A driving scene data mining method

Abstract

The application relates to a driving scene data mining method, which comprises the following steps: S10. An initial model and a to-be-screened data pool are acquired, and forward prediction is performed on the to-be-screened data pool; S20. Data mining is performed on the prediction result, and a value score corresponding to each sample is obtained; S30. Value samples in the to-be-screened data pool are sorted according to the value score to obtain a value sample set; S40. Part of the value samples in the value sample set are selected as a new data set of this round of iteration, and a model of this round of iteration is trained; S50. The remaining value samples in the value sample set are taken as an iteration to-be-screened data pool, and forward prediction is performed on the iteration to-be-screened data pool; S60. Step S20 is returned until only valueless samples or samples with very small values are left in the iteration to-be-screened data pool, and a final iteration model and a final iteration data set are obtained. The data mining process of the application uses an active learning method to perform model iteration and data set expansion, and the model iteration speed is accelerated, and the model can reach higher accuracy under the same data amount.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining, in particular to a driving scene data mining method. BACKGROUND

[0002] In recent years, with the advent of the big data era shown by various statistical reports, the development of new data acquisition and transmission technologies has brought exponential growth of data, and relevant research shows that data can accelerate model iteration, and the mass production and landing deployment of models both depend on large data sets.

[0003] The dependence of models on large data sets has brought challenges to labels. In order to accelerate the model iteration speed and meet the project budget and time budget, high demands are put forward for the quality and quantity of labels, and the acquisition time and budget cost of some types of data and labels are high, such as:

[0004] (1) Special working condition scenes where data is difficult to obtain, such as data at the moment of a car accident;

[0005] (2) High data acquisition cost or sensor cost, such as radar point cloud;

[0006] (3) Time-consuming label acquisition, such as panoramic semantic segmentation.

[0007] Traditional algorithms are based on passive learning to do iteration, and each time the data set to be calibrated is randomly sampled; relevant research shows that compared with active learning which mines data according to data value each time, traditional algorithms need to use 40% to 50% more samples to reach the same accuracy as active learning.

[0008] The core part of active learning is the data mining algorithm. Some mining methods are based on feature dimensionality reduction analysis, which needs to invade the model inside or modify the output interface of the model, is not user-friendly, has low platform reusability, and has high algorithm complexity; some mining methods based on shadow mode in the autonomous driving scene need to be used after the algorithm is deployed on the vehicle side, and the whole data link of the perception layer, decision layer and control layer needs to be opened, which cannot be used for early-stage algorithm development alone, and has high cost and long cycle; in addition, some data mining algorithms are designed for specific algorithms and cannot be migrated to other tasks at a small cost, and have low universality. SUMMARY

[0009] The present application provides a driving scene data mining method to solve the problems of excessive dependence on big data sets, difficulty in algorithm landing, exceeding project budget, how to efficiently use big data, reduce GPU and CPU consumption, reduce storage resource consumption, how to automatically mine Hardcase and Badcase to help automatically build regression test ISSUE set, how to automatically mine vehicle data, realize online training of model, how to automatically mine cloud data, reduce data collection demand, and how to develop efficient and universal data mining algorithm.

[0010] To solve the above technical problems, the technical scheme of the present application is as follows:

[0011] A driving scene data mining method comprises the following steps:

[0012] S10. An initial model and a to-be-screened data pool are obtained, and the initial model is used to make forward prediction on the to-be-screened data pool;

[0013] S20. A data mining algorithm is used to mine the prediction results to obtain a value score corresponding to each sample;

[0014] S30. The value samples in the to-be-screened data pool are sorted according to the value scores to obtain a value sample set;

[0015] S40. Part of the value samples in the value sample set are selected as a new data set of the current iteration, and a model of the current iteration is trained;

[0016] S50. The remaining value samples in the value sample set are used as an iteration to-be-screened data pool, and the model of the current iteration is used to make forward prediction on the iteration to-be-screened data pool;

[0017] S60. Return to step S20 until only valueless samples or samples with very small value are left in the iteration to-be-screened data pool, and obtain a final iteration model and a final iteration data set.

[0018] Further, as a preferred technical scheme, step S10 specifically comprises:

[0019] S101. An original image sequence is obtained, and an image deduplication algorithm is used to deduplicate the original image sequence to obtain a deduplicated data set;

[0020] S102. Part of the samples in the deduplicated data set are selected as an initial data set, and an initial model is trained;

[0021] S103. The remaining samples in the deduplicated data set are used as a to-be-screened data pool, and the initial model is used to make forward prediction on the to-be-screened data pool.

[0022] Further, as a preferred technical solution, step S101 specifically comprises:

[0023] S1011. Extracting a feature vector of each picture sample in the original image sequence;

[0024] S1012. Calculating the similarity between any picture sample and all picture samples according to the feature vector;

[0025] S1013. Judging whether the similarity score exceeds a preset threshold, if yes, deleting the picture sample with similarity exceeding the preset threshold to obtain a deduplicated dataset.

[0026] Further, as a preferred technical solution, the image deduplication algorithm is based on a neural network, and step S101 specifically comprises:

[0027] A CNN network is used to extract a 1024-dimensional feature vector of each picture sample in the original dataset.

[0028] Further, as a preferred technical solution, the similarity measurement index is a normalized inner product value between two feature vectors;

[0029] The similarity is calculated by the following formula:

[0030] K(X,Y)=(X,Y) / (||X||*||Y||);

[0031] Wherein, K represents the similarity score, X and Y represent the 1024-dimensional feature vector, (X,Y) represents the vector inner product, ||X|| and ||Y|| represent the vector module.

[0032] Further, as a preferred technical solution, the data mining algorithm is based on multi-model comparison data mining, that is, the data mining algorithm designs different first and second models based on at least one of the calculation scale or the model structure, and step S20 specifically comprises:

[0033] S201. Inputting the prediction result into the first model and the second model respectively to obtain corresponding first and second detection results;

[0034] S202. One-to-one matching the detection boxes in the first and second detection results to obtain an index parameter of matching the first and second detection results;

[0035] S203. Obtaining a value score corresponding to the current sample according to the index parameter.

[0036] Further, as a preferred technical solution, in step S202, the first detection result is taken as the true value, and the second detection result is taken as the detection value; or, the second detection result is taken as the true value, and the first detection result is taken as the detection value.

[0037] The index parameters include false positives, true positives, and false negatives;

[0038] The step S202 specifically includes:

[0039] The area intersection-over-union algorithm is used to match the detection boxes in the first detection result and the second detection result one by one, to obtain false positives, true positives, and false negatives matched between the first detection result and the second detection result.

[0040] Further, as a preferred technical solution, the step S203 specifically includes:

[0041] S2031. Calculate the precision and recall of the first detection result and the second detection result according to the false positives, true positives, and false negatives matched between the first detection result and the second detection result;

[0042] S2032. Calculate the value score corresponding to the current sample according to the precision and recall of the first detection result and the second detection result.

[0043] Further, as a preferred technical solution, the precision is the ratio of true positives to the sum of true positives and false positives;

[0044] The recall is the ratio of true positives to the sum of true positives and false negatives;

[0045] The value score is the difference between the product of the precision and the recall and 1.

[0046] Further, as a preferred technical solution, the data mining algorithm is based on a single model forward data mining, that is, the data mining algorithm is based on a third model, and the step S20 specifically includes:

[0047] S204. Input the prediction result into the third model to obtain a third detection result;

[0048] S205. Obtain the confidence of all detection boxes in the third detection result;

[0049] S206. Determine whether the confidence is in the middle range. If not, the detection box score is 0. If yes, perform the next step;

[0050] S207. Calculate the area of the detection box, and calculate the score of the detection box according to the area;

[0051] S208. Add up the scores of all detection boxes to obtain the value score corresponding to the current sample.

[0052] Further, as a preferred technical solution, the step S207 specifically includes:

[0053] The area of the detection frame is calculated, the area type of the current detection frame is obtained according to the area profile, and the score of the detection frame is calculated according to the related configuration information of the area type and the area of the detection frame.

[0054] Further, as a preferred technical solution, the related configuration information of the area type includes a score lower limit, a score upper limit, an area lower limit and an area upper limit, and the score of the detection frame is calculated by the following formula:

[0055]

[0056] Wherein, box_score represents the score of the detection frame, s1 is the score lower limit, s2 is the score upper limit, a1 is the area lower limit, a2 is the area upper limit, and area is the area of the detection frame.

[0057] Compared with the prior art, the beneficial effects of the technical solution of the present application are:

[0058] The data mining process of the present application uses an active learning method to perform model iteration and data set expansion, which speeds up the model iteration speed, and under the same data amount, the model can reach higher accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 It is a step flow chart of embodiment 1 of the present application.

[0060] Figure 2 It is a step flow chart of the image deduplication algorithm of embodiment 1 of the present application.

[0061] Figure 3 It is a deduplication result schematic diagram of embodiment 1 of the present application.

[0062] Figure 4 It is a step flow chart of the multi-model comparison data mining algorithm of embodiment 2 of the present application.

[0063] Figure 5 It is a step flow chart of the single model forward data mining algorithm of embodiment 3 of the present application.

[0064] The drawings are only used for illustrative description, and cannot be understood as a limitation on the present patent; in order to better illustrate the embodiments, some components in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted; the same or similar reference numerals correspond to the same or similar components; the positional relationship described in the drawings is only used for illustrative description, and cannot be understood as a limitation on the present patent. DETAILED DESCRIPTION

[0065] The preferred embodiments of the present application are described in detail below with reference to the accompanying drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the scope of protection of the present application can be more clearly defined.

[0066] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it should be understood that if the orientations or positional relationships indicated by the terms "upper", "lower", "left", "right", "top", "bottom", "inner", "outer" and the like are based on the orientations or positional relationships shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore the terms describing the positional relationship in the drawings are only used for exemplary illustration, and cannot be understood as a limitation on the present patent.

[0067] In addition, if the terms "first", "second", and the like are used only for descriptive purposes, mainly for distinguishing different devices, elements or components (the specific types and structures may be the same or different), and are not used to indicate or imply the relative importance and quantity of the indicated devices, elements or components, and cannot be understood as indicating or implying the relative importance.

[0068] Embodiment 1

[0069] The embodiment discloses a driving scene data mining method, which is for a 2D target detection task in an automatic driving front-view camera scene, or other scenes or other tasks, and a detection result thereof includes an image sequence or a video, and a model iteration and dataset expansion framework based on active learning is designed, as shown in the method Figure 1 , comprising the steps of:

[0070] S10. Obtain an initial model and a data pool to be screened, and use the initial model to make a forward prediction on the data pool to be screened.

[0071] This step is specifically:

[0072] S101. Obtain an original image sequence, and use an image deduplication algorithm to deduplicate the original image sequence to obtain a deduplicated dataset.

[0073] In this step, since there may be redundant information in the original image sequence, such as redundant data caused by insufficient video frame extraction or static scenes, it is necessary to perform deduplication processing, and the image deduplication algorithm in this step is used to remove samples with very similar image information in the original image sequence, so as to avoid the existence of redundant samples.

[0074] In this step, the image deduplication algorithm used is a neural network-based image deduplication algorithm, therefore, this step refers to Figure 2 , and specifically includes:

[0075] S1011. Extract the feature vector of each picture sample in the original image sequence.

[0076] This step is specifically:

[0077] The 1024-dimensional feature vector of each picture sample in the original data set is extracted using the CNN network as the information representation vector of the picture sample.

[0078] S1012. Calculate the similarity between any picture sample and all picture samples according to the feature vector.

[0079] In this step, the similarity measurement index is the normalized inner product value between two feature vectors, and the similarity is the normalized inner product value, so the similarity is calculated by the following formula:

[0080] K(X,Y) = (X,Y) / (||X|| * ||Y||);

[0081] Where K represents the similarity score, X and Y represent the 1024-dimensional feature vector, (X,Y) represents the vector inner product, and ||X|| and ||Y|| represent the vector norm.

[0082] S1013. Determine whether the similarity score exceeds the preset threshold. If so, delete the picture sample with a similarity score exceeding the preset threshold to obtain a deduplicated data set.

[0083] The specific implementation process of steps S1012 and S1013 is as follows: calculate the similarity between any picture sample and all picture samples according to the feature vector. If the similarity score exceeds the preset threshold, it is considered that there is information redundancy, and all picture samples with a similarity score exceeding the threshold are deleted; repeat steps S1012 and S1013 until all picture samples are processed to obtain a deduplicated data set.

[0084] In this step, the preset threshold is set to 0.85-0.95. The deduplicated data set obtained by setting the preset threshold to different values has different numbers of samples.

[0085] For example, see Figure 3 Take 38006 pictures as the original image sequence, i.e., the original data set, and set different similarity preset thresholds for picture deduplication. Four similarity preset thresholds are tested, i.e., 0.95, 0.92, 0.9, and 0.85, and four deduplicated data sets are obtained, respectively containing 24964, 15534, 11454, and 5417 deduplicated pictures.

[0086] In the embodiment, 0.9 is finally selected as the preset threshold of similarity considering the project budget and the sparsity of the dataset, and 11454 pieces of deduplicated dataset are obtained and sent to the next step of calculation.

[0087] In some embodiments, the image deduplication algorithm adopted is a Hash-based deduplication algorithm. Therefore, the original image sequence is deduplicated by using the Hash-based deduplication algorithm to obtain the deduplicated dataset. The Hash-based deduplication algorithm is a prior art, and therefore, in the embodiment, the process of deduplicating the original image sequence by using the Hash-based deduplication algorithm is not described in detail.

[0088] S102. Select part of the samples in the deduplicated dataset as the initial dataset to train the initial model.

[0089] The step is specifically:

[0090] Randomly select picture samples in the deduplicated dataset that meet the project budget, send them to artificial marking, and use them as the initial dataset to train the initial model, that is, the initial version of the 2D target detection model.

[0091] Referring to the above example, 11454 pieces of deduplicated dataset are obtained. Therefore, in this step, 8325 picture samples are randomly selected as the initial dataset, and the initial model is trained after artificial calibration to obtain the initial model.

[0092] S103. Use the initial model to make forward prediction on the remaining samples in the deduplicated dataset as the to-be-screened data pool.

[0093] That is, the initial model is used to make forward prediction on each sample in the to-be-screened data pool.

[0094] S20. Use a data mining algorithm to mine the prediction results to obtain the value score corresponding to each sample.

[0095] That is, the forward prediction result of each sample is input into the data mining algorithm to obtain the value score corresponding to each sample.

[0096] S30. Sort the value samples in the to-be-screened data pool according to the value score to obtain the value sample set.

[0097] In this step, the value samples in the to-be-screened data pool are sorted in descending order according to the value score to obtain the value sample set.

[0098] S40. Select part of the value samples in the value sample set as the newly added dataset of the current iteration to train the model of the current iteration.

[0099] The step is specifically:

[0100] From the value sample set, the value samples that meet the project budget are taken in order as the newly added data set of this round of iteration, and the model of this round of iteration is trained using the newly added data set of this round of iteration.

[0101] Referring to the above example, 8325 picture samples are randomly selected as the initial data set, and after manual labeling, the initial version model is trained to obtain the initial model. Then, 3000 most valuable samples are mined from the value sample set as the newly added data set of this round of iteration.

[0102] S50. The remaining value samples in the value sample set are used as the iteration data pool to be screened, and the forward prediction of the iteration data pool to be screened is performed using the model of this round of iteration.

[0103] S60. Return to step S20 until only valueless samples or samples with very small value are left in the iteration data pool to be screened, and obtain the final iteration model and the final iteration data set.

[0104] That is, until there is no valuable sample in the 11454 pictures in the above example, 4163 most valuable pictures are finally generated as the final iteration data set, and the model is trained with the superposition of the data set of the last round and the final iteration data set to obtain the model of this round of iteration, that is, the final iteration model.

[0105] The data mining process of the embodiment uses an active learning method to perform model iteration and data set expansion, which speeds up the model iteration speed, and under the same data amount, the model can reach higher accuracy.

[0106] Embodiment 2

[0107] The embodiment discloses a driving scene data mining method, and specifically discloses a data mining algorithm. In the embodiment, the data mining algorithm is based on multi-model comparison data mining, that is, the data mining algorithm is based on at least a first model and a second model with different computing scales, such as a server complex model and an edge simple model; or the data mining algorithm is based on at least a first model and a second model with the same computing amount but different model structure designs, such as an anchor-based model and a model without anchors.

[0108] In the embodiment, the two models often appear in the algorithm verification stage, so the data mining algorithm of the embodiment will be used with a high probability in the middle of the project iteration.

[0109] The core idea of multi-model comparison data mining is that the detection results of different models for the same sample may differ, and the reason for the difference is that the models have different capabilities. The samples with differences are regarded as high-value samples, and the models can learn from each other and promote each other.

[0110] But there are differences in the sample also has a priority problem, that is, to give the value of the assessment score, according to score to select the optimal sample, in this embodiment, each sample corresponding to the value of the score calculation, that is, the calculation of score, see Figure 4 , by the following steps:

[0111] S201. The prediction results are input into the first model and the second model respectively to obtain corresponding first detection results and second detection results.

[0112] In this step, the forward prediction result of each sample is input into the first model and the second model respectively to obtain the first detection result and the second detection result of the current sample.

[0113] S202. The detection boxes in the first detection result and the second detection result are matched one by one to obtain index parameters matched by the first detection result and the second detection result.

[0114] In this step, the first detection result can be taken as the true value, and the second detection result can be taken as the detection value; or the second detection result can be taken as the true value, and the first detection result can be taken as the detection value.

[0115] Meanwhile, in this step, the index parameters include false positives, true positives and false negatives.

[0116] Therefore, this step specifically includes:

[0117] The detection boxes in the first detection result and the second detection result are matched one by one by using an area intersection over union (IOU) algorithm to obtain false positives (false positives, i.e. fp), true positives (true positives, i.e. tp) and false negatives (false negatives, i.e. fn) matched by the first detection result and the second detection result.

[0118] S203. Obtain a value score (score) corresponding to the current sample according to the index parameters.

[0119] This step specifically includes:

[0120] S2031. Calculate the precision and recall of the first detection result and the second detection result according to the false positives (fp), true positives (tp) and false negatives (fn) matched by the first detection result and the second detection result.

[0121] S2032. Calculate a value score corresponding to the current sample according to the precision and recall of the first detection result and the second detection result.

[0122] In this step, the calculation of the precision, the recall and the value score corresponding to the current sample is specifically as follows:

[0123] The precision is the ratio of the true positive (tp) to the sum of the true positive (tp) and the false positive (fp), i.e.,

[0124] The recall is the ratio of the true positive (tp) to the sum of the true positive (tp) and the false negative (fn), i.e.,

[0125] The value score is the difference between the product of the precision and the recall and 1, i.e., score = 1-precision*recall.

[0126] Through the above steps, the value score corresponding to each sample is obtained.

[0127] In this step, the value score ranges from 0 to 1, which can be understood as an evaluation of the difference in results. If the difference between the first detection result and the second detection result is large, the precision and the recall are smaller, the value score is larger, and the sample value is larger.

[0128] After obtaining the value score corresponding to each sample by using the data mining algorithm of the present embodiment, step S30 is continued to obtain a final iteration model and a final iteration data set.

[0129] Embodiment 3

[0130] The present embodiment discloses a driving scene data mining method, and specifically discloses a data mining algorithm. In the present embodiment, the data mining algorithm is based on single-model forward data mining, i.e., the data mining algorithm is based on a third model.

[0131] In the embodiment, the core idea of the single model forward data mining is that the confidence of the detection box reflects the confidence of the model for the detection result. The detection box with low confidence can be filtered out by threshold post-processing, which has little effect on the result. The detection box with high confidence means that the model has high confidence for the detection box, which is almost correct result. The detection box with medium confidence is the most valuable, because the model is not very confident and a little noise disturbance can easily give an error result.

[0132] Based on the above analysis, a single model forward data mining algorithm based on confidence is designed to obtain the value score corresponding to each sample, i.e. score, as shown in Figure 5 , including the following steps:

[0133] S204. Input the prediction result into the third model to obtain the third detection result.

[0134] S205. Obtain the confidence of all detection boxes in the third detection result.

[0135] S206. Determine whether the confidence is in the middle range. If not, the score of the detection box is 0. If yes, execute the next step.

[0136] In this step, the middle range of the confidence is set to [0.4, 0.7], so as to determine whether the confidence of the detection box is in the range of [0.4, 0.7]. If not, the score of the current detection box is 0. If yes, execute the next step.

[0137] S207. Calculate the area of the detection box, and calculate the score of the detection box according to the area.

[0138] This step is specifically:

[0139] Calculate the area of the detection box. According to the area configuration file, the area type of the current detection box is obtained. According to the related configuration information of the area type and the area of the detection box, the score of the detection box is calculated.

[0140] In the embodiment, the area configuration file refers to Table 1 below:

[0141] Area Type Area Value Lower Bound Area Value Upper Bound Score Lower Bound Score Upper Bound 'ignore' 0 100 0.0 0.0 'tiny' 100 256 0.0 0.0 'small' 256 1024 0.0 0.5 'medium' 1024 9216 0.5 1.0 'large' 9216 1e8 1.0 1.0

[0142] Table 1

[0143] The related configuration information of the area type includes a score lower limit, a score upper limit, an area lower limit, an area upper limit, and thus the score of the detection box is calculated by the following formula:

[0144]

[0145] wherein, box_score represents the score of the detection box, s1 is the score lower limit, s2 is the score upper limit, a1 is the area lower limit, a2 is the area upper limit, and area is the area of the detection box.

[0146] S208. The scores of all the detection boxes are accumulated to obtain the value score corresponding to the current sample.

[0147] The step specifically includes:

[0148] The scores of all the detection boxes of the current sample are accumulated to obtain the value score corresponding to the current sample.

[0149] In the step, the value score ranges from 0 to +∞, and the greater the value score, the greater the sample value.

[0150] After obtaining the value score corresponding to each sample by using the data mining algorithm of the embodiment, step S30 is continued to be executed to obtain the final iteration model and the final iteration dataset.

[0151] Obviously, the above embodiments of the present application are merely examples for clearly illustrating the present application, and are not intended to limit the implementation manners of the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the implementation manners do not need to be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A driving scene data mining method, characterized in that, The method comprises the steps of: S10. obtaining an initial model and a data pool to be screened, and performing forward prediction on the data pool to be screened by using the initial model; S20. performing data mining on the prediction result by using a data mining algorithm to obtain a value score corresponding to each sample; S30. sorting the value samples in the data pool to be screened according to the value score to obtain a value sample set; S40. selecting part of the value samples in the value sample set as a new data set of the current iteration, and training to obtain a model of the current iteration; S50. taking the remaining value samples in the value sample set as an iteration data pool to be screened, and performing forward prediction on the iteration data pool to be screened by using the model of the current iteration; S60. returning to step S20 until only valueless samples or samples with very small value are left in the iteration data pool to be screened, and obtaining a final iteration model and a final iteration data set; Step S10 specifically comprises: S101. obtaining an original image sequence, and performing deduplication on the original image sequence by using an image deduplication algorithm to obtain a deduplicated data set; S102. selecting part of the samples in the deduplicated data set as an initial data set, and training to obtain an initial model; S103. taking the remaining samples in the deduplicated data set as a data pool to be screened, and performing forward prediction on the data pool to be screened by using the initial model; The data mining algorithm is based on multi-model comparison data mining, that is, the data mining algorithm is based on at least a first model and a second model with different calculation scales, and step S20 specifically comprises: S201. inputting the prediction result into the first model and the second model respectively to obtain corresponding first detection result and second detection result; S202. one-to-one matching the detection boxes in the first detection result and the second detection result to obtain an index parameter matched by the first detection result and the second detection result; S203. obtaining a value score corresponding to the current sample according to the index parameter.

2. The driving scene data mining method of claim 1, wherein, Step S101 specifically comprises: S1011. extracting a feature vector of each picture sample in the original image sequence; S1012. calculating the similarity between any one picture sample and all picture samples according to the feature vector; S1013. judging whether the similarity score exceeds a preset threshold, and if so, deleting the picture sample with the similarity exceeding the preset threshold to obtain a deduplicated data set.

3. The method of claim 2, wherein, The image deduplication algorithm is based on a neural network, and step S101 specifically comprises: extracting a 1024-dimensional feature vector of each picture sample in the original data set by using a CNN network.

4. The driving scene data mining method of claim 2, wherein, The similarity measurement index is a normalized inner product value between two feature vectors; The similarity is calculated by the following formula: K(X,Y) = (X,Y) / (||X||*||Y||); Wherein, K represents the similarity score, X and Y represent the 1024-dimensional feature vector, (X,Y) represents the vector inner product, and ||X|| and ||Y|| represent the vector module.

5. The method of claim 1, wherein, In step S202, the first detection result is taken as the true value, and the second detection result is taken as the detection value; or the second detection result is taken as the true value, and the first detection result is taken as the detection value; The index parameter includes false positive, true positive and false negative; Step S202 specifically comprises: The area intersection-over-union algorithm is used to match the bounding boxes in the first detection result and the second detection result one by one, to obtain false positives, true positives and false negatives matched between the first detection result and the second detection result.

6. The driving scene data mining method of claim 5, wherein, Step S203 specifically includes: S2031. Calculate the precision and recall of the first detection result and the second detection result according to the false positives, true positives and false negatives matched between the first detection result and the second detection result; S2032. Calculate the value score corresponding to the current sample according to the precision and recall of the first detection result and the second detection result.

7. The driving scene data mining method of claim 6, characterized in that, the precision is the ratio of the true positives to the sum of the true positives and the false positives; the recall is the ratio of the true positives to the sum of the true positives and the false negatives; the value score is the difference between the product of the precision and the recall and 1.

Images