Data processing method and equipment

Abstract

The embodiment of the invention provides a data processing method and equipment. The method comprises the following steps: inputting map difference data into a data processing model to obtain a map difference vector of the map difference data; the data processing model is obtained by training according to a map sample cluster and marking information of the map sample cluster, the map sample cluster is obtained by clustering map samples, the marking information of the map samples in the same map sample cluster is the same, and the map difference data is determined according to the difference between the uploaded data and the map data; and according to the similarity between the contrast map vector of the contrast map data and the map difference vector, screening out false alarm data in the map difference data to obtain target difference data. The method is used for improving the data processing effect.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data processing method and device. Background Art

[0002] Autonomous driving is one of the key technologies of modern automobiles, which places higher demands on the real-time and accuracy of maps. Traditional high-precision map update methods can no longer meet these requirements.

[0003] At present, the industry generally believes that crowdsourcing-based methods will become the mainstream of high-precision map updates. Crowdsourcing updates mainly extract road vectorized data through the model on the vehicle and upload it to the cloud. Subsequently, the map is automatically updated through steps such as classification, matching, and aggregation. Therefore, the quality of the vectorized data uploaded by the vehicle to the cloud directly determines the effect of the final map update, which in turn affects the performance of the autonomous driving function. However, since the model is affected by the quality of the training samples during training, the model cannot extract high-quality vectorized data when it is used. Therefore, when processing the data, the recognition effect of false alarm data that does not belong to the map data in the data is not good, resulting in poor results for subsequent map data updates. Summary of the invention

[0004] The embodiments of the present application provide a data processing method and device to improve the update effect of map data.

[0005] In a first aspect, an embodiment of the present application provides a model training method, comprising:

[0006] The map difference data is input into a data processing model to obtain a map difference vector of the map difference data; the data processing model is trained according to the map sample cluster and the annotation information of the map sample cluster, the map sample cluster is obtained by clustering the map samples, wherein the map samples in the same map sample cluster have the same annotation information, and the map difference data is determined according to the difference between the uploaded data and the map data;

[0007] According to the similarity between the reference map vector and the map difference vector of the reference map data, false positive data in the map difference data are screened out to obtain target difference data.

[0008] In a second aspect, an embodiment of the present application provides a data processing device, including:

[0009] An input module is used to input the map difference data into a data processing model to obtain a map difference vector of the map difference data; the data processing model is trained according to map sample clusters and annotation information of the map sample clusters, and the map sample clusters are obtained by clustering the map samples, wherein at least one map sample cluster includes more than two map samples;

[0010] The determination module is used to determine the target difference data in the map difference data according to the similarity between the reference map vector and the map difference vector of the reference map data.

[0011] In a third aspect, an embodiment of the present application provides an electronic device, including: a memory, a processor;

[0012] Memory stores computer-executable instructions;

[0013] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above-mentioned various possible implementations.

[0014] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the various possible implementation methods as described above.

[0015] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the various possible implementation methods described above when executed by a processor.

[0016] The data processing method and device provided in the embodiment of the present application obtain a map difference vector of the map difference data by inputting map difference data into a data processing model, and determine the target difference data in the map difference data according to the similarity between the reference map vector of the reference map data and the map difference vector, wherein, since the annotation information of map samples in the same map sample cluster is the same, the annotation efficiency of the samples can be improved when the model is trained, thereby reducing the workload of manual annotation, and by ensuring that the samples in the same cluster have consistent annotation information, the influence of outliers can be reduced, thereby reducing the noise level in the training samples, and thus reducing the occurrence of problems such as missed recognition or wrong recognition, thereby enabling the data processing model to learn a wider range of changes and different forms of expression, thereby enhancing its robustness to changes in input data, so that when the data processing model is used, it can more accurately extract the map difference vector of the map difference data, thereby enabling the effect of screening out false positive data in the map difference data according to the similarity between the reference map vector of the reference map data and the map difference vector to be better, thereby achieving an improved update effect on the map data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0018] Figure 1 A schematic diagram of a scenario of the data processing method provided in this application;

[0019] Figure 2 Schematic diagram of the data processing method provided for this application Figure 1 ;

[0020] Figure 3 Schematic diagram of the data processing method provided for this application Figure 2 ;

[0021] Figure 4 Schematic diagram of the data processing method provided for this application Figure 3 ;

[0022] Figure 5 A flowchart of another data processing method provided for this application;

[0023] Figure 5a A schematic diagram of the sample pair construction method provided for this application;

[0024] Figure 5b A schematic diagram of the process of clustering images provided in this application;

[0025] Figure 5c Schematic diagram of the effective clusters provided for this application;

[0026] Figure 5d A schematic diagram of the cluster classification process provided for this application;

[0027] Figure 6 A schematic diagram of the structure of the data processing device provided by this application;

[0028] Figure 7 A schematic diagram of the structure of the electronic device provided in this application.

[0029] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0030] First, the terms involved in this application are explained:

[0031] Map crowdsourcing updates refer to the process of continuously updating and improving map information through geographic data and feedback provided by a large number of users or volunteers. This data can include new roads, landmarks, traffic signs, etc., which can be collected and verified through crowdsourcing, which can improve the accuracy and timeliness of maps and reduce the high cost and time delay of traditional map updates.

[0032] Clustering is an unsupervised learning method used to group objects in a data set so that objects in the same group are more similar to each other, while objects in different groups are less similar, thereby discovering the inherent structure and patterns of the data.

[0033] Cluster Feature refers to the statistical or geometric properties used to describe a cluster (i.e. a group of similar data points) in cluster analysis.

[0034] In the process of crowdsourcing data collection, the vehicle terminal first identifies and extracts vector elements such as road signs, arrows, and markings. Due to the limitation of sample coverage, although existing work is still committed to improving the recognition ability of the vehicle side to reduce the proportion of misidentification and missed identification, it is unrealistic to completely avoid these problems. In addition, subsequent map matching and other mapping tool chains may also introduce certain misjudgments and missed judgments. Therefore, the current mainstream map crowdsourcing update methods mainly include two types: one is to improve the recognition ability on the end, and provide the error samples to the end model for training through manual processing results, thereby improving the recognition ability of the model; the second is to use vector semantic filtering. According to the detection results on the end, various vector features are used to formulate rules to filter misidentification and missed identification. The modeling process of these two methods is offline. After the modeling is completed, the new model replaces the old model to achieve improved recognition capabilities.

[0035] However, these methods have the following problems: the iteration cycle of the on-end model is long, and more labeled samples are required for model training. Problems discovered in one cycle often need to wait until the next cycle to be improved, and similar errors will recur before they are improved. In addition, it is difficult to distinguish whether an individual is misidentified based on the existing vector semantic features, and it is necessary to increase the feature dimension, such as the image dimension.

[0036] The data processing method provided in the embodiment of the present application obtains annotation information for model training by clustering map samples and annotating the clustered map sample clusters, thereby ensuring that there is at least one annotation information for describing more than two map samples. During model training, these samples serve as multiple positive samples, providing features under multiple perspectives and conditions, thereby improving the generalization and robustness of the trained model. At the same time, the workload of manual review and correction can be reduced by annotating the map sample clusters, thereby improving the efficiency of manual work and making the entire process more automated. In addition, when the obtained data processing model is used to extract map difference vectors, the obtained map difference vectors can be highly generalized and distinguishable, and can support fast and accurate similarity calculation and classification tasks while maintaining image semantic information, thereby making the target difference data determined by the similarity between the map vector and the map difference vector more accurate, thereby improving the processing effect of the map difference data.

[0037] The technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0038] Figure 1 A schematic diagram of a scenario of the data processing method provided in this application, such as Figure 1 As shown, the specific application scenario of the present application can be a crowdsourcing update tool system, which can be a server, wherein the server can include a crowdsourcing update tool chain, job results, a production feedback analysis module, and an FP filtering module based on image similarity, and the production feedback analysis module includes an image representation model, a clustering analysis unit, and a historical sample similarity unit. Among them, the crowdsourcing update tool chain is used to obtain change elements and states, and the FP filtering module based on image similarity classifies the changes. After filtering non-real changes, the job results are submitted to manual confirmation, and the manual job results are used in the production feedback analysis module.

[0039] The data processing method in the embodiment of the present application does not impose any special restrictions on the implementation method of the execution subject, as long as the execution subject can input the map difference data into the data processing model to obtain the map difference vector of the map difference data; the data processing model is trained based on the map sample clusters and the annotation information of the map sample clusters, and the map sample clusters are obtained by clustering the map samples, wherein there is at least one map sample cluster including more than two map samples; and the target difference data in the map difference data is determined based on the similarity between the reference map vector and the map difference vector of the reference map data.

[0040] Figure 2 Schematic diagram of the data processing method provided for this application Figure 1 ,like Figure 2 As shown, the method includes:

[0041] S201. Inputting the map difference data into a data processing model to obtain a map difference vector of the map difference data; the data processing model is trained based on map sample clusters and annotation information of the map sample clusters, the map sample clusters are obtained by clustering map samples, wherein the map samples in the same map sample cluster have the same annotation information, and the map difference data is determined based on the difference between the uploaded data and the map data.

[0042] Among them, map difference data may refer to the uploaded data of individuals identified by sensors, and the data obtained after differential matching with the map data of individuals already on the map, wherein individuals may refer to objects on the map, which may include streets, stations, signboards and other objects. In an embodiment of the present application, map difference data may include changes in map elements and changes in status. Among them, changes in map elements include photos of signs, markings, arrows, etc., and status changes include added status, deleted status and updated status, wherein the added status indicates elements that appear in the map data but do not exist in the historical map data. The deleted status indicates elements that do not exist in the map data but exist in the historical map data. The updated status indicates elements that exist in the map data and also exist in the historical map data, but whose attributes or positions have changed.

[0043] For example, sensors (such as cameras) capture the location information of billboards on the street in real time. By comparing the data of billboards with billboards recorded in the map database, differential matching is used to calculate the difference data representing the changes in the billboard style.

[0044] The data processing model may refer to a model for extracting a map difference vector from the map difference data, wherein the model may include an image encoder. The map difference vector may be obtained by inputting the map difference data into the image encoder in the data processing model through encoding, decoding and other processes.

[0045] In the embodiment of the present application, the data processing model can be trained based on the map sample clusters and the annotation information of the map sample clusters, wherein:

[0046] A map sample cluster may refer to a sample cluster obtained after clustering map samples, and a map sample in a map sample cluster may refer to a map image or photo uploaded by a user or a vehicle for model training. For example, a map sample may include images or photos of various objects on a road, such as road traffic signs, billboards, street lights, etc. Clustering map samples may refer to clustering according to the features of the map samples, thereby forming a number of map sample clusters with similar environmental conditions.

[0047] Annotation information may refer to information for marking a map sample and characterizing map elements. Annotation information includes, but is not limited to, information such as the type, location, and attributes of an element. Annotation information may be a text description, a label, or a combination of labels, used to indicate specific features of an element in a map sample. In an embodiment of the present application, since the clustered sample clusters all belong to the same object, when annotating, the annotation information of the sample cluster can be used as the annotation information of each map sample in the sample cluster. This greatly improves annotation efficiency compared to annotating individual photos one by one.

[0048] In the embodiment of the present application, the annotation information can be obtained by manual annotation and automatic annotation.

[0049] When the annotation information can be obtained by manual annotation, the clustered map sample clusters can be displayed to the user, and the annotation information of the map sample clusters can be obtained by responding to the user's annotation of the map sample clusters.

[0050] When the annotation information can be obtained by automatic annotation, it can be generated according to the vector feature data of the map sample in the reference map. For example, when the obtained map sample is a sample of an object in an existing map, the natural language organized by the vector feature data can be used to annotate the annotation information of the map sample. For example, when a map sample in the obtained map sample cluster has corresponding vector feature data in the reference map, the annotation information of the map sample cluster can be annotated according to the natural language organized by the vector feature data.

[0051] Among them, the reference map may refer to a point cloud map, satellite map or three-dimensional map that has been constructed and verified. The reference map may be used as a standard or benchmark for matching and comparing with new map samples.

[0052] Vector feature data may refer to the data of a sub-image in a reference map that matches a map sample. For example, when the sub-image in a reference map that matches a map sample is a billboard, the vector feature data may include geometric information such as the landmark location and outline of the billboard. Vector feature data may be obtained from a reference map, or by matching key points in a map sample with corresponding points in a reference map using a matching algorithm, thereby extracting corresponding vector features.

[0053] Generating annotation information of a map sample based on vector feature data may refer to a process of automatically generating annotation information of a map sample by parsing the vector feature data. For example, when vector feature data of building A is obtained, and the vector feature data includes location information and use information of the building, the annotation information may be obtained by field extraction, for example, building A: location (100, 200), use "residential".

[0054] Therefore, by matching the map sample with the reference map to obtain the vector feature data at the corresponding position from the reference map, and automatically generating the annotation information of the map sample based on the vector feature data, the automatic annotation of the map sample is achieved, and the accuracy and efficiency of the annotation are improved.

[0055] The same annotation information of map samples in the same map sample cluster may mean that since the map sample cluster may have multiple photos, when each photo is used as a map sample, there is a situation where one annotation information corresponds to multiple photos. For example, when the object is a traffic light, the map sample cluster may include map samples of traffic lights taken from different angles or in different weather conditions, and the annotation information may refer to the description of the traffic light.

[0056] After obtaining the map sample cluster and annotation information, training data is constructed according to the map samples and annotation information in the map sample cluster, so that the data processing model can be trained. Since each annotation information may correspond to multiple map samples, during the model training process, features under multiple perspectives and conditions can be provided, thereby improving the generalization ability and robustness of the trained model, making the data processing model more effective in use.

[0057] S202: Filter out false positive data in the map difference data according to the similarity between the reference map vector and the map difference vector of the reference map data to obtain target difference data.

[0058] The reference map data may refer to pre-processed and annotated standard map data, which is used to identify changes in the update data by comparing the similarity between its vector and the map difference vector. In the embodiment of the present application, the data types of the reference map data can be divided into positive change data (TP) that can be identified, false positive data (FP) that is identified as an error, and data that cannot be identified (unkonwn).

[0059] The similarity between the reference map vector and the map difference vector may refer to the distance or matching degree between the reference map vector and the map difference vector in the feature space, and is used to measure their similarity. By comparing the similarity between the reference map vector and the map difference vector, it is possible to determine which of the map difference data are positive change data that can be identified, false alarm data that are identified as errors, and data that cannot be identified.

[0060] False positive data may represent data in the uploaded data that does not belong to the object described in the map data. For example, the object of the uploaded data is a billboard. When the billboard does not correspond to any billboard in the map data, the uploaded data is false positive data.

[0061] The target difference data may refer to identified positive change data and / or unidentifiable data, wherein when the target difference data is positive change data, the map can be automatically updated through crowdsourcing through the positive change data; when the target difference data is false alarm data, the false alarm data can be eliminated from the map difference data for subsequent update operations; when the target difference data is unidentifiable data, it can be further processed manually by transferring it to manual means.

[0062] Therefore, the data processing method provided in the embodiment of the present application can generate more accurate map difference vectors by inputting the map difference data into the trained data processing model. Subsequently, by comparing the similarity of these difference vectors with the existing control map vectors, it is determined whether the map difference reflects the real changes in the actual situation. Based on this judgment, it can be decided to take subsequent processing measures such as crowdsourcing update, elimination or manual review for the identified target difference data. This method not only improves the accuracy of map data update, but also effectively reduces the extra workload of users.

[0063] Figure 3 Schematic diagram of the data processing method provided for this application Figure 2 ,like Figure 3 As shown, in this embodiment Figure 2 Based on the embodiment, a method for training a data processing model is described in detail, and the method may further include:

[0064] S301, performing a first clustering process on the map samples according to the first similarity and the number of neighbors between the map sample vectors to obtain an initial sample cluster, where the map sample vector is a map vector of the map sample;

[0065] S302 . Perform a second clustering process on the map samples in the initial sample cluster according to the cluster features of the initial sample cluster to obtain a map sample cluster.

[0066] The first similarity may refer to the similarity between map sample vectors.

[0067] The number of neighbors refers to the number of nearest neighbor sample points selected for each sample point in the clustering algorithm. When performing the first clustering process based on the first similarity between map sample vectors, the number of neighbors determines the number of nearest neighbor samples considered for each map sample vector in the clustering process. By selecting an appropriate number of neighbors, it can be ensured that each sample point and its most similar neighbors form an initial sample cluster together, thereby better capturing the local structure and similarity of the data.

[0068] In the embodiment of the present application, by setting the requirements for the first similarity and the number of neighbors, the first clustering process can be performed on the map samples to obtain the initial sample clusters. When performing the first clustering process, in order to ensure that the map samples belonging to the same class of objects or the same object are classified into the same cluster, the requirement for the first similarity can be set to be smaller and the requirement for the number of neighbors can be set to be larger. Alternatively, the first clustering process can be directly performed as needed to obtain the sample clusters.

[0069] The same type of objects can refer to objects with similar shapes and types. That is, when clustering, map samples with similar shapes or similar image contents (for example, signboards of different types but similar shapes) can be classified as the same type of objects and preliminarily clustered to form initial sample clusters. Subsequently, when further clustering the map samples in these initial sample clusters, by increasing the requirements for similarity and number of neighbors, signboards representing the same type can be clustered more accurately, and finally a purer and more accurate sample cluster can be obtained.

[0070] Cluster features may refer to statistical or geometric properties used to describe the initial sample clusters in cluster analysis. In the embodiment of the present application, since the first clustering process is to ensure that map samples belonging to the same class of objects or the same object are classified into the same cluster as coarse clusters, in order to ensure the clustering effect and make the clustering effect of map samples in each map sample cluster better, the map samples may be subjected to a second clustering process through cluster features, thereby more finely dividing and optimizing the structure of the initial sample clusters and obtaining more reasonable map sample clusters.

[0071] In the embodiment of the present application, the cluster features of the initial sample cluster include the map sample quantity, non-core points, and core point similarity between core points in the initial sample cluster;

[0072] According to the cluster features of the initial sample cluster, the map samples in the initial sample cluster are subjected to a second clustering process to obtain a map sample cluster, including:

[0073] When the map sample size meets the first sample size requirement, or when there are non-core points in the initial sample cluster, or when the similarity of core points in the initial sample cluster meets the core point similarity requirement, a second clustering process is performed on the map samples in the initial sample cluster according to the second similarity to obtain a sample cluster, wherein the second similarity is greater than the first similarity.

[0074] The map sample size in the initial sample cluster may refer to the number of map samples in the initial sample cluster, and the sample size of the initial sample cluster may reflect the size of the cluster. The first sample size requirement may refer to the requirement for the size of the initial sample cluster, that is, when the map sample size in the initial sample cluster is greater than the first sample size, the map samples in the initial sample cluster are subjected to a second clustering process, wherein the sample size requirement may be set in advance by the user as needed.

[0075] A core point refers to a data point in cluster analysis that has a number of neighboring points exceeding a preset threshold in its neighborhood. When the number of other data points contained in a data point's specified neighborhood exceeds the minimum neighbor number threshold, the data point can be determined to be a core point. Conversely, when the number of other data points contained in a data point's specified neighborhood does not exceed a certain minimum neighbor number threshold, the data point can be determined to be a non-core point.

[0076] The core point similarity between core points may refer to the core point similarity between core points in cluster analysis. The core point similarity requirement may refer to the requirement for the mean value of the core point similarity, that is, when the mean value of the core point similarity is greater than the preset mean value, the map samples in the initial sample cluster are subjected to a second clustering process. In an embodiment of the present application, the mean value of the core point similarity may be obtained by calculating the average value of the similarity values ​​between all core points. For example, the similarity between each pair of core points in the initial sample cluster may be calculated, and then the mean value of the similarity between the core points may be obtained based on the similarity between each pair of core points.

[0077] In the embodiment of the present application, the mean similarity between core points can characterize the overall similarity between core points in the initial sample cluster. If the mean similarity between core points is high, it means that these core points are very close in the feature space, and the data points in the cluster are relatively consistent and compact. On the contrary, if the mean similarity is low, it means that the differences between core points are large, and the data points in the cluster are relatively dispersed.

[0078] Therefore, when the sample size in the initial sample cluster is large, there are non-core points, or the core point similarity meets the core point similarity requirements, it means that the map samples within the cluster may belong to different objects, so the initial sample cluster needs to be processed again.

[0079] Among them, the map samples in the initial sample cluster are subjected to the second clustering processing. In the process of obtaining the sample cluster, the second clustering processing can be performed in sequence on the three angles of map sample quantity, non-core points, and core points in the above scheme, or one or more of them can be selected to perform the second clustering processing.

[0080] In the embodiment of the present application, the initial sample clusters can be subdivided according to a second similarity higher than the first similarity. Thus, by subdividing these clusters, the local features and intrinsic patterns of the data can be captured more accurately, avoiding classifying data points of different properties into the same cluster, thereby improving the quality and interpretability of the clustering results.

[0081] In the embodiment of the present application, after performing a first clustering process on the map samples according to the first similarity and the number of neighbors between the map sample vectors to obtain the initial sample clusters, the method further includes:

[0082] When the map sample size does not meet the second sample size requirement, the initial sample clusters are merged according to the cluster center similarity between the cluster centers in the initial sample clusters to obtain the map sample clusters;

[0083] When there are discrete map samples representing discrete points in the map samples, the discrete map samples are merged into the initial sample cluster according to the vector similarity between the discrete map samples and the map samples in the initial sample cluster to obtain a map sample cluster;

[0084] When map samples belonging to the same object exist in multiple initial sample clusters, the initial sample clusters are merged according to the vector similarities between the map samples in the multiple initial sample clusters to obtain a map sample cluster.

[0085] Among them, the second sample size can be the same as the first sample size or different (the second sample size is smaller than the first sample size). When the sample size of the initial sample cluster is smaller than the second sample size, in order to ensure the clustering effect, the initial sample clusters can be merged according to the cluster center similarity between the cluster centers in the initial sample clusters.

[0086] The cluster center may refer to the core point with the highest similarity among the core points in the initial sample cluster.

[0087] Cluster center similarity can refer to the similarity between cluster centers of two different initial sample clusters. Cluster center similarity can be determined by Euclidean distance, cosine similarity, etc. For example, when cluster center similarity is determined by cosine similarity, cosine similarity satisfies:

[0088]

[0089] Among them, C a is the cluster center of the initial sample cluster a, C b is the cluster center of the initial sample cluster b.

[0090] In the embodiment of the present application, by comparing the similarity of two cluster centers, if their similarity exceeds a preset threshold, it indicates that the two clusters are very close in features and may belong to the same category or group, so the two initial sample clusters can be merged to obtain a sample cluster.

[0091] Discrete points can represent map samples that are not clustered into the initial sample clusters during clustering, that is, discrete map samples.

[0092] The vector similarity between the discrete map sample and the map sample in the initial sample cluster may refer to the similarity between the vector of the discrete map sample and the vector of the map sample in the initial sample cluster. By determining the vector similarity, incorporating the discrete map sample into the initial sample cluster may refer to, for a sample that has not yet been classified into any cluster, if there is a classified sample with a similarity exceeding a preset threshold, and these classified samples each belong to a different cluster, then adding the non-clustered sample to the cluster where the sample whose similarity meets the similarity requirement is located, for example, when there are two initial sample clusters that meet the similarity requirement, then adding the discrete map sample to the two initial sample clusters respectively.

[0093] The map samples belonging to the same object existing in multiple initial sample clusters may refer to the same map sample appearing in multiple different initial sample clusters. For example, photo A exists in both the first initial sample cluster and the second initial sample cluster.

[0094] According to the vector similarity between map samples in multiple initial sample clusters, the initial sample clusters are merged to obtain sample clusters, which can refer to determining the merging of initial sample clusters according to the maximum similarity between vectors of map samples in different initial sample clusters, wherein when the maximum similarity is greater than the maximum similarity threshold, the information representing the same object contained in the initial sample clusters is highly similar, so the initial sample clusters can be merged, thereby effectively reducing the cluster overlap problem caused by the uneven distribution of map samples, and making the clustering results more accurate and clear.

[0095] In an embodiment of the present application, a scheme for merging discrete map samples to obtain a map sample cluster and a scheme for merging initial sample clusters to obtain a map sample cluster can be executed in parallel with a scheme for performing a second clustering process on the map samples in the initial sample cluster, or can be executed sequentially or selectively based on a preset order according to the situation of the initial sample cluster.

[0096] S303, obtaining a first entropy value according to the similarity between the map sample vector and the text vector, and the true value of the similarity between the map sample vector and the text vector, wherein the map sample vector is a vector of a map sample in a map sample cluster, and the text vector is a vector of annotation information;

[0097] S304, obtaining a second entropy value according to the similarity between the text vector and the map sample vector and the true value of the similarity between the text vector and the map sample vector;

[0098] S305, determining a loss function of an initial model according to the first entropy value and the second entropy value;

[0099] S306. According to the loss function, the initial model is adjusted to obtain a data processing model.

[0100] The text vector may be a vector obtained by encoding the annotation information. When encoding, the annotation information may be input into a preset encoder to obtain the text vector. In some embodiments, the preset encoder may be an initial model encoder or other encoder for extracting text vectors.

[0101] The similarity between the map sample vector and the text vector, and the similarity between the text vector and the map sample vector can be determined by calculating the cosine similarity, Euclidean distance, etc. between the two vectors. The above similarity can be used to evaluate the matching degree between the map sample and the annotation information. Among them, the similarity between the map sample vector and the text vector can refer to the similarity between each map sample vector and the text vector, and the similarity between the text vector and the map sample vector can refer to the similarity between each text vector and the map sample vector.

[0102] In the embodiment of the present application, when the similarity is determined by cosine similarity, the similarity between the map sample vector and the text vector may satisfy:

[0103]

[0104] The similarity between the text vector and the map sample vector can satisfy:

[0105]

[0106] Among them, I i can be the i-th map sample vector, T i can be the i-th text vector, T i T Can be the transposed vector of the text vector, I i T Can refer to the transposed vector of the map sample vector.

[0107] In order to facilitate the processing of the similarity between the map sample vector and the text vector, and the similarity between the text vector and the map sample vector, the two similarities may be standardized before determining the first entropy value and the second entropy value, wherein the formula after the similarity between the map sample vector and the text vector is standardized may satisfy:

[0108]

[0109] The formula after normalizing the similarity between the text vector and the map sample vector can satisfy:

[0110]

[0111] Among them, x can be a map sample vector, x j can be the jth map sample vector, y can be a text vector, yj can be the jth text vector, s(x,y j ) can be the map sample vector x and the text vector y j The similarity score of s(y,x j ) can be the text vector y and the map sample vector x j The similarity score of , t can be an adjustable parameter.

[0112] The true value of the similarity between a map sample vector and a text vector, and the true value of the similarity between a text vector and a map sample vector may both refer to numerical values ​​indicating whether the map sample vector and the text vector are similar in actual circumstances. In this embodiment of the present application, when the annotation information corresponding to the text vector is used to describe the map sample corresponding to the map sample vector, the true value of the similarity between the map sample vector and the text vector is 1; when the annotation information corresponding to the text vector is not used to describe the map sample corresponding to the map sample vector, the true value of the similarity between the map sample vector and the text vector is 0.

[0113] The entropy value can characterize the difference between the similarity and the true value of the similarity, wherein the first entropy value can refer to the similarity between the map sample vector and the text vector, and the difference between the true value of the similarity between the map sample vector and the text vector, and the second entropy value can refer to the similarity between the text vector and the map sample vector, and the difference between the true value of the similarity between the text vector and the map sample vector.

[0114] After determining the similarity between the map sample vector and the text vector, and the similarity between the text vector and the map sample vector, a loss function may be determined to train the model.

[0115] Among them, the loss function may refer to a function used in machine learning and statistics to evaluate the difference between the model prediction value and the actual value. It quantifies the degree of error in the model prediction and is usually defined as a non-negative value. The smaller the value, the closer the model's prediction result is to the actual result. In some embodiments, the loss function may include mean square error (MSE) and cross-entropy loss.

[0116] In the embodiment of the present application, when the loss function is the cross entropy loss, the loss function of the initial model is determined to satisfy the following conditions based on the first entropy value and the second entropy value:

[0117] L = 0.5*(cross_entropy(p x2y (x),q x2y (x))+

[0118] cross_entropy(p y2x (y),q y2x (y)));

[0119] Among them, p x2y (x) can refer to the similarity between the map sample vector and the text vector, p y2x (y) can refer to the similarity between the text vector and the map sample vector, q x2y (x) can refer to the true value of the similarity between the map sample vector and the text vector, q y2x (y) may refer to the true value of the similarity between the text vector and the map sample vector.

[0120] Among them, since there is at least one annotation information for describing more than two map samples, in order to reflect the relationship between the map sample vector and the text vector, in the embodiment of the present application, the KL divergence can be selected to determine the loss function, wherein the loss function of the initial model can satisfy:

[0121] L=0.5*(KL(p x2y (x),q x2y (x))+KL(p y2x (y),q y2x (y)));

[0122] Among them, L represents the loss function of the initial model; p x2y (x) represents the similarity between the text vector and the map sample vector, q x2y (x) represents the true value of the similarity between the text vector and the map sample vector; p y2x (y) represents the similarity between the text vector and the map sample vector, q y2x (y) represents the true value of the similarity between the text vector and the map sample vector.

[0123] Therefore, when using KL divergence to deal with multiple positive samples in contrastive learning, it can better measure the difference between the predicted distribution and the true distribution of multiple positive samples, provide more stable gradient information, and make full use of the information of all positive samples. Therefore, compared with the cross entropy loss function, it can more comprehensively optimize the model performance.

[0124] In an embodiment of the present application, the initial model may be a model for annotating images and text information, and the model may include an image encoder and a text encoder, wherein the image encoder is used to extract map sample vectors from map samples, and the text encoder is used to extract text vectors of annotated information. In an embodiment of the present application, the initial model may be a CLIP (Contrastive Language-Image Pre-Training) model, and the CLIP model may combine natural language supervision signals and use contrastive learning to extract vectorized representations of documents and images.

[0125] After determining the loss function, the initial model can be adjusted according to the loss function to obtain a data processing model. Among them, after determining the loss function, the parameters of the initial model can be adjusted by applying an optimization algorithm (such as stochastic gradient descent, Adam, etc.) to minimize the loss function. For example: the loss between the model prediction output and the actual label can be calculated first; then the gradient of the parameters is calculated based on the loss function, that is, the partial derivative of the loss with respect to each parameter; then the model parameters are updated in the direction of the negative gradient, and the learning rate is usually used to control the update step size. Repeat the above process, gradually adjust the model weights and biases, so that the model can better fit the training data and reduce the prediction error. Therefore, by continuously adjusting the initial model, a fully trained data processing model can be obtained, which can provide accurate and robust prediction performance on new data.

[0126] In the embodiment of the present application, both the map sample vector and the text vector can be vectors obtained after being input into the initial model. Therefore, by using both the map sample vector and the text vector as vectors obtained after being input into the initial model, the fusion of multiple sample data can be realized, thereby improving the richness and accuracy of feature representation and enhancing the prediction performance and generalization ability of the model.

[0127] Figure 4 Schematic diagram of the data processing method provided for this application Figure 3 ,like Figure 4 As shown, in this embodiment Figure 2 Based on the embodiment, a method for determining target difference data in map difference data according to the similarity between the map image vector and the reference map vector is described in detail. The method may further include:

[0128] S401. Determine initial difference data in the map difference data that meets a preset similarity requirement and target comparison data in the comparison map data that meets a preset similarity requirement according to the similarity between the map image vector and the comparison map vector.

[0129] The preset similarity requirement can be used to determine the initial difference data and target reference map data that are similar to the map difference data and the reference map data, wherein the preset similarity requirement can refer to being lower than a first similarity threshold and being higher than a second similarity threshold, wherein when the similarity between the map image vector and the reference map vector is lower than the first similarity threshold, it can be represented that it is impossible to determine whether the map difference data and the reference map data belong to the same object, and the first initial difference data cannot be determined. Being higher than the second similarity threshold can refer to the second initial difference data that can be determined and has a high similarity with the object to which the target reference map data belongs. Data between the first similarity threshold and the second similarity threshold can be determined as low similarity data.

[0130] In the embodiment of the present application, the initial difference data includes the first initial difference data and / or the second initial difference data, wherein the first initial difference data cannot be determined whether it is identifiable and correctly changed data, so the staff needs to make a judgment again to determine whether to update the existing map, and the second initial difference data is related to the control cluster where the control map data is located, so it can be directly determined to be identifiable and correctly changed data, and the data of the existing map can be directly updated. The data between the first similarity threshold and the second similarity threshold can be determined to be irrelevant to the control cluster where the control map data is located, so it can be processed manually in the form of error reporting.

[0131] In the embodiment of the present application, the first similarity threshold and the second similarity threshold can be set by the staff as needed.

[0132] S402: When the target comparison data is comparison map data belonging to the same comparison cluster, determine the target difference data in the initial difference data according to the data volume of the target comparison data and the type of the comparison cluster.

[0133] The control cluster may refer to a cluster obtained by clustering the control map data, wherein the control map data in different control clusters correspond to different objects. In the embodiment of the present application, the target control data is the control map data belonging to the same control cluster, which may refer to the target control data belonging to the same control cluster.

[0134] In the embodiment of the present application, according to the data volume of the target comparison data and the type of the comparison cluster, determining the target difference data in the initial difference data may include:

[0135] If the data amount of the target control data in the control cluster is less than the preset data amount, determining the initial difference data as the target difference data;

[0136] If the data volume of the target comparison data is greater than or equal to the preset data volume, the target difference data in the initial difference data is determined according to the type of the comparison cluster.

[0137] Among them, the data amount of the target control data in the control cluster is less than the preset data amount, which may mean that it is impossible to determine whether the target object corresponding to the initial difference data can be identified through the control cluster, so the initial difference data needs to be determined as the target difference data that cannot be determined whether it is identifiable.

[0138] The data volume of the target comparison data being greater than or equal to the preset data volume may indicate that the initial difference data is similar to the comparison cluster, so that the target difference data meeting the type requirement in the initial difference data may be determined according to the type of the comparison cluster.

[0139] The preset data volume can be used to determine whether the control cluster is related to the target control data. For example, when there are 10 photos in the control cluster that meet the similarity threshold with the initial difference data, and the preset data volume requirement is 5, it can be determined that the control cluster is related to the initial difference data. Thus, according to the type of the control cluster, the target difference data in the initial difference data is determined. When there are 3 photos in the control cluster that meet the similarity threshold with the initial difference data, it can be determined that the control cluster is not related to the initial difference data, and the initial difference data is determined to be target difference data that cannot be determined to be identifiable.

[0140] In the embodiment of the present application, according to the type of the reference cluster, the type of the initial difference data can also be divided into: 1. positive change data (TP) that can be identified; 2. false alarm data (FP) that is identified as an error; 3. and unidentifiable data (unkonwn), wherein, when the type of the initial difference data is positive change data (TP) that can be identified and / or unidentifiable data (unkonwn), the initial difference data is the target difference data.

[0141] S403. When the target control data is control map data belonging to different control clusters and the data volume of the target control data in the control cluster meets the preset data volume requirement, the target difference data in the initial difference data is determined according to the type of the target control cluster in the control cluster and the data volume of the target control data in the control cluster.

[0142] Among them, since there may be multiple control clusters, when performing comparison, there may be target control data in multiple control clusters that meet the similarity requirements with the initial difference data. Therefore, the target difference data in the initial difference data can be determined based on the data volume of the target control data in the multiple control clusters and the type of the control clusters.

[0143] Among them, the target control cluster with the largest amount of data in the control cluster can be determined first according to the amount of data of the target control data in the control cluster, and then the target difference data in the initial difference data can be determined according to the type of the target control cluster. For example, when there are 10 photos in the first control cluster that meet the similarity threshold with the initial difference data, and there are 7 photos in the second control cluster that meet the similarity threshold with the initial difference data, and the preset data amount requirement is 5 photos, then the first control cluster can be determined as the target control cluster, and the second control cluster can be determined as other control clusters, and the target difference data in the initial difference data can be determined according to the type of the target control cluster.

[0144] In this embodiment of the present application, the control cluster is a map sample cluster; and / or the number of objects corresponding to the control map data in the control cluster meets the requirement of the preset number of objects, and the proportion of the control map data belonging to the same type meets the preset proportion requirement.

[0145] In order to ensure that the control cluster has universality and high confidence, the number of objects in the control cluster must meet the object number requirement, and the proportion of control map data of the same type in the control cluster must meet the preset proportion requirement.

[0146] The objects in the control cluster may refer to the shooting targets of the photos in the control cluster, and the number of objects in the control cluster may refer to the number of shooting targets. For example, when the shooting targets include billboards and traffic signs, the number of objects in the control cluster is 2.

[0147] The proportion of the control map data belonging to the same type in the control cluster may refer to whether the proportion of TP, FP or unkonwn data is greater than a threshold.

[0148] In the embodiment of the present application, the control cluster may be a cluster obtained after the model performs clustering processing on images in the historical map data, or may be a map sample cluster obtained through steps S301 to S302 in the embodiment of the present application.

[0149] Among them, when the type of the target difference data is positive change data, in order to further ensure the accuracy of the target difference data identification and avoid the occurrence of erroneous data, in the embodiment of the present application, after determining the target difference data in the map difference data according to the similarity between the map image vector and the reference map vector, the method further includes:

[0150] Determine a target error object that meets the distance requirement among the error objects according to the position of the target object corresponding to the target difference data in the map and the position of the error object in the historical map;

[0151] When the vector similarity between the historical map data of the target error object and the target difference data meets the first vector similarity requirement, adjusting the type of the target difference data to false alarm data;

[0152] and / or,

[0153] According to the position and vector features of the target object in the map, determine the map object that meets the position distance requirements and vector feature requirements in the map;

[0154] When the vector similarity between the map data of the map object and the target difference data meets the second vector similarity requirement, the type of the target difference data is adjusted to false positive data.

[0155] Among them, error objects can refer to objects that have been marked as non-positive change data in the previous data processing process. These objects are usually identified by previous clustering analysis or other methods and are considered to be features that are not included in the map or non-map features.

[0156] In data processing, by comparing the positions of these error objects, we can find the target error object that meets the distance requirement among the error objects, and by comparing the vector similarity between the historical map data of the target error object and the target difference data, we can effectively reduce the recurrence of the same type of errors. That is, if the target object is very similar to the error object in position and appearance, then the target object may also be an error object, so its category can be marked as false alarm data. In this way, the accuracy and efficiency of data processing can be improved and repeated errors can be avoided.

[0157] In addition, in order to reduce the number of errors caused by the reported position and attribute accuracy errors that do not match the existing map elements, the judgment of the map object is added. By judging whether the object corresponding to the target difference data appears repeatedly in the map, the false positive data in the target difference data is determined.

[0158] The map object may refer to an object on the map whose position meets the position distance requirement and whose vector features are consistent with the target object. After determining the map object, if the map object and the target object are the same, the data representing the target object is duplicated marked data, so the type of the target difference data can be adjusted to false positive data.

[0159] For example, a street lamp on a road appears once within a radius of 10m. If other street lamps (map objects) appear within 10m of the target object, and if the vector similarity between the map data of the street lamp and the target difference data meets the second vector similarity requirement, it indicates that the target difference data is repeated false alarm data.

[0160] In the embodiment of the present application, the execution order and execution method of the technical solution A1 for determining whether the vector similarity between the historical map data of the target error object and the target difference data meets the first vector similarity requirement, and the technical solution A2 for determining whether the vector similarity between the map data of the map object and the target difference data meets the second vector similarity requirement can be arbitrary. For example, only the technical solution A1 or the technical solution A2 can be executed; the technical solution A1 can be executed first, and then the technical solution A2; the technical solution A2 can be executed first, and then the technical solution A1; and the technical solution A1 and the technical solution A2 can be executed simultaneously.

[0161] The data processing method provided in the embodiment of the present application improves the accuracy of map updates by sending target difference data to a trained data processing model to obtain a more accurate map image vector. At the same time, by combining multiple data sources of control clusters, error objects, and map objects for joint judgment, the number of classifiable individuals is further increased while ensuring the accuracy of map updates, thereby improving manual work efficiency.

[0162] Figure 5 A flowchart of another data processing method provided in this application is shown below. Figure 5a A schematic diagram of the sample pair construction method provided for this application, Figure 5b A schematic diagram of the process of clustering images provided in this application, Figure 5c A schematic diagram of the effective cluster provided for this application, Figure 5d A schematic diagram of the cluster classification process provided for this application.

[0163] like Figure 5a As shown, the method includes:

[0164] S501: Modeling stage.

[0165] Among them, the modeling stage includes: image representation model training and cluster analysis.

[0166] Image representation model training may refer to: training a CLIP model for acquiring a map domain, wherein the obtained image representation model is used to extract image features, and samples used for training may include photos of individual elements to be predicted, individual elements with historical changes, and individual elements already on the map.

[0167] Among them, the CLIP model can be used to obtain the vector representation of the image, which is used to calculate the similarity between images; the CLIP model can also use the mother library data through multimodal learning to quickly obtain the corresponding text description of the individual, and construct image-text pairs to train the model, thereby avoiding additional manual labeling work; in cases such as text-to-image retrieval and scene retrieval, the CLIP model can effectively solve the problem of establishing associations between multimodal images and text.

[0168] Among them, when using the CLIP model for training, in order to better represent the characteristics of map elements, map domain elements are added to the general domain CLIP model for tuning, and the data set used can be composed of two parts: the image side and the text side.

[0169] When obtaining the content on the image side, the original photo of the reported result in the user end can be cropped by using the detection frame on the user end, thereby obtaining the input on the image side.

[0170] When obtaining the content on the text side, for the reported results that match the reference map, the vector features at the corresponding positions on the reference map can be organized into natural language, which can be used to quickly generate large-scale annotations. For the reported results that do not match the reference map, the CLIP model can be used to obtain image representations, cluster photos, and annotate photos in the same cluster using the same manual annotation method.

[0171] Therefore, through the above two marking methods, the present application greatly improves the marking efficiency compared to the method of marking individual photos one by one.

[0172] In the embodiment of the present application, after obtaining the data set for training, the loss function can be determined by constructing sample pairs, where:

[0173] Figure 5a A schematic diagram of the sample pair construction method provided for this application, such as Figure 5a As shown, in the prior art, when setting samples for training the CLIP model, each annotation content corresponds to only one image, that is, Figure 5a As shown in the annotation a in , in the sample pair matrix constructed in the current data set that needs to be batched, only the sample pairs on the diagonal of the matrix are positive samples, and the rest are negative samples. Therefore, when determining the loss function, only the cross entropy loss function is used to calculate the difference between the predicted similarity matrix and the labels of only the diagonal positive samples.

[0174] In the embodiment of the present application, one text corresponds to multiple images, that is, Figure 5a As shown in the annotation b in , the matrix does not only have positive samples on the diagonal. Therefore, the KL divergence can handle the situation of multiple positive samples, so that it can better reflect the relationship between the similarity matrix and multiple positive sample annotations, thereby improving the accuracy of model training.

[0175] Cluster analysis can be the process of clustering and labeling sample images when training an image representation model. In an embodiment of the present application, the cluster analysis process may include: image feature extraction, similarity calculation, clustering, and cluster classification.

[0176] Herein, image feature extraction may refer to extracting a vector representation of an image using a CLIP model.

[0177] Similarity calculation may refer to calculating the similarity between images using vector representation and cosine similarity.

[0178] After determining the similarity between images, the images can be clustered. In the embodiment of the present application, the DBSCAN clustering algorithm can be used for clustering. Figure 5bAs shown in the figure, the process includes: 1. Coarse clustering: Based on the similarity between images, the DBSCAN clustering algorithm is used. In order to make the number of photos in the clusters obtained by clustering as large as possible, a lower similarity and a larger number of neighbors can be selected. 2. Cluster feature extraction: Extract the features of the clusters obtained by coarse clustering to determine whether the clusters need to be subdivided. The features include: the number of samples in the cluster, the number of core points, the number of edge points, the mean similarity between core points, the mean similarity between edge points, and the mean similarity between core points and edge points. 3. Cluster subdivision: When the number of samples in the cluster is less than the threshold, or all samples in the cluster are core points, or the mean similarity between core points in the cluster is greater than the similarity threshold, there is no need to subdivide the cluster. If the cluster needs to be refined, the similarity threshold can be increased, and the original cluster can be clustered again using the DBSCAN clustering algorithm to obtain the subdivision results. 4. Cluster merging: For clusters with a small sample size, determine whether they can be merged with other clusters. The judgment method includes: obtaining the cluster center of the cluster, where the cluster center is the core point with the highest similarity between core points in the current cluster. Calculate the similarity between cluster centers. If the current cluster has a cluster with a similarity exceeding the threshold, merge it with the cluster with the most similarity. Otherwise, do not merge it. 5. Clustering judgment: For unclustered samples, if there is a sample with a similarity exceeding the threshold, and the sample has its corresponding cluster, add the current unclustered sample to the cluster with the sample with the greatest similarity. Otherwise, the current sample still does not belong to any cluster. 6. Individual merging: If the photos of an individual exist in multiple clusters, if the maximum similarity of the photos of individuals in two clusters exceeds the threshold, merge the two clusters.

[0179] Cluster classification refers to judging the clusters obtained after cluster analysis. When the number of individuals in the cluster obtained after cluster analysis is greater than the threshold, and the TP ratio or FP ratio is greater than the threshold, the cluster is determined to be a valid cluster that meets the confidence requirements and high-frequency requirements. Figure 5c The merged clusters can be labeled for model training, and the confirmed valid clusters can be compared with the cluster classification in the prediction stage.

[0180] S502: prediction stage.

[0181] Among them, the prediction stage includes: obtaining changing elements and status, cluster classification, historical sample similarity and map deduplication.

[0182] Among them, Figure 5dAs shown in the figure, the cluster classification process may include: 1. Low similarity filtering: Filter the samples whose similarity between the photos under the individual and the valid clusters is lower than the similarity threshold. If there is no cluster with similarity to the threshold under the current individual after filtering, the classification of the current individual cannot be confirmed based on the existing valid clusters, and its category is classified as unknown. 2. Count the proportion of related cluster categories: Take photos as units, and count the category proportions of related photos in all related clusters under the individual. If the proportion of TP or FP in the related photos is less than the threshold, it is considered that there is confusion between the current individual and the TP and FP clusters of the known valid clusters. To ensure accuracy, such individuals are classified as unknown. 3. Determine the cluster: To ensure accuracy, only clusters with related sample quantities greater than the threshold are retained in the same cluster. If there is no related cluster after filtering, it is considered that the correlation between the current individual and the related cluster is not significant, and it is classified as unknown. If there are multiple clusters after filtering, the current individual is assigned to the cluster with the highest similarity. 4. Determine the category: The category of the valid cluster to which the current individual belongs is used as the category of the current individual. After cluster classification, the samples can be divided into TP, FP and unknown.

[0183] The similarity process of historical samples can include: 1. Calculate the distance between the current non-FP individual and the historical FP individual. The distance within a certain range is regarded as a neighboring point. If there is no neighboring point, the category of the current non-FP individual remains unchanged. 2. Similarity calculation: If there is a neighboring point, calculate the similarity of all photos under the current non-FP individual and all photos under the neighboring point. 3. Individual classification: If the maximum similarity between the current individual and the neighboring point is greater than the similarity threshold, it is considered that the current non-FP individual is similar to the historical FP individual, and the category of the current non-FP individual is assigned to FP.

[0184] The map deduplication process may include: 1. Acquisition of adjacent maps: Calculate the distance between the current non-FP individual and the map individual. Figure 1 The ones that are consistent are regarded as adjacent points. For different elements, the compared vector features are different. For example, for signs, compare whether the shapes of the current non-FP individual and the map individual are the same. For arrows, compare whether the subdivision types of the current non-FP individual and the map individual are consistent. 2. Similarity calculation: If there are adjacent points, calculate the similarity between the current non-FP individual and the adjacent map individual. 3. Duplicate determination: Calculate the similarity of all photos under the current non-FP individual and all photos of the adjacent points on the map. If the maximum similarity is greater than the similarity threshold, it is considered that the current non-FP individual has a similar map individual, and the category of the current non-FP individual is assigned to FP.

[0185] In another data processing method provided in an embodiment of the present application, the adaptation of the CLIP model in the map field includes the generation method of training samples and the construction method of positive samples used in contrastive learning to calculate losses. Compared with the CLIP model in the general field, the model after adding map element tuning has improved the representation of map elements, and the distinction between individuals of different categories is higher. In addition, compared with single-step DBSCAN clustering, multi-step clustering obtains higher homogeneity of elements in the cluster, which improves the precision and recall of subsequent classification tasks. Therefore, when predicting, by combining cluster classification, historical sample similarity and map samples, the accuracy is ensured while increasing the number of classifiable individuals, thereby improving the efficiency of manual work.

[0186] As shown in the following table:

[0187]

[0188] For data with predicted categories, the precision recall of TP reached 97%, the precision recall of FP reached 96%, and the rest of the data was unknown. The overall coverage of TP was 12.09%, and the overall coverage of FP was 19.32%. After adding classification recognition, the overall data judgment was reduced by about 14.08%. For map crowdsourcing updates, this not only improves the accuracy of existing change data, but also is expected to reduce 14% of manual workload.

[0189] Figure 6 A schematic diagram of the structure of the data processing device provided in this application, such as Figure 6 As shown, the data processing device 60 provided in this embodiment includes:

[0190] An input module 601 is used to input the map difference data into a data processing model to obtain a map difference vector of the map difference data; the data processing model is trained according to map sample clusters and annotation information of the map sample clusters, and the map sample clusters are obtained by clustering the map samples, wherein at least one map sample cluster includes more than two map samples;

[0191] The determination module 602 is used to determine the target difference data in the map difference data according to the similarity between the reference map vector and the map difference vector of the reference map data.

[0192] In a possible implementation, the input module 601 may also be specifically used for:

[0193] According to the first similarity and the number of neighbors between the map sample vectors, a first clustering process is performed on the map samples to obtain an initial sample cluster, where the map sample vector is a map vector of the map sample;

[0194] According to the cluster features of the initial sample cluster, a second clustering process is performed on the map samples in the initial sample cluster to obtain a map sample cluster.

[0195] In a possible implementation, the cluster features of the initial sample cluster include the map sample quantity, non-core points, and core point similarity between core points in the initial sample cluster; the input module 601 may also be specifically used for:

[0196] When the map sample size meets the first sample size requirement, or when there are non-core points in the initial sample cluster, or when the similarity of core points in the initial sample cluster meets the core point similarity requirement, a second clustering process is performed on the map samples in the initial sample cluster according to the second similarity to obtain a sample cluster, wherein the second similarity is greater than the first similarity.

[0197] In a possible implementation, the input module 601 may also be specifically used for:

[0198] When the map sample size does not meet the second sample size requirement, the initial sample clusters are merged according to the cluster center similarity between the cluster centers in the initial sample clusters to obtain the map sample clusters;

[0199] When there are discrete map samples representing discrete points in the map samples, the discrete map samples are merged into the initial sample cluster according to the vector similarity between the discrete map samples and the map samples in the initial sample cluster to obtain a map sample cluster;

[0200] When map samples belonging to the same object exist in multiple initial sample clusters, the initial sample clusters are merged according to the vector similarities between the map samples in the multiple initial sample clusters to obtain a map sample cluster.

[0201] In a possible implementation, the input module 601 may also be specifically used for:

[0202] According to the similarity between the map sample vector and the text vector, and the true value of the similarity between the map sample vector and the text vector, a first entropy value is obtained, where the map sample vector is a vector of the map sample in the map sample cluster, and the text vector is a vector of the annotation information;

[0203] Obtaining a second entropy value according to the similarity between the text vector and the map sample vector and the true value of the similarity between the text vector and the map sample vector;

[0204] Determine a loss function of the initial model according to the first entropy value and the second entropy value;

[0205] According to the loss function, the initial model is adjusted to obtain the data processing model.

[0206] In a possible implementation, the determining module 602 may also be specifically configured to:

[0207] Determining, according to the similarity between the map image vector and the reference map vector, initial difference data in the map difference data that meets a preset similarity requirement, and target reference map data in the reference map data that meets a preset similarity requirement;

[0208] When the target comparison map data is comparison map data belonging to the same comparison cluster, determining the target difference data in the initial difference data according to the data volume of the target comparison map data and the type of the comparison cluster;

[0209] When the target control map data is control map data belonging to different control clusters and the data volume of the target control map data in the control cluster meets the preset data volume requirement, the target difference data in the initial difference data is determined according to the type of the target control cluster in the control cluster and the data volume of the target control map data in the control cluster.

[0210] In a possible implementation, when the type of the target difference data is positive change data, the determination module 602 may also be specifically configured to:

[0211] Determine a target error object that meets the distance requirement among the error objects according to the position of the target object corresponding to the target difference data in the map and the position of the error object in the historical map;

[0212] When the vector similarity between the historical map data of the target error object and the target difference data meets the first vector similarity requirement, adjusting the type of the target difference data to false alarm data; and / or, determining the map object in the map that meets the position distance requirement and the vector feature requirement according to the position and vector feature of the target object in the map;

[0213] When the vector similarity between the map data of the map object and the target difference data meets the second vector similarity requirement, the type of the target difference data is adjusted to false positive data.

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

[0215] Figure 7 This is a schematic diagram of the structure of the electronic device provided in this application. Figure 7As shown, the electronic device 70 provided in this embodiment includes: at least one processor 701 and a memory 702. Optionally, the device 70 also includes a communication component 703. Among them, the processor 701, the memory 702 and the communication component 703 are connected via a bus 704. In the specific implementation process, at least one processor 701 executes the computer execution instructions stored in the memory 702, so that at least one processor 701 executes the above method. The specific implementation process of the processor 701 can be referred to the above method embodiment, and its implementation principle and technical effect are similar, and this embodiment will not be repeated here.

[0216] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in the invention may be directly embodied as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor. The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory. The bus may be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the buses in the drawings of the present application are not limited to only one bus or one type of bus.

[0217] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0218] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

Claims

1. A data processing method, characterized in that: include: Inputting the map difference data into a data processing model to obtain a map difference vector of the map difference data; the data processing model is trained based on map sample clusters and annotation information of the map sample clusters, the map sample clusters are obtained by clustering map samples, wherein the map samples in the same map sample cluster have the same annotation information, and the map difference data is determined based on the difference between the uploaded data and the map data; According to the similarity between the reference map vector of the reference map data and the map difference vector, false positive data in the map difference data is screened out to obtain target difference data.

2. The method according to claim 1, characterized in that The method further comprises: According to the first similarity and the number of neighbors between the map sample vectors, a first clustering process is performed on the map samples to obtain an initial sample cluster, wherein the map sample vector is a map vector of the map sample; According to the cluster features of the initial sample cluster, a second clustering process is performed on the map samples in the initial sample cluster to obtain a map sample cluster.

3. The method according to claim 2, characterized in that The cluster features of the initial sample cluster include the map sample quantity, non-core points, and core point similarity between core points in the initial sample cluster; The step of performing a second clustering process on the map samples in the initial sample cluster according to the cluster features of the initial sample cluster to obtain a map sample cluster includes: When the map sample quantity meets the first sample quantity requirement, or when the non-core point exists in the initial sample cluster, or when the core point similarity in the initial sample cluster meets the core point similarity requirement, a second clustering process is performed on the map samples in the initial sample cluster according to a second similarity to obtain a sample cluster, wherein the second similarity is greater than the first similarity.

4. The method according to claim 2, characterized in that: After performing a first clustering process on the map samples according to the first similarity and the number of neighbors between the map sample vectors to obtain an initial sample cluster, the method further includes: When the map sample size does not meet the second sample size requirement, merging the initial sample clusters according to the cluster center similarities between the cluster centers in the initial sample clusters to obtain a map sample cluster; When there are discrete map samples representing discrete points in the map samples, according to the vector similarity between the discrete map samples and the map samples in the initial sample cluster, the discrete map samples are merged into the initial sample cluster to obtain a map sample cluster; When map samples belonging to the same object exist in multiple initial sample clusters, the initial sample clusters are merged according to the vector similarities between the map samples in the multiple initial sample clusters to obtain a map sample cluster.

5. The method according to claim 1, characterized in that Before inputting the map difference data into the data processing model to obtain the map difference vector of the map difference data, the method further includes: Obtaining a first entropy value according to the similarity between a map sample vector and a text vector and a true value of the similarity between the map sample vector and the text vector, wherein the map sample vector is a vector of a map sample in the map sample cluster, and the text vector is a vector of the annotation information; Obtaining a second entropy value according to the similarity between the text vector and the map sample vector and a true value of the similarity between the text vector and the map sample vector; Determining a loss function of the initial model according to the first entropy value and the second entropy value; According to the loss function, the initial model is adjusted to obtain a data processing model.

6. The method according to claim 1, characterized in that The annotation information is obtained in response to the user's annotation processing of the map sample cluster; Or, generated according to the vector feature data of the map sample in the reference map.

7. The method according to any one of claims 1 to 6, characterized in that The step of determining target difference data in the map difference data according to the similarity between the map image vector and the reference map vector comprises: Determining, according to the similarity between the map image vector and the reference map vector, initial difference data in the map difference data that meets a preset similarity requirement, and target reference map data in the reference map data that meets the preset similarity requirement; When the target comparison map data is comparison map data belonging to the same comparison cluster, determining the target difference data in the initial difference data according to the data volume of the target comparison map data and the type of the comparison cluster; When the target control map data is control map data belonging to different control clusters and the data volume of the target control map data in the control cluster meets the preset data volume requirement, the target difference data in the initial difference data is determined according to the type of the target control cluster in the control cluster and the data volume of the target control map data in the control cluster.

8. The method according to any one of claims 1 to 6, characterized in that When the type of the target difference data is positive change data, After determining the target difference data in the map difference data according to the similarity between the map image vector and the reference map vector, the method further includes: Determine, according to the position of the target object corresponding to the target difference data in the map and the position of the error object in the historical map, the target error object that meets the distance requirement among the error objects; When the vector similarity between the historical map data of the target error object and the target difference data meets a first vector similarity requirement, adjusting the type of the target difference data to false alarm data; and / or, Determining a map object in the map that meets the position distance requirement and the vector feature requirement according to the position and vector feature of the target object in the map; When the vector similarity between the map data of the map object and the target difference data meets a second vector similarity requirement, the type of the target difference data is adjusted to false positive data.

9. The method according to any one of claims 1 to 6, characterized in that: The control cluster is the map sample cluster; and / or, The number of objects corresponding to the reference map data in the reference cluster meets a preset object number requirement, and the proportion of reference map data belonging to the same type meets a preset proportion requirement.

10. An electronic device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 9.

Images