103] In practical applications, raw natural driving log data can be transmitted offline to the central storage system via the vehicle-side acquisition link, or it can be imported into the processing environment by the local cache node through batch processing tasks. In specific implementation, multiple data segments that can be processed later can be formed based on the raw natural driving log data, so that each data segment in the generated dataset to be mined contains trajectory location sequences and driving state sequences in multiple dimensions.

[0104] S102: For each data segment, determine the trajectory feature vector of the data segment based on the driving state sequence of multiple dimensions in the data segment.

[0105] In this step, the trajectory feature vector, also known as the trajectory fingerprint, is used to represent the vectorized result of the features of each data segment. Its role is to transform the driving state information of the data segment into a feature representation that is easy to calculate and compare, so as to support the subsequent effective segment selection by the information theory-based pruning algorithm.

[0106] Understandably, the core of this step lies in transforming the original, heterogeneous, and potentially varying lengths of multidimensional driving state sequences into unified, comparable trajectory feature vectors. Using fixed-length vectors establishes a consistent basis for comparison across different data segments. Furthermore, by simultaneously incorporating multiple dimensions of driving state changes, rather than relying on a single collision rate, violation count, or trajectory distance indicator, information such as cruising, following, acceleration / deceleration, steering, and interaction complexity within the segments can be co-encoded. This allows for a more accurate identification of target datasets with higher information content in subsequent screening stages.

[0107] S103: Based on the trajectory feature vector of each data segment and the information theory-based pruning algorithm, the effective segments of the dataset to be mined are selected to obtain the target mining dataset.

[0108] In this step, the information theory-based pruning algorithm is used to filter data segments in the dataset to be mined. Its role is to find the subsets with more information as the target datasets for mining.

[0109] In practice, an iterative pruning process can be adopted. First, initial seed fragments are selected from the dataset to be mined. Then, the information contribution of each fragment relative to the current retained set is progressively evaluated among the unretained fragments. Fragments with higher information contributions are added to the retained set, and the evaluation values ​​of the remaining fragments are updated simultaneously. Fragments with lower information contributions are deleted. This process continues until the preset retention conditions or the calculation budget limit are reached. For example, during the pruning process, the information contribution of data fragments can be quantified based on information theory metrics such as information entropy, conditional entropy, and mutual information to obtain evaluation values.

[0110] Based on the above analysis, this step incorporates trajectory feature vectors and information contribution evaluation into the screening process. This ensures that the screening results no longer rely on a single threshold rule or perform coarse deduplication solely based on trajectory distance. Instead, it retains data segments that better represent the overall driving state distribution and complex behavioral evolution under controlled scale conditions. For a large number of approximate cruise segments, the algorithm will gradually remove the parts with lower information contribution due to their significant information overlap. For a small number of segments with unique state combinations, even if their trajectory length is short or they do not meet the significance requirements of traditional single indicators, they can still be retained because they provide additional information, thus obtaining the target mining dataset.

[0111] In practical applications, target mining datasets can be directly output to training data management platforms as data inputs for imitation learning, reward / value model training, and offline evaluation.

[0112] The driving trajectory data mining method, device, and medium provided in this application obtain a dataset containing multiple data segments from the original natural driving log data. For each data segment, a trajectory feature vector is determined by combining the trajectory position sequence and the driving state sequence of multiple dimensions. Based on the trajectory feature vector of each data segment and the information theory-based pruning algorithm, effective segments are selected from the dataset to be mined. This enables the mining of a target dataset with high information content from massive logs, thereby meeting the requirement of high-quality trajectory data for autonomous driving model training.

[0113] Furthermore, Figure 2 This is a flowchart illustrating a specific driving trajectory data mining method provided in Embodiment 2 of this application. Based on the above embodiments, the dataset to be mined in this embodiment includes sub-datasets corresponding to different scenario events, such as... Figure 2 As shown, the method provided in this embodiment includes:

[0114] S201: Obtain the dataset to be mined based on the original natural driving log data.

[0115] The dataset to be mined includes multiple data segments, each of which includes a trajectory location sequence and a multi-dimensional driving state sequence.

[0116] In this embodiment, the dataset to be mined consists of subsets corresponding to different scenario events. In actual processing, the original natural driving log data needs to be categorized based on scenario events to form multiple subsets corresponding to different scenario events, thus forming the dataset to be mined.

[0117] Among them, scene events are used to identify the driving context to which the original data fragment belongs. For example, scene events include highway cruising, adjacent vehicle cut-in interaction, passing through construction sections, ramp merging and diverging, non-motorized vehicles crossing the road, and waiting for a red light at a straight intersection.

[0118] S202: For each data segment, determine the trajectory feature vector of the data segment based on the driving state sequence of multiple dimensions in the data segment.

[0119] In one possible implementation, this step includes steps 2.1 to 2.2 as follows:

[0120] Step 2.1: For each dimension of the driving state sequence, extract the feature values ​​of the data segments in the dimension based on the driving state sequence.

[0121] In this step, for each dimension of the driving state sequence, the original sampled values ​​within that dimension are first time-aligned and interval-normalized, and then the corresponding feature values ​​are extracted based on the driving state sequence.

[0122] In practice, the feature value can be composed of the statistics, changes, extreme value positions or state proportions of that dimension. For example, it can represent the average level, fluctuation range, continuous change trend and the duration of a specific state, respectively.

[0123] For example, the characteristic values ​​corresponding to the speed sequence are, for example, the average speed and / or the speed standard deviation; the characteristic values ​​corresponding to the acceleration sequence are, for example, at least one of the average acceleration, the standard deviation of acceleration, and the standard deviation of lateral acceleration; the characteristic values ​​corresponding to the steering wheel angle sequence are, for example, the average steering wheel angle; the average time to collision (TTC) corresponding to the relative distance / speed sequence with surrounding vehicles, etc.

[0124] Step 2.2: Based on the feature values ​​of the data fragments in different dimensions, construct the trajectory feature vector of the data fragments.

[0125] In this step, after obtaining the feature values ​​corresponding to each dimension, the feature values ​​in different dimensions are concatenated or mapped according to a preset dimensional order to construct a trajectory feature vector of fixed length.

[0126] Specifically, each component in the vector corresponds to a different driving state dimension, ensuring that different data segments have a consistent vector form within the same representation space. If the feature value of a certain dimension contains multiple sub-indicators, they can be expanded in order of sub-indicators and incorporated into the vector, allowing the vector to simultaneously retain both local change information and overall state information for that dimension. Furthermore, the dimensional length of the trajectory feature vector can be pre-set based on the number of driving state dimensions during implementation; this application does not impose specific limitations on this.

[0127] As a specific example, this step may include the following process:

[0128] 1) For any feature dimension, obtain all feature values ​​of each data segment in the dataset to be mined on that feature dimension.

[0129] 2) Based on all the feature values, use equal-interval binning or quantile binning to divide the feature dimension into multiple intervals.

[0130] 3) Based on the correspondence between the feature values ​​of each data segment and the corresponding intervals of each feature dimension, determine the interval index corresponding to each data segment in each feature dimension.

[0131] Taking equal-interval binning as an example, the method for determining the index of data segment A on feature dimension a includes: first, calculating the global minimum and global maximum values ​​of all feature values ​​on dimension a; then, uniformly dividing the numerical span between the maximum and minimum values ​​into B (e.g., 16) intervals of equal width; determining the interval range into which the feature values ​​of data segment A on feature dimension a fall as the target interval; and determining the interval index of the target interval as the interval index of data segment A on feature dimension a, where the interval code corresponding to each interval is an integer from 0 to B-1. It should be understood that this equal-interval binning method only uniformly divides the numerical range and does not constrain the number of samples contained in each interval.

[0132] Alternatively, taking the quantile binning method as an example, the method for determining the index of data segment A on feature dimension a includes: arranging the feature values ​​of all data segments under feature dimension a in ascending order to form an ordered sequence; counting the number of samples in the ordered sequence that are less than or equal to the feature value of data segment A, dividing this number by the total number of samples and multiplying by one hundred to obtain the quantile rank percentage corresponding to that feature value; dividing the obtained quantile rank percentage by one hundred, multiplying it by the preset number of intervals B, and discarding the decimal part of the calculation result, keeping only the integer part, to obtain the interval index of data segment A on feature dimension a. It should be understood that using the quantile binning method ensures that the number of data segments within each interval is approximately equal.

[0133] 4) For any data segment, concatenate the interval indices of the data segment on each feature dimension to obtain a fixed-length integer vector, which serves as the trajectory feature vector of the data segment.

[0134] For example, the trajectory feature vector obtained by discretizing the 6-dimensional features of a certain data segment is: [3, 12, 7, 1, 15, 8]. It can be understood that the closer the trajectory feature vectors of two data segments are, the more similar the driving behaviors of the two data segments are.

[0135] Optionally, the scene type encoding (such as one-hot encoding) corresponding to the data fragment can also be used simultaneously to construct the trajectory feature vector.

[0136] The implementation of steps 2.1 to 2.2 above allows the trajectory feature vector to fully reflect the behavioral differences and evolutionary characteristics of the segments, thereby improving the effectiveness of subsequent screening steps.

[0137] S203: For each scene event, if the number of data fragments in the subset corresponding to the scene event is greater than the preset quota number corresponding to the scene event, then perform an information theory-based pruning algorithm on the multiple data fragments corresponding to the scene event until the number of data fragments equals the preset quota number, and obtain multiple target data fragments corresponding to the scene data.

[0138] S204: If the number of data segments corresponding to a scene event is less than or equal to the preset quota number corresponding to the scene event, then each data segment corresponding to the scene event will be determined as the target data segment.

[0139] In this embodiment, it is necessary to count the number of data segments in each subset and then compare this number with the corresponding preset quota. When the number of segments corresponding to a certain scene event exceeds the quota, the information redundancy between segments is calculated by combining the trajectory feature vectors of each data segment, and highly redundant segments are removed one by one according to the pruning criterion based on information theory until the number of remaining segments is equal to the quota.

[0140] In other words, the target percentage of each scenario event can be preset, and the total number of data segments in the dataset to be mined can be counted. For each scenario event, the product of the target percentage of the scenario event and the total number of data segments is determined as the preset quota for that scenario event. If the number of data segments in the corresponding subset of the scenario event is less than or equal to the preset quota, all data segments in the subset are retained without pruning. If the number of data segments in the corresponding subset of the scenario event is greater than the preset quota, the subset needs to be pruned, and the number of segments retained should be equal to the preset quota for the scenario event.

[0141] It is understandable that this embodiment can explicitly guarantee the lower limit of difficult scenarios in the dataset through scenario-based quotas, thereby effectively solving the problem in the prior art where simple high-speed cruising scenarios account for too high a proportion in the data, while long-tail difficult scenarios (such as complex interactions, construction areas, and extreme weather) account for too low a proportion, resulting in a serious deviation between the training data distribution and the healthy distribution.

[0142] In one possible implementation, the execution method of "performing an information-theory-based pruning algorithm on multiple data fragments corresponding to the scene event until the number of data fragments equals the preset quota" in step S203 can be implemented based on a greedy iterative algorithm, specifically including:

[0143] Step a, construct the algorithm iteration set S'_m, and initialize the algorithm iteration set to an empty set, that is, construct S'_m=∅.

[0144] The algorithm iteration set is used to hold the fragments retained in each round of pruning. After being initialized to an empty set, it enables the pruning process to accumulate from the unretained state.

[0145] In step b, during the first iteration (i.e., k=1), for each data segment in the subset S_m, the distribution deviation between the data segment and the subset is calculated based on the trajectory feature vector of the data segment. The data segment with the smallest distribution deviation D({c}, S_m) in the subset is transferred to the algorithm iteration set and enters the next iteration.

[0146] Distribution bias is used to characterize the difference between the distribution of a data segment and the overall distribution of a subset of data. It can be measured and calculated using the maximum mean distance or KL divergence.

[0147] Taking the maximum mean distance method as an example, the steps to calculate the distribution deviation between any two non-empty datasets include:

[0148] 1) For the first dataset (e.g., a dataset that only includes data segment c) and the second dataset (e.g., a sub-dataset), obtain the trajectory feature vectors corresponding to all data segments in both datasets;

[0149] 2) The Gaussian kernel is selected as the kernel function to measure feature similarity. The calculation rule of the Gaussian kernel is as follows: first, calculate the square of the Euclidean distance between the two feature vectors, divide the value by the square of twice the preset kernel bandwidth parameter, take the negative value of the result, and then perform exponential operation to obtain the similarity value between the two features.

[0150] 3) Calculate the first part of the baseline value: Traverse all samples within the first dataset, pair the samples together, calculate the Gaussian kernel similarity for each pair, and sum up the similarity values ​​of all pairs; divide the sum by the square of the total number of samples in the first dataset to obtain the first part of the baseline value.

[0151] 4) Calculate the second part of the baseline value: Traverse all samples within the second dataset, pair the samples together and calculate the Gaussian kernel similarity, and summarize all pairing results; divide the sum of the sums by the square of the total number of samples in the second dataset to obtain the second part of the baseline value.

[0152] 5) Calculate the cross-correction value: Cross-pair each feature of the first dataset with each feature of the second dataset, calculate the Gaussian kernel similarity one by one and accumulate them; multiply the accumulated result by negative two, and then divide it by the product of the total number of samples in the first dataset and the total number of samples in the second dataset to obtain the cross-correction value;

[0153] 6) Combine the three values: Add the first part of the baseline value, the second part of the baseline value, and the cross-correction value together. The final sum is the distribution deviation of the two datasets obtained by measuring the maximum mean distance.

[0154] The process in step b can be understood as finding the most representative data segment (that is, the one most closely related to the distribution of the dataset) from the subset.

[0155] Step c: For each remaining data segment in the subset, based on the trajectory feature vector of the data segment, calculate the information entropy increment and distribution deviation increment brought about by adding the data segment to the algorithm iteration set, and calculate the first score of the data segment based on the information entropy increment and distribution deviation increment.

[0156] Step d: Transfer the top number of data segments with the highest scores in the subset to the algorithm iteration set and proceed to the next iteration.

[0157] Step e, repeat steps c and d, until the number of fragments in the algorithm iteration set equals the preset quota.

[0158] The information entropy increment is used to characterize the change in the amount of information in the set after adding candidate fragments. A larger value indicates that the corresponding data fragment fills the sparse region in the feature space, making the frequency distribution more uniform and the dataset containing the data fragment more diverse. Specifically, the calculation method involves calculating the first information entropy and the first distribution deviation of the algorithm iteration set without added data fragments, and the second information entropy and the second distribution deviation of the algorithm iteration set with added data fragments. The difference between the second information entropy and the first information entropy is determined as the information entropy increment corresponding to that data fragment.

[0159] As a specific example, in the information theory-based pruning algorithm, the method for calculating the information entropy of any non-empty dataset includes the following steps 1) to 3):

[0160] 1) Discretize the multiple feature dimensions corresponding to the non-empty dataset into multiple multi-dimensional joint intervals.

[0161] Among them, non-datasets are, for example, the aforementioned set of algorithm iterations without data fragments, or the set of algorithm iterations with data fragments.

[0162] 2) For each multidimensional joint interval, count the number of data segments in the non-empty dataset whose trajectory feature vectors fall into the multidimensional joint interval, and determine the empirical frequency of the non-empty dataset in the multidimensional joint interval based on this number.

[0163] Among them, the multi-dimensional joint interval is a discrete interval obtained by jointly dividing multiple feature dimensions, which is used to transform the continuous feature space into a statistically available distribution space; the empirical frequency is used to represent the relative frequency of occurrence of data segments in each multi-dimensional joint interval, and serves as the probability basis for calculating information entropy.

[0164] 3) Calculate the information entropy of the non-empty dataset based on the empirical frequency corresponding to each multi-dimensional joint interval.

[0165] In practical implementation, firstly, based on the trajectory feature vectors of each data segment in the non-empty dataset, the corresponding multiple feature dimensions are defined with defined boundaries. Then, the discrete results on each dimension are jointly combined to form multiple non-overlapping multi-dimensional joint intervals. The boundaries can be determined based on the feature value range, sample distribution density, or preset quantiles, ensuring that the joint intervals cover the main distribution area of ​​the trajectory feature vectors. Subsequently, the number of data segments falling into each joint interval is counted, and this number is divided by the total number of data segments in the non-empty dataset to obtain the empirical frequency of that joint interval. Based on the empirical frequencies of all joint intervals, the information entropy of the non-empty dataset is calculated according to the information entropy calculation expression.

[0166] Specifically, the empirical probability of any multi-dimensional joint interval can be calculated as follows: count the total number of data segments in the non-empty dataset T whose feature values ​​all fall into the j-th joint interval, divide this number by the total number of data segments in set T, and obtain the empirical probability of the multi-dimensional joint interval. The information entropy can be calculated as follows: for each joint interval, take the logarithm of the empirical probability of the dataset in that interval, and multiply it by the empirical probability of the interval itself; sum up all the multiplication results corresponding to all joint intervals, and finally take the negative of the summation value. The result is the information entropy of the dataset.

[0167] As calculated in steps 1) to 3) above, by jointly mapping multidimensional features to discrete intervals, the distribution pattern of the dataset can be quantitatively expressed in the form of frequency, and information entropy can be obtained accordingly. The calculated information entropy result reflects the distribution uncertainty of the non-empty dataset in multiple feature dimensions, and can be used as a basis for judging the information content and representativeness of the dataset, thereby providing a unified quantitative standard for effective segment selection.

[0168] Furthermore, the distribution deviation increment is used to characterize the change in distribution structure after adding candidate fragments. A smaller value indicates that the subset distribution is closer to the whole set distribution after adding the data fragment, and the dataset retaining the data fragment has higher fidelity. Specifically, it is calculated by calculating the first distribution deviation between the algorithm iteration set without the added data fragment and the subset before the iteration begins, and the second distribution deviation between the algorithm iteration set with the added data fragment and the subset before the iteration begins. The second distribution deviation and the first distribution deviation of the first information entropy are then determined as the distribution deviation increment corresponding to the data fragment.

[0169] The distributional bias between any two non-empty datasets has been introduced earlier and will not be repeated here.

[0170] The first score is used to comprehensively rank candidate segments. Specifically, it can be obtained by weighted summation of the information entropy increment and distribution bias component corresponding to the data segment. For example, the first score can be calculated as the difference between the product of the information entropy increment, the distribution bias increment, and the balance coefficient λ. The balance coefficient can be determined according to the actual situation. For example, the balance coefficient can take any value from 0.01 to 1.0, and this application does not impose any restrictions on this; for example, it can be 0.1. It should be understood that the balance coefficient is used to control the trade-off between the two factors. A smaller balance coefficient tends to emphasize improving the diversity of the retained dataset, while a larger balance coefficient tends to emphasize the fidelity of the retained dataset.

[0171] In other words, during subsequent iterations of the first iteration (i.e., when k > 1), for each remaining data segment in the subset, the increase in information entropy and the increase in distribution bias resulting from adding the data segment to the algorithm iteration set are calculated, and the two are combined to form a first score. Then, a predetermined number of segments are added to the set according to the scores from high to low, until the predetermined quota is reached. The predetermined number can be determined based on the required level of detail and efficiency of the selection, and this application does not impose any restrictions on it. For example, the predetermined number can be 1, 2, or 3.

[0172] The greedy iterative pruning method proposed in steps a to e above achieves a hierarchical optimization effect during the selection process: the initial retained fragments selected in the initialization phase can fit the overall distribution of the corresponding subset dataset, and in each subsequent iteration, when adding new fragments, both the incremental information contribution and the consistency of the global distribution are considered simultaneously. This approach can ensure that the final dataset has higher information content and complete scene representativeness (i.e., dataset fidelity) while strictly controlling the total size of the output fragments, effectively enhancing the supporting role of the mined dataset in driving behavior modeling.

[0173] S205: Summarize the target data fragments corresponding to events in different scenarios to obtain the target mining dataset.

[0174] In this step, the target data fragments from different scenario events need to be aggregated to form a target mining dataset with controlled size and balanced scenario coverage. Meanwhile, the retained dataset possesses a high amount of information due to the execution of an information-theoretic pruning algorithm.

[0175] The driving trajectory data mining method provided in this application, through scene-level quota allocation, can explicitly guarantee the lower limit proportion of difficult scenarios in the dataset, avoiding simple scenarios dominating the dataset and improving the scene distribution balance of the target dataset. Simultaneously, by performing an information-theoretic pruning algorithm on each excess subset of scene events, the method can control the overall data size while increasing the amount of information retained in the dataset. When the trajectory data obtained by this method is used in the model training set, the model can fully learn the characteristics of various rare and difficult driving scenarios, alleviating the model bias problem caused by imbalanced training sample categories, and effectively improving the model's ability to identify and generalize decisions regarding rare and dangerous conditions.

[0176] Furthermore, Figure 3 This is a schematic diagram of a specific driving trajectory data mining process provided in Embodiment 3 of this application. Based on the above embodiments, as follows... Figure 3 As shown, before step 102 of embodiment one or step S202 of embodiment two, the method provided in this application includes:

[0177] S301: For each raw data segment in the natural driving log data, determine the scene event corresponding to the raw data segment, and determine the time window of the raw data segment based on the scene event.

[0178] In this step, the time window is used to define the effective time range around the scene event, and the segmented data fragments are used to carry more granular local driving processes within the scene.

[0179] In actual processing, for each raw data segment, the corresponding scene event can be identified by combining the road type where the vehicle is located, changes in traffic participants, or longitudinal and lateral control characteristics of the vehicle during the data segment collection period, and the time window can be determined according to the preset duration or boundary conditions of the scene event.

[0180] For example, for simple and standard scenario events (such as cut-in interactions, merging and diverging at construction ramps, waiting at a red light at a straight intersection, etc.), the time window can be formed by extending A seconds forward and backward, with the time of the event as the central node; for complex scenario events (such as passing through a construction section, etc.), the time window can be formed by extending B seconds forward and backward, with B being greater than A.

[0181] Optionally, for more complex scenarios and events, the time window can be dynamically adjusted according to the complexity of the specific event.

[0182] In addition, if the time windows of any two events in the natural driving log data overlap by more than M%, the original data segments corresponding to the two events will be merged into one original data segment and marked as multiple scene events. The merged original data segment can be distributed in a subset of each corresponding scene event.

[0183] In one possible implementation, before step S301 is executed, the method of this application embodiment may further include: removing original data segments in the natural driving log data that do not meet preset hard constraints, wherein the hard constraints include at least one of hard constraints in the safety dimension, hard constraints in the compliance dimension, and hard constraints in the availability dimension of the positioning module or sensor fusion module.

[0184] In other words, to ensure the usability of data segments, each raw data segment in the natural driving log data undergoes a hard-constraint-based elimination process. Specifically, segments that do not meet the hard constraints can be eliminated based on safety warnings, traffic violation signs, location quality indicators, and fusion status indicators recorded in the log.

[0185] For example, hard constraints in the safety dimension can be designed based on indicators related to personal safety, such as collision distance, TTC, and maximum lateral acceleration; while compliance dimensions can be determined based on indicators of behaviors that violate traffic regulations, such as whether traffic signals are ignored, whether speeding, or whether lane lines are crossed.

[0186] For example, if the collision distance is less than a safety threshold (e.g., 2 meters) in the driving scenario corresponding to the original data segment, or the speed exceeds the speed limit by more than 10%, or a red light signal is violated, or the positioning module or sensor fusion module is unavailable, then the original data segment will be removed from the natural driving log data.

[0187] Understandably, the massive amounts of real-world driving data relied upon for model training come from drivers with different driving styles and habits. The learning process focuses on the vehicle's dynamic behavior in interaction with various road users, aiming to cultivate "excellent drivers." Therefore, extracting superior driving data from real-world driving data is crucial for model training.

[0188] The method provided in this implementation first filters out data with key issues such as safety and compliance through hard constraints. This allows the training data obtained from this filtering to guide the model to learn more about smooth, safe, and reasonable driving behaviors, weaken the behavioral bias caused by poor driving samples, and improve the safety and rationality of the model's decision-making under various complex traffic interactions.

[0189] S302: The original data segment is divided into segments using a time window to obtain the segmented data segments.

[0190] S303: Combine the segmented data fragments belonging to the same scene event to form a subset of the corresponding scene event, and summarize multiple subsets to obtain the dataset to be mined.

[0191] In the specific implementation, the original data segment is divided according to the time window determined in step S301, so that one original data segment can generate one or more segmented data segments. Then, the segmentation results under the same scene event are merged to form a subset, and finally all subsets are summarized to obtain the dataset to be mined.

[0192] The aforementioned steps S301 and S303 determine the time window for segmenting the corresponding data segment based on the specific scenario event type. This can match the interaction duration and evolution pattern of various driving scenarios, accurately extract data segments containing the complete scenario cause, interaction process, and ending state, and avoid losing key interaction features due to excessively short windows or introducing a large amount of irrelevant and redundant driving data due to excessively long windows.

[0193] S304: For any two data segments in the dataset to be mined, if the cosine similarity between the trajectory feature vectors of the two data segments is greater than a preset threshold, then the two data segments are determined to be a redundant pair.

[0194] In this step, the preset threshold can be determined according to the requirements of the degree of redundancy removal in actual application, for example, it can be 95%, and this application does not impose specific restrictions on it.

[0195] S305: Perform redundant data fragment removal on redundant pairs in the dataset to be mined.

[0196] It is understandable that the dataset to be mined obtained through the preprocessing process in steps S301 to S305 above can be used to perform subsequent steps such as over-judgment and pruning of data fragments corresponding to scene events.

[0197] In one possible implementation, any data segment in the redundant pair can be removed.

[0198] In one possible implementation, the method provided in this application embodiment may further include the following: calculating a second score for each data segment in the dataset to be mined based on preset evaluation indicators for each evaluation dimension. The evaluation dimensions include at least one of a comfort dimension, a task consistency dimension, and a data rationality dimension.

[0199] Specifically, the indicators corresponding to the comfort dimension are used to evaluate the comfort level of drivers and passengers in the driving scenario corresponding to the data segment. The comfort level can be characterized by acceleration and deceleration smoothness, lateral offset stability, and ride comfort. The corresponding indicators include, for example, longitudinal acceleration change rate (jerk), lateral acceleration change rate, and steering wheel angular velocity. For example, the smaller the absolute value of the longitudinal acceleration corresponding to the data segment and the smoother the steering wheel angular velocity, the higher the comfort dimension score.

[0200] The metrics corresponding to the task consistency dimension are used to evaluate the degree of consistency between the data content of the data segment and the established driving task, scenario goal and behavioral constraints in the driving scenario corresponding to the data segment. The corresponding metrics include, for example, whether the navigation route is followed and whether the lane change intention is completed.

[0201] The metrics corresponding to the data rationality dimension are used to evaluate the rationality of the data content of a data segment. These metrics may include, for example, the rationality of the data in terms of timestamp monotonicity, multi-sensor consistency, and frame drop rate. For example, data segments with strictly increasing timestamps and clock jumps have a higher score in the data rationality dimension; and / or, data segments with the maximum deviation between the timestamps of each sensor exceeding a threshold (e.g., 50ms) have a lower score in the data rationality dimension; and / or, data segments with a percentage of valid frames per unit time lower than a preset frame drop rate threshold (e.g., 95%) have a lower score in the data rationality dimension.

[0202] Optionally, the evaluation dimension may also include an efficiency dimension, the corresponding indicators of which are used to evaluate the traffic efficiency of driving behavior corresponding to the data segment. The corresponding indicators include, for example, the frequency of invalid start-stop and stop, and the extra idling waiting time.

[0203] The second score is calculated based on the scores of the data segment on each evaluation dimension.

[0204] In practical implementation, at least one corresponding evaluation index can be preset for the comfort dimension, task consistency dimension, and data rationality dimension. Each evaluation index is then mapped to a sub-score with a unified dimension. For example, the index values ​​can be normalized to between 0 and 1 and then weighted and summarized to obtain the second score for the corresponding data segment. For any data segment, if multiple evaluation dimensions are set simultaneously, the sub-scores of each dimension can be synthesized according to preset weights to obtain the second score for that data segment.

[0205] The weights can be set according to the emphasis of the training task on comfort, task consistency and rationality, and this application does not limit them.

[0206] Based on the scoring steps described above, a specific implementation of step S305 may include: removing data segments with lower second scores from the dataset to be mined in redundant pairs.

[0207] In other words, after any two data segments in the dataset to be mined are identified as redundant pairs, their second scores are compared. The data segment with the lower second score is marked as an object to be removed and deleted from the dataset to be mined, while the data segment with the higher second score is retained as the basis for subsequent data mining based on the pruning algorithm.

[0208] The method provided in this implementation introduces a second scoring system oriented towards comfort, task consistency, and data rationality. This allows redundancy removal to move beyond solely relying on trajectory similarity and instead incorporates a comprehensive selection based on the overall quality of the segments, retaining data segments with superior overall performance from pairs of redundant segments. This ensures that the dataset to be mined after redundancy removal maintains a higher overall sample quality, fully reflecting the superior attributes of the retained data in terms of driving behavior and data integrity. Simultaneously, it enhances the supporting value of the retained segments for autonomous driving scenario modeling and model training, guiding the model to learn more standardized, smoother, and more reliable driving behaviors.

[0209] The driving trajectory data mining method provided in this application first removes raw data segments that do not meet various preset hard constraints, filtering out data with sensor failures, abnormal positioning, non-compliance, and safety defects from the source, ensuring the basic reliability of the input data while achieving preliminary mining of excellent driving data. Next, it configures trajectory extraction time windows differently according to different event types in various scenarios, fully covering the entire process of occurrence, interaction, and conclusion of events in each scenario, achieving accurate and regular extraction of effective data segments, and constructing a standardized and highly complete dataset to be mined. Finally, it identifies highly repetitive samples based on the high-dimensional feature similarity of each segment in the dataset to be mined, performing batch removal of redundant segments to reduce dataset storage and training redundancy overhead. Furthermore, this embodiment introduces a multi-dimensional soft scoring mechanism in the redundant segment removal stage, comprehensively considering dimensions such as data rationality, driving comfort, driving task consistency, and traffic efficiency to quantitatively score paired redundant segments, directly removing inferior segments with low overall scores, and prioritizing the retention of high-quality data segments with more standardized and stable driving behavior. This ensures that while removing redundancy, subsequent models can learn even better driving behaviors.

[0210] Figure 4 This is a schematic diagram of the structure of a driving trajectory data mining device provided in Embodiment 4 of this application, as shown below. Figure 4 As shown, the driving trajectory data mining device 40 provided in this embodiment includes:

[0211] The acquisition module 401 is used to acquire the dataset to be mined based on the original natural driving log data; the dataset to be mined includes multiple data segments; each data segment includes a trajectory location sequence and a driving state sequence in multiple dimensions;

[0212] The determination module 402 is used to determine the trajectory feature vector of each data segment based on the driving state sequence of multiple dimensions in the data segment.

[0213] Processing module 403 is used to filter effective segments of the dataset to be mined based on the trajectory feature vector of each data segment and an information theory-based pruning algorithm to obtain the target mining dataset.

[0214] The driving trajectory data mining device 40 provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0215] Figure 5 This is a schematic diagram of the structure of a driving trajectory data mining device provided in Embodiment 5 of this application, as shown below. Figure 5 As shown, based on the above embodiments, the driving trajectory data mining device 40 provided in this embodiment further includes:

[0216] The first rejection module 404 includes:

[0217] The determination unit is used to determine two data segments as redundant pairs if the cosine similarity between the trajectory feature vectors of the two data segments is greater than a preset threshold.

[0218] The removal unit is used to remove redundant data fragments from redundant pairs in the dataset to be mined.

[0219] The scoring module 405 is used to calculate the second score of each data segment in the dataset to be mined based on the preset evaluation indicators for each evaluation dimension; wherein the evaluation dimension includes at least one of the comfort dimension, task consistency dimension and data rationality dimension;

[0220] The second elimination module 406 is used to eliminate raw data segments in the natural driving log data that do not meet the preset hard constraints;

[0221] Among them, hard constraints include at least one of the following: hard constraints in the security dimension, hard constraints in the compliance dimension, and hard constraints in the availability dimension of the positioning module or sensor fusion module.

[0222] In one possible implementation, the determining module 402 is specifically used for:

[0223] For each dimension of the driving state sequence, feature values ​​of the data segments in that dimension are extracted based on the driving state sequence;

[0224] Based on the feature values ​​of the data fragments in different dimensions, the trajectory feature vector of the data fragments is constructed.

[0225] In one possible implementation, the dataset to be mined includes sub-datasets corresponding to events in different scenarios; the processing module 403 is specifically used for:

[0226] For each scene event, if the number of data fragments in the subset corresponding to the scene event is greater than the preset quota number corresponding to the scene event, then an information theory-based pruning algorithm is executed on multiple data fragments corresponding to the scene event until the number of data fragments equals the preset quota number, thus obtaining multiple target data fragments corresponding to the scene data.

[0227] If the number of data segments corresponding to a scene event is less than or equal to the preset quota number corresponding to the scene event, then each data segment corresponding to the scene event will be determined as the target data segment.

[0228] The target mining dataset is obtained by summarizing the target data fragments corresponding to events in different scenarios.

[0229] In one possible implementation, the processing module 403 is specifically used for:

[0230] Step a, construct the algorithm iteration set, and initialize the algorithm iteration set to an empty set;

[0231] Step b: In the first iteration, for each data segment in the subset, calculate the distribution deviation between the data segment and the subset based on the trajectory feature vector of the data segment; transfer the data segment with the smallest distribution deviation in the subset to the algorithm iteration set and enter the next iteration.

[0232] Step c: For each remaining data segment in the subset, based on the trajectory feature vector of the data segment, calculate the information entropy increment and distribution deviation increment brought about by adding the data segment to the algorithm iteration set, and calculate the first score of the data segment based on the information entropy increment and distribution deviation increment.

[0233] Step d: Transfer the top number of data segments with the highest scores in the subset to the algorithm iteration set and proceed to the next iteration.

[0234] Step e, repeat steps c and d, until the number of fragments in the algorithm iteration set equals the preset quota.

[0235] In one possible implementation, the information entropy calculation method for any non-empty dataset in the information-theory-based pruning algorithm includes:

[0236] Discretize the multiple feature dimensions corresponding to the non-empty dataset into multiple multi-dimensional joint intervals;

[0237] For each multidimensional joint interval, count the number of data segments in the non-empty dataset whose trajectory feature vectors fall into the multidimensional joint interval, and determine the empirical frequency of the non-empty dataset in the multidimensional joint interval based on the number of segments.

[0238] The information entropy of the non-empty dataset is calculated based on the empirical frequency corresponding to each multi-dimensional joint interval.

[0239] In one possible implementation, the first rejection module 404 is specifically used for:

[0240] Remove the data segments with the second lowest scores from the redundant pairs from the dataset to be mined.

[0241] In one possible implementation, the acquisition module 401 is specifically used for:

[0242] For each raw data segment in the natural driving log data, determine the scene event corresponding to the raw data segment, and determine the time window of the raw data segment based on the scene event;

[0243] The original data segment is divided into segments using a time window to obtain the segmented data segments;

[0244] Data fragments belonging to the same scene event are combined to form a subset of the corresponding scene event, and multiple subsets are aggregated to obtain the dataset to be mined.

[0245] The driving trajectory data mining device 40 provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0246] Figure 6 This is a schematic diagram of the structure of a computer device provided in Embodiment Six of this application. Figure 6 As shown, the computer device 50 provided in this embodiment includes at least one processor 501 and a memory 502. Optionally, the computer device 50 further includes a communication component 503. The processor 501, memory 502, and communication component 503 are connected via a bus 504.

[0247] In a specific implementation, at least one processor 501 executes computer execution instructions stored in memory 502, causing at least one processor 501 to perform the above-described method.

[0248] The specific implementation process of processor 501 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0249] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0250] The memory may include read-only memory and random access memory. The memory may be volatile or non-volatile, or may include both. Non-volatile memory may include read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory may include random access memory (RAM), which serves as an external cache. Many forms of RAM are available by way of example, but not limitation. Examples include Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Sync Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0251] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0252] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method.

[0253] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, implement the above-described method.

[0254] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as SRAM, EEPROM, EPROM, PROM, ROM, magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0255] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside within an ASIC. Alternatively, the processor and the readable storage medium can exist as discrete components in a device.

[0256] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0257] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0258] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0259] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0260] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0261] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

[0262] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for mining driving trajectory data, characterized in that, The method includes: Based on the original natural driving log data, a dataset to be mined is obtained; the dataset to be mined includes multiple data segments; each data segment includes a trajectory location sequence and a driving state sequence in multiple dimensions; For each data segment, a trajectory feature vector is determined based on the driving state sequence of multiple dimensions in the data segment; Based on the trajectory feature vector of each data segment and an information-theoretic pruning algorithm, the dataset to be mined is filtered for effective segments to obtain the target mining dataset.

2. The method according to claim 1, characterized in that, Determining the trajectory feature vector of the data segment based on the multi-dimensional driving state sequence in the data segment includes: For each dimension of the driving state sequence, feature values ​​of the data segment in that dimension are extracted based on the driving state sequence; Based on the feature values ​​of the data segment in different dimensions, the trajectory feature vector of the data segment is constructed.

3. The method according to claim 1, characterized in that, The dataset to be mined includes sub-datasets corresponding to different scene events; the effective fragment selection of the dataset to be mined is performed based on the trajectory feature vector of each data fragment and an information-theoretic pruning algorithm to obtain the target mining dataset, including: For each scene event, if the number of data fragments in the subset corresponding to the scene event is greater than the preset quota number corresponding to the scene event, then an information theory-based pruning algorithm is executed on multiple data fragments corresponding to the scene event until the number of data fragments is equal to the preset quota number, thereby obtaining multiple target data fragments corresponding to the scene data. If the number of data segments corresponding to the scene event is less than or equal to the preset quota number corresponding to the scene event, then each data segment corresponding to the scene event is determined as a target data segment. The target mining dataset is obtained by summarizing the target data fragments corresponding to events in different scenarios.

4. The method according to claim 3, characterized in that, The step of performing an information-theoretic pruning algorithm on multiple data fragments corresponding to the scene event until the number of data fragments equals the preset quota includes: Step a, construct the algorithm iteration set, and initialize the algorithm iteration set as an empty set; Step b: In the first iteration, for each data segment in the subset, the distribution deviation between the data segment and the subset is calculated based on the trajectory feature vector of the data segment; the data segment with the smallest distribution deviation in the subset is transferred to the algorithm iteration set and enters the next iteration. Step c: For each remaining data segment in the subset, based on the trajectory feature vector of the data segment, calculate the information entropy increment and distribution deviation increment brought about by adding the data segment to the algorithm iteration set, and calculate the first score of the data segment based on the information entropy increment and distribution deviation increment; Step d: Transfer the top preset number of data segments with the highest scores in the subset to the algorithm iteration set, and proceed to the next iteration; Step e, repeat steps c and d until the number of fragments in the algorithm iteration set equals the preset quota.

5. The method according to any one of claims 1 to 4, characterized in that, In the information-theory-based pruning algorithm, the method for calculating the information entropy of any non-empty dataset includes: The multiple feature dimensions corresponding to the non-empty dataset are discretized into multiple multi-dimensional joint intervals; For each multi-dimensional joint interval, the number of data segments in the non-empty dataset whose trajectory feature vectors fall into the multi-dimensional joint interval is counted, and the empirical frequency of the non-empty dataset in the multi-dimensional joint interval is determined based on the number of segments. The information entropy of the non-empty dataset is calculated based on the empirical frequency corresponding to each multi-dimensional joint interval.

6. The method according to any one of claims 1 to 4, characterized in that, Before obtaining the target mining dataset by filtering effective fragments from the multiple data fragments using the trajectory feature vector of each data fragment and an information-theoretic pruning algorithm, the method further includes: For any two data segments in the dataset to be mined, if the cosine similarity between the trajectory feature vectors of the two data segments is greater than a preset threshold, then the two data segments are determined to be a redundant pair. Redundant data fragments are removed from the redundant pairs in the dataset to be mined.

7. The method according to claim 6, characterized in that, The method further includes: Based on the preset evaluation indicators for each evaluation dimension, a second score is calculated for each data segment in the dataset to be mined; wherein, the evaluation dimension includes at least one of the comfort dimension, task consistency dimension, and data rationality dimension; Accordingly, the redundant data fragment removal process for redundant pairs in the dataset to be mined includes: The data segment with the lower second score in the redundant pair is removed from the dataset to be mined.

8. The method according to claim 3 or 4, characterized in that, The process of obtaining the dataset to be mined based on the original natural driving log data includes: For each raw data segment in the natural driving log data, determine the scene event corresponding to the raw data segment, and determine the time window of the raw data segment based on the scene event; The original data segment is segmented using the time window to obtain the segmented data segment; The segmented data fragments belonging to the same scene event are collectively constructed into a subset of the corresponding scene event, and multiple subsets are aggregated to obtain the dataset to be mined.

9. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Remove the original data segments in the natural driving log data that do not meet the preset hard constraints; The hard constraints include at least one of the following: hard constraints in the security dimension, hard constraints in the compliance dimension, and hard constraints in the availability dimension of the positioning module or sensor fusion module.

10. A computer device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-9.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-9.

Images