A method and device for extracting implicit intention data from natural driving data sets

Abstract

The present invention discloses a method and device for extracting implicit intention data from a natural driving dataset, comprising: step S1, obtaining reference vehicle data and candidate vehicle data; step S2, training an Adaboost classifier based on the reference vehicle data and non-reference data and non-candidate vehicle data in the natural driving dataset; step S3, obtaining driving behavior similarity between the reference vehicle and the candidate vehicle based on driving information of the reference vehicle data and the candidate vehicle data; step S4, if the driving behavior similarity is greater than a preset driving behavior similarity threshold, proceeding to step S5; otherwise, determining that the candidate vehicle does not have the intention to perform the driving behavior; and step S5, inputting the driving information of the candidate vehicle that is greater than the preset driving behavior similarity threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has the implicit intention to perform the driving behavior.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving technology, and in particular relates to a method and device for extracting implicit intention data from a natural driving data set. Background Art

[0002] Autonomous driving technology is rapidly developing and has garnered widespread attention and research from both industry and academia, both domestically and internationally. This has led to the emergence of a large number of naturalistic driving datasets, providing real-world data support for autonomous driving research and development. Effectively leveraging this data can improve autonomous driving technology from multiple perspectives, such as enhancing the accuracy and safety of decision-making modules and improving the passenger experience. However, some naturalistic driving datasets contain driving data with implicit intent, meaning that the driver intends a specific driving action but fails to execute it due to unknown factors such as the traffic environment and the vehicle's state. Effectively mining this data can significantly enhance the human-like experience of autonomous driving systems.

[0003] Patent CN105045900A discloses a method for obtaining summary information based on data similarity. This invention uses the longest common subsequence method to calculate the similarity between the first and second data in the sorting position. If the similarity exceeds a threshold, the data in the second position is deleted and the summary information of the data in the first position is obtained. This invention is mainly used in the field of text processing and has a limited application scenario.

[0004] Patent CN105247430A discloses a data processing method based on data similarity. This invention detects the aging of industrial robots by comparing the similarity between command data and servo data. However, the similarity calculation method in this invention uses a simple formula and is not suitable for extracting scene data with complex relationships.

[0005] The lack of sufficient utilization of naturalistic driving datasets has prevented the transfer of the vast amount of driving experience implicit in the data, failing to fully support improvements in the safety and stability of autonomous driving systems. Furthermore, the human-like experience of autonomous driving systems is limited. Currently, there is a lack of methods for extracting driving data that specifically captures the implicit intentions in naturalistic driving datasets. Furthermore, data extraction methods from other fields cannot be directly transferred due to their limited considerations and insufficient information. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and device for extracting implicit intention data from a natural driving dataset. The method and device comprehensively consider the motion information of a single vehicle and the information of its environment, and can, to a certain extent, explore the nonlinear relationship between various factors while avoiding overfitting. The method is easy to use and can fully explore the natural driving dataset to provide data support for the development of autonomous driving systems.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for extracting implicit intent data from a natural driving dataset comprises the following steps:

[0009] Step S1, obtaining reference vehicle data and candidate vehicle data;

[0010] Step S2: training an Adaboost classifier based on the benchmark vehicle data and the non-benchmark vehicle data and non-candidate vehicle data in the natural driving dataset;

[0011] Step S3: obtaining the driving behavior similarity between the reference vehicle and the candidate vehicle based on the driving information of the reference vehicle data and the candidate vehicle data;

[0012] Step S4: If the driving behavior similarity is greater than the preset driving behavior similarity threshold, proceed to step S5; otherwise, determine that the candidate vehicle has no intention to perform the driving behavior;

[0013] Step S5: Input the candidate vehicle driving information that is greater than a preset driving behavior similarity threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has an implicit intention to perform the driving behavior.

[0014] Preferably, in step S1, vehicle data that performs the driving behavior is extracted from a natural driving data set according to the driving behavior as reference vehicle data.

[0015] Preferably, in step S2, based on the driving position and time of the reference vehicle, vehicle data with similar positions and times are obtained from the non-reference vehicle data as candidate vehicle data.

[0016] Preferably, step S2 includes:

[0017] Mark the benchmark vehicle data as positive data, and mark the non-benchmark vehicle data and non-candidate vehicle data as negative data;

[0018] The positive and negative data are combined into a training data set to train the Adaboost classifier.

[0019] The present invention also provides a device for extracting implicit intention data from a natural driving data set, comprising:

[0020] An acquisition module, used to acquire benchmark vehicle data and candidate vehicle data;

[0021] A training module, configured to train an Adaboost classifier based on the benchmark vehicle data and non-benchmark vehicle data and non-candidate vehicle data in the natural driving dataset;

[0022] a calculation module, configured to obtain a driving behavior similarity between the reference vehicle and the candidate vehicle based on driving information of the reference vehicle data and the candidate vehicle data;

[0023] a judgment module, configured to proceed to step S5 if the driving behavior similarity is greater than a preset driving behavior similarity threshold, and otherwise determine that the candidate vehicle does not have the intention to perform the driving behavior;

[0024] The extraction module is used to input the driving information of candidate vehicles with a driving behavior similarity greater than a preset threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has an implicit intention to perform the driving behavior.

[0025] Preferably, in the acquisition module, vehicle data performing the driving behavior is extracted from a natural driving data set according to the driving behavior as the benchmark vehicle data.

[0026] Preferably, in the acquisition module, based on the driving position and time of the reference vehicle, vehicle data with similar positions and times are acquired from the non-reference vehicle data as candidate vehicle data.

[0027] Preferably, the training module includes:

[0028] a marking unit, configured to mark the reference vehicle data as positive example data, and mark the non-reference vehicle data and non-candidate vehicle data as negative example data;

[0029] The training unit is used to form a training data set from positive example data and negative example data to train the Adaboost classifier.

[0030] The technical solution of this invention mines the correlation between identical / similar driving behaviors and identical / similar driving intentions and integrates this correlation into data extraction of implicit intentions. This method has clear logic and is easy to implement, making it widely applicable to mining traffic big data. The extracted data can be used in autonomous driving decision-making modules to improve the generation of complex decision-making models, thereby enhancing the accuracy and effectiveness of decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of a method for extracting implicit intention data from a natural driving dataset according to an embodiment of the present invention;

[0032] Figure 2 A flowchart of a method for extracting implicit intention data from a natural driving dataset according to another embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0035] The definitions of the formula symbols involved in the embodiments of the present invention are shown in Table 1:

[0036] Table 1

[0037]

[0038]

[0039] Example 1:

[0040] like Figure 1 As shown, an embodiment of the present invention provides a method for extracting implicit intention data from a natural driving dataset, comprising the following steps:

[0041] Step S1, obtaining reference vehicle data and candidate vehicle data;

[0042] Step S2: Train an Adaboost classifier based on the baseline vehicle data and non-baseline vehicle data and non-candidate vehicle data in the natural driving dataset; wherein the data in the natural driving dataset is divided into three categories: 1. Baseline vehicle data, which is data on vehicles that have generated the "implicit intent" as described in the text and have performed the corresponding behavior; 2. Candidate vehicle data, which is data on vehicles that may have generated the "implicit intent" as described in the text but have not performed the corresponding behavior; 3. Non-baseline vehicle data and non-candidate vehicle data, which is data on vehicles that have not generated the "implicit intent" as described in the text and have not performed the corresponding behavior;

[0043] Step S3: obtaining the driving behavior similarity between the reference vehicle and the candidate vehicle based on the driving information of the reference vehicle data and the candidate vehicle data;

[0044] Step S4: If the driving behavior similarity is greater than the preset driving behavior similarity threshold, proceed to step S5; otherwise, determine that the candidate vehicle has no intention to perform the driving behavior;

[0045] Step S5: Input the candidate vehicle driving information that is greater than a preset driving behavior similarity threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has an implicit intention to perform the driving behavior.

[0046] As an implementation of the embodiment of the present invention, in step S1, vehicle data that performs the driving behavior is extracted from a natural driving data set according to the driving behavior as reference vehicle data.

[0047] As an implementation of an embodiment of the present invention, in step S2, based on the driving position and time of the reference vehicle, vehicle data with similar positions and times are obtained from the non-reference vehicle data as candidate vehicle data.

[0048] As an implementation of the embodiment of the present invention, step S2 includes:

[0049] Mark the benchmark vehicle data as positive data, and mark the non-benchmark vehicle data and non-candidate vehicle data as negative data;

[0050] The positive and negative data are combined into a training data set to train the Adaboost classification.

[0051] Example 2:

[0052] An embodiment of the present invention provides a method for extracting implicit intention data from a natural driving dataset, comprising the following steps:

[0053] Step 1: Establish an implicit intention reasoning model as follows, so as to obtain the direct correlation between the same / similar driving behavior and the same / similar driving intention, and establish an implicit intention driving data extraction method based on this correlation. The specific process is as follows: Figure 2 shown.

[0054] 1) Existing Phenomenon (PH): The candidate vehicle (v2) and the benchmark vehicle (v1) have the same motion state when they are at the same / similar location and in the same / similar vehicle topology network;

[0055] 2) Based on PH, there are two hypotheses for the driving intention of candidate vehicle v2:

[0056] A1), v2 has the same intention as v1;

[0057] A2) v2 has different intentions from v1.

[0058] 3) In step 2), the two intention hypotheses (A1 hypothesis and A2 hypothesis) are completely random. If the intention information of v2 is completely unknown, we can first assume that P(A1) = P(A2) = 1 / 2;

[0059] 4) Combining 2) and 3) we can get the conditional probability of PH (the existing phenomenon described in step 1) under two assumptions:

[0060]

[0061]

[0062] Among them, M i Represents the attributes describing the vehicle's motion state, such as acceleration and speed, and n represents the number of attributes describing the vehicle's motion state;

[0063] 5) Then the posterior probability is as follows:

[0064]

[0065]

[0066] 6) The ratio of the two posterior probabilities mentioned above:

[0067]

[0068] And R>>1, that is, when the phenomenon PH exists, hypothesis A1 is much more likely than hypothesis A2. Therefore, according to this inference model, when two vehicles traveling in a similar location and within a similar vehicle topology network exhibit the same driving behavior, it can be inferred that they have the same driving intention. This conclusion can be generalized to similar situations. That is, when two vehicles traveling in a similar location and within a similar vehicle topology network exhibit similar driving behavior, it can be inferred that they have similar driving intentions. The higher the behavioral similarity, the higher the similarity of the driving intentions.

[0069] Step 2: For a driving behavior Act, extract the vehicle data that performs the behavior from the natural driving data set, and use this part of the data as the benchmark vehicle data, including the location information (loc vi ), motion status information and the multi-vehicle location topology information in,

[0070]

[0071]

[0072]

[0073] Among them, vel vi 、acc vi They represent the speed and acceleration of vehicle vi respectively; A flag vector indicating whether there is a vehicle at the location node corresponding to other vehicles in the multi-vehicle topology where vehicle vi is located; A vector representing the relative distance between vehicle vi and other vehicles in the multi-vehicle position topology; The vector representing the relative velocity between vehicle vi and other vehicles in the multi-vehicle position topology; The vector representing the relative acceleration between vehicle vi and other vehicles in the multi-vehicle position topology in which it is located; FV, LVF, LVB, RVF, and RVB represent the vehicle in front of the described vehicle, the vehicle in front of the left lane, the vehicle behind the left lane, the vehicle in front of the right lane, and the vehicle behind the right lane, respectively; f i Indicates whether there is a traffic vehicle at the location node corresponding to traffic vehicle i in the multi-vehicle topology structure where the vehicle is located. Define f i The values ​​are as follows (see Table 1 for the symbols / meanings in the text):

[0074]

[0075]

[0076] Where Δloc vj-vk represents the relative position between vehicle vj and vehicle vk. For example, when (vj, vk) = (vi, LVF), Δloc vj-vk It is expressed as the relative position between vi and the vehicle in front of vi in ​​the left lane (i.e., LVF). The specific expression is as follows:

[0077] Δloc vj-vk =loc vj -loc vk

[0078]

[0079] Among them, Δvel vj-vk represents the relative speed between vehicle vj and vehicle vk. For example, when (vj, vk) = (vi, LVF), Δvel vj-vk It is expressed as the relative speed between vi and the vehicle in front of vi in ​​the left lane (i.e., LVF). The specific expression is as follows:

[0080] Δvel vj-vk =vel vj -vel vk

[0081]

[0082] Among them, Δacc vj-vkrepresents the relative acceleration between vehicle vj and vehicle vk. For example, when (vj, vk) = (vi, LVF), Δacc vj-vk It is expressed as the relative acceleration between vi and the vehicle in front of vi in ​​the left lane (i.e., LVF). The specific expression is as follows:

[0083] Δacc vj-vk =acc vj -acc vk

[0084] Step 3: Extract the candidate vehicle data from the non-reference vehicle data in the natural driving dataset, which is close in location and time to the reference vehicle data. This data contains the necessary information of the candidate vehicle, namely, its location information, motion state information, and its multi-vehicle position topology information.

[0085] Step 4: Mark the benchmark vehicle data as positive data, and mark the non-benchmark vehicle data and non-candidate vehicle data as negative data;

[0086] Step 5: Combine the positive and negative data into a training data set and train the Adaboost classifier;

[0087] Step 6: For each pair of reference vehicle (v1) data and its corresponding candidate vehicle (v2) data, calculate the similarity of driving behavior in different dimensions, including the similarity of vehicle positions (sim loc ), multi-vehicle topology similarity (sim top_struc ) and attribute similarity (sim top_val ), single vehicle motion similarity (sim mot ), and define the similarity calculation method as follows:

[0088]

[0089]

[0090]

[0091]

[0092] Among them, σ represents the degree of concentration of the calculation model's definition of "close location", Jaccard' is the modified Jaccard similarity, and Pearson is the Pearson correlation coefficient. Their calculation methods are as follows:

[0093]

[0094]

[0095]

[0096]

[0097] Step 7: Combine the multi-dimensional similarities of driving behaviors to obtain the overall driving behavior similarity (sim), which is expressed as follows:

[0098] sim=sim loc ·sim top_struc ·sim top_val ·sim mot

[0099] Step 8: Compare the overall driving behavior similarity with a preset similarity threshold. If the similarity is less than the threshold, the data is considered to be driving behavior data without implicit intention. If the similarity is greater than the threshold, continue.

[0100] Step 9: Input the candidate vehicle data with driving behavior similarity greater than the threshold into the Adaboost classifier obtained in step 5. If the output is a positive example, the vehicle data has the intention to perform the driving behavior Act determined in step 2; otherwise, it does not have this intention.

[0101] The method for extracting data revealing implicit intent from natural driving datasets, as described in the present invention, is based on an inference model that maps driving behavior to driving intent. This method can mine large amounts of data revealing implicit intent from existing datasets for use in autonomous driving calculations, decision-making, and other models. This allows autonomous vehicles to learn from the hidden human driver experience within the data, improving both decision accuracy and safety while enhancing the passenger experience. The method boasts clear logic and is easy to implement. Furthermore, the method leverages a rich source of natural driving big data, making it applicable to other autonomous driving decision-making, control, and calculation modules requiring big data support.

[0102] The technical effects of the embodiments of the present invention are as follows:

[0103] 1. Established a reasoning model for the relationship between driving behavior and implicit driving intention;

[0104] 2. Based on the inference model, a framework for extracting implicit intent data based on driving similarity is constructed. This framework also takes into account the vehicle and traffic environment factors of the vehicle under investigation, making it highly universal.

[0105] 3. During data extraction, vehicle data is selected based on driving similarity, and an Adaboost classifier is added for further screening and confirmation. This combined structure can effectively identify nonlinear relationships while avoiding overfitting.

[0106] Example 3:

[0107] An embodiment of the present invention further provides a device for extracting implicit intention data from a natural driving dataset, comprising:

[0108] An acquisition module, used to acquire benchmark vehicle data and candidate vehicle data;

[0109] A training module, configured to train an Adaboost classifier based on the benchmark vehicle data and non-benchmark vehicle data and non-candidate vehicle data in the natural driving dataset;

[0110] a calculation module, configured to obtain a driving behavior similarity between the reference vehicle and the candidate vehicle based on driving information of the reference vehicle data and the candidate vehicle data;

[0111] a judgment module, configured to proceed to step S5 if the driving behavior similarity is greater than a preset driving behavior similarity threshold, and otherwise determine that the candidate vehicle does not have the intention to perform the driving behavior;

[0112] The extraction module is used to input the driving information of candidate vehicles with a driving behavior similarity greater than a preset threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has an implicit intention to perform the driving behavior.

[0113] As an implementation manner of the embodiment of the present invention, in the acquisition module, vehicle data performing the driving behavior is extracted from the natural driving data set according to the driving behavior as the benchmark vehicle data.

[0114] As an implementation of an embodiment of the present invention, in the acquisition module, based on the driving position and time of the reference vehicle, vehicle data with similar positions and times are acquired from the non-reference vehicle data as candidate vehicle data.

[0115] As an implementation of an embodiment of the present invention, the training module includes:

[0116] a marking unit, configured to mark the reference vehicle data as positive example data, and mark the non-reference vehicle data and non-candidate vehicle data as negative example data;

[0117] The training unit is used to form a training data set from positive example data and negative example data to train the Adaboost classifier.

[0118] The above is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope described in the present invention should be covered by the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for extracting implicit intention data from a natural driving dataset, characterized in that: The following steps are involved: Step S1, obtaining reference vehicle data and candidate vehicle data; Step S2: training an Adaboost classifier based on the benchmark vehicle data and non-benchmark data and non-candidate vehicle data in the natural driving dataset; Step S3: obtaining the driving behavior similarity between the reference vehicle and the candidate vehicle based on the driving information of the reference vehicle data and the candidate vehicle data; Step S4: If the driving behavior similarity is greater than the preset driving behavior similarity threshold, proceed to step S5; otherwise, determine that the candidate vehicle has no intention to perform the driving behavior; Step S5: Input the driving information of the candidate vehicle with a driving behavior similarity greater than a preset driving behavior similarity threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has an implicit intention to perform the driving behavior; In step S1, vehicle data that performs the driving behavior is extracted from the natural driving data set as benchmark vehicle data based on the driving behavior; and vehicle data with similar positions and times are obtained from the non-benchmark vehicle data as candidate vehicle data based on the driving position and time of the benchmark vehicle.

2. The method for extracting implicit intention data from a natural driving dataset according to claim 1, wherein: Step S2 includes: Mark the benchmark vehicle data as positive data, and mark the non-benchmark vehicle data and non-candidate vehicle data as negative data; The positive and negative data are combined into a training data set to train the Adaboost classification.

3. A device for extracting implicit intention data from a natural driving dataset, characterized in that: include: An acquisition module, used to acquire benchmark vehicle data and candidate vehicle data; A training module, configured to train an Adaboost classifier based on the benchmark vehicle data and non-benchmark data and non-candidate vehicle data in the natural driving dataset; a calculation module, configured to obtain a driving behavior similarity between the reference vehicle and the candidate vehicle based on driving information of the reference vehicle data and the candidate vehicle data; a judgment module, configured to proceed to step S5 if the driving behavior similarity is greater than a preset driving behavior similarity threshold, and otherwise determine that the candidate vehicle does not have the intention to perform the driving behavior; An extraction module is used to input the driving information of candidate vehicles with a driving behavior similarity greater than a preset threshold into the trained Adaboost classifier. If the output result is a positive example, the candidate vehicle has an implicit intention to perform the driving behavior; According to the driving behavior, vehicle data that performs the driving behavior is extracted from the natural driving data set as the benchmark vehicle data; based on the driving position and time of the benchmark vehicle, vehicle data with similar position and time are obtained from the non-benchmark vehicle data as candidate vehicle data.

4. The device for extracting implicit intention data from a natural driving dataset according to claim 3, wherein: The training module includes: a marking unit, configured to mark the reference vehicle data as positive example data, and mark the non-reference vehicle data and non-candidate vehicle data as negative example data; The training unit is used to form a training data set from positive example data and negative example data to train the Adaboost classifier.

Images