Scene extraction method based on improved density peak clustering

Abstract

The invention relates to the technical field of automatic driving, in particular to a scene extraction method based on improved density peak clustering, and the method comprises the following steps: S1, carrying out the normalization processing of a first scene data set, and obtaining a second scene data set; s2, respectively calculating a density value of each scene data in the second scene data set by using an improved density formula; s3, dividing the second scene data set into a plurality of clusters based on an improved density peak clustering algorithm; and S4, outputting a set formed by cluster centers of all clusters as a typical scene set, so that the accuracy of typical scene extraction can be greatly improved, and the stability and adaptability are high.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a scene extraction method based on improved density peak clustering. Background Art

[0002] New energy and intelligent connected cars are not only the core driving force for the transformation and upgrading of my country's automobile industry, but also the key breakthrough for my country to compete for a leading position in the global automobile field. As the core technology of new energy and intelligent connected cars, ensuring the high reliability of autonomous driving technology is the top priority.

[0003] The update and iteration of autonomous driving technology requires the use of a test scenario library, but the existing test scenario library is not typical enough to fully cover the complex traffic environment in the real world, resulting in low accuracy in the extraction of typical scenes.

[0004] In the related art, typical scene extraction methods include scene extraction methods based on expert experience, scene extraction methods based on machine learning, and scene extraction methods based on clustering. In the scene extraction methods based on expert experience, the extraction of typical scenes is too dependent on the subjectivity of experts, and it is easy to miss potential high-value typical scenes. In the scene extraction methods based on machine learning, the scene extraction methods are unstable and unexplainable. The scene extraction methods based on clustering are currently a hot direction, but most of the scene extraction methods based on clustering are difficult to determine parameters and are unstable. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a scene extraction method based on improved density peak clustering, which can greatly improve the accuracy of typical scene extraction and has high stability and adaptability.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention provides a scene extraction method based on improved density peak clustering, comprising the following steps:

[0008] S1: normalize the first scene data set to obtain a second scene data set;

[0009] S2: using an improved density formula, respectively calculating the density value of each scene data in the second scene data set;

[0010] S3: dividing the second scene data set into multiple clusters based on an improved density peak clustering algorithm;

[0011] S4: The set of cluster centers of all clusters is output as a typical scene set;

[0012] The calculation formula of the density value is:

[0013]

[0014] Among them, d ij Represents the Euclidean distance between two scene data; d c Represents the cutoff distance taking all d ij The 95% quantile, cutoff distance represents the calculated density value ρ i How large a range of data points should be considered when calculating density? zero_one(t) represents the selection function. When calculating density, zero_one(t) selects the point i whose distance is less than d. c point;

[0015] The calculation formula for zero_one(t) is:

[0016]

[0017] The step S3 comprises:

[0018] First, set the cluster merging difficulty adjustment parameter W, and then repeat the cluster generation process until any scene data is assigned to at least one cluster, then merge all mergeable clusters according to the cluster merging rule, and finally adjust the cluster merging difficulty adjustment parameter W, and repeat the above generation process until the best result is obtained;

[0019] In the cluster generation process, the scene data with the largest density value is first selected from the unassigned scene data as the cluster center, and then the scene data with decreasing adjacent density values ​​in all the remaining scene data are found from the cluster center and included in the cluster. The definition of scene data with decreasing adjacent density values ​​is:

[0020] If there exists an ordered sequence consisting of multiple data points in the scene dataset P (P i ,P x1 ,P x2 ,…P xn ,P j ), satisfying d ix1 ≤d c And i ≥ρ x1 ;d xnj ≤d c And xn ≥ρ j ; 1≤k<n, there is

[0021] and Then the data point P j is the data point P i The adjacent density decreasing point.

[0022] By adopting the above technical solution, in the process of extracting typical scenes, the dimensional expression can be converted into dimensionless by normalizing the scene data to reduce the difficulty of data processing. Then, the density value of the scene data is calculated using the improved density formula, which can accurately quantify the correlation between the scene data and highlight the representative points. Then, the scene data is accurately divided into multiple clusters without manual point selection through the improved density peak clustering algorithm. Finally, the set composed of the cluster centers of multiple clusters is output to obtain the typical scene set. Due to the density-based clustering algorithm used in the scene extraction process, the density calculation formula and the clustering process are targetedly improved, which can greatly improve the accuracy of typical scene extraction, and has high stability and adaptability.

[0023] At the same time, according to experiments, the density calculation formula can optimally quantify the correlation between scene data and improve the accuracy of scene extraction. By adjusting the cluster merging difficulty adjustment parameter W, the difficulty of cluster merging can be set to determine the best clustering effect. In addition, in the clustering process, the point with the largest density value is repeatedly selected as the cluster center and its adjacent points with decreasing density values ​​are recursively assigned, thereby realizing cluster clustering without the manual selection of cluster centers, greatly reducing the errors caused by manual intervention, and having high adaptability to natural scene data where it is difficult to select cluster centers.

[0024] Optionally, the first scene data set includes multiple feature values, including traffic environment feature values ​​and scene subject feature values, the traffic environment feature values ​​include at least one of weather type, lighting conditions, road type, and traffic density, and the scene subject feature values ​​include vehicle type and vehicle position of the tested vehicle and the target vehicle.

[0025] By adopting the above technical solution, scene extraction can be performed based on at least one of weather type, lighting conditions, road type, and traffic density.

[0026] Optionally, the step S1 includes:

[0027] Normalizing the traffic environment characteristic values ​​and the scene subject characteristic values, and then mapping the normalized traffic environment characteristic values ​​and the scene subject characteristic values ​​to a range of 0 to 1 to obtain a second scene data set including a plurality of scene data;

[0028] The normalization formula is:

[0029]

[0030] Among them, x' represents the normalized result value of the eigenvalue, x represents the eigenvalue, and μ represents the standard deviation of the eigenvalue.

[0031] By adopting the above technical solution, normalization processing of the eigenvalues ​​can be achieved, and mapping the eigenvalues ​​to a range of 0 to 1 can effectively eliminate the dimension effect.

[0032] Optionally, the value range of the cluster merging difficulty adjustment parameter W is {0, 0.1, 0.2...1.0}, and the initial value is 1.0; for any cluster C x With cluster C y The merging rule is: if there is a data point P i , there is P i ∈C x And P i ∈C y ; Data point P i The density value is greater than that of cluster C x W*|C x | data points, and is larger than cluster C y W*|C y | data points, cluster C x With cluster C y merge.

[0033] By adopting the above technical solution and adjusting the cluster merging difficulty adjustment parameter W, it is possible to obtain extraction results of different scenarios as needed, thereby improving the diversity of extraction results.

[0034] Optionally, step S3 further includes:

[0035] The strategy for adjusting the cluster merging difficulty adjustment parameter W is: if the number of generated clusters is less than expected, increase the cluster merging difficulty adjustment parameter W; if the number of generated clusters is more than expected, decrease the cluster merging difficulty adjustment parameter W.

[0036] By adopting the above technical solution and adjusting the cluster merging difficulty adjustment parameter W, the number of clusters can be effectively controlled to obtain diversified clustering results.

[0037] In summary, the present invention at least includes the following beneficial technical effects:

[0038] 1. The density peak clustering algorithm is used for typical scene extraction in autonomous driving technology. The algorithm characteristics can greatly improve the extraction accuracy and have high adaptability to natural scene data with poor consistency.

[0039] 2. This method is oriented to application scenarios and improves the traditional density peak clustering algorithm, eliminating the step of manually selecting cluster centers and greatly reducing the errors caused by manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flowchart of a scene extraction method based on improved density peak clustering;

[0041] Figure 2 This is a clustering effect diagram of a scene extraction method based on improved density peak clustering;

[0042] Figure 3 A typical scene set extracted by a scene extraction method based on improved density peak clustering. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] The terms used in the following embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to be used as limitations of the present invention. As used in the specification of the present invention and the appended claims, the singular expressions "one", "a kind of", "said", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear contrary indication in its context. It should also be understood that the term "and / or" used in the present invention refers to and includes any or all possible combinations of one or more listed items. The term "exemplary" means "used as an example, embodiment or illustrative", and any embodiment described as "exemplary" here does not have to be interpreted as being superior to or better than other embodiments. The terms "first" and "second" are only used for descriptive purposes, and cannot be understood as implying or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" can expressly or implicitly include one or more of these features, and in the description of the embodiments of the present invention, unless otherwise specified, the meaning of "multiple" is two or more.

[0045] The embodiment of the present invention provides a scene extraction method based on improved density peak clustering.

[0046] refer to Figure 1-Figure 3 , a scene extraction method based on improved density peak clustering, comprising the following steps:

[0047] S1: Normalize the first scene data set to obtain a second scene data set.

[0048] The first scene data set includes multiple feature values ​​x, and the feature values ​​x include traffic environment feature values ​​and scene subject feature values. The traffic environment feature values ​​include: weather type, lighting conditions, road type, traffic density and other feature values, and the scene subject feature values ​​include: vehicle type, vehicle position and other feature values ​​of the tested vehicle and the target vehicle.

[0049] First, the traffic environment feature values ​​and the scene subject feature values ​​are normalized, and then the normalized traffic environment feature values ​​and the scene subject feature values ​​are mapped to the range of 0 to 1 to eliminate the dimension effect, and a second scene data set P to be clustered including multiple scene data is obtained.

[0050] The normalization process uses the Z-score (standard score) method, and the normalization formula is:

[0051]

[0052] Where x' represents the normalized result value of the eigenvalue x, x represents the eigenvalue, and μ represents the standard deviation of the eigenvalue.

[0053] It should be understood that for any scene data, there is a corresponding point to be processed. That is, for any eigenvalue x, there is a corresponding point to be processed. That is, for the second scene data set P, there is a corresponding point set to be processed.

[0054] S2: Using the improved density formula, respectively calculate the density value of each scene data in the second scene data set.

[0055] Calculate the density value ρ of each scene data in the scene data set P i , density value ρ i It will serve as an important basis for subsequent clustering, and the density value ρ i The calculation formula is:

[0056]

[0057] Among them, d ij Represents the Euclidean distance between two scene data; d c Represents the cutoff distance taking all d ij The 95% quantile, cutoff distance represents the calculated density value ρ i How large a range of data points should be considered when calculating density? zero_one(t) represents the selection function. When calculating density, zero_one(t) selects the point i whose distance is less than d. c The calculation formula of zero_one(t) is:

[0058]

[0059] It should be understood that the process of calculating the density value is a process of processing the points to be processed into data points, so as to facilitate subsequent clustering.

[0060] S3: Divide the second scene data set into multiple clusters based on an improved density peak clustering algorithm.

[0061] refer to Figure 1 and Figure 2 In the clustering process, first set the cluster merging difficulty adjustment parameter W to adjust the difficulty of cluster merging. Repeat the cluster generation process until any scene data is assigned to at least one cluster, that is, any data point is assigned to at least one cluster. Then merge all mergeable clusters according to the cluster merging rules. Finally, adjust the cluster merging difficulty adjustment parameter W according to the scene extraction strategy, and repeat the above generation process until the best result is obtained.

[0062] The value range of the cluster merging difficulty adjustment parameter W is {0, 0.1, 0.2…1.0}. The larger the value, the more difficult it is to merge the clusters, and the more clusters are generated. The initial value is 1.0.

[0063] In the process of cluster generation, the density value ρ is first selected from the unassigned scene data. i The largest one is taken as the cluster center, and then the adjacent density values ​​ρ in all the remaining scene data are found from the cluster center. i The scene data is decremented and incorporated into the cluster.

[0064] The definition of the second scene data with decreasing adjacent density values ​​is:

[0065] If there exists an ordered sequence consisting of multiple data points in the scene dataset P (P i ,P x1 ,P x2 ,…P xn ,P j ), satisfying d ix1 ≤d c And i ≥ρ x1 ;d xnj ≤d c And xn ≥ρ j ; 1≤k<n, there is and Then the data point P j is the data point P i The adjacent density decreasing point.

[0066] Among them, for any cluster C x With cluster C y The merging rule is: if there is a data point P i , there is P i ∈C x And P i ∈C y ; Data point P i The density value is greater than that of cluster C x W*|C x| data points, and is larger than cluster C y W*|C y | data points, cluster C x With cluster C y merge.

[0067] The strategy for adjusting the cluster merging difficulty adjustment parameter W is as follows: if the number of generated clusters is less than expected, the cluster merging difficulty adjustment parameter W is increased; if the number of generated clusters is more than expected, the cluster merging difficulty adjustment parameter W is decreased.

[0068] S4: Output the set of cluster centers of all clusters as a typical scene set.

[0069] refer to Figure 3 ,In this step, since the value range of the cluster merging difficulty adjustment parameter W is {0, 0.1, 0.2…1.0}, a total of 11 numbers, traverse the 11 numbers and output the visual clustering results, and take the best result as the output.

[0070] As described above, the above embodiments are only used to introduce the technical solution of the present invention in detail, but the description of the above embodiments is only used to help understand the method and core idea of ​​the present invention and should not be understood as limiting the present invention. Changes or substitutions that can be easily thought of by technicians in this technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A scene extraction method based on improved density peak clustering, characterized in that: The following steps are involved: S1: normalize the first scene data set to obtain a second scene data set; S2: using an improved density formula, respectively calculating the density value of each scene data in the second scene data set; S3: dividing the second scene data set into multiple clusters based on an improved density peak clustering algorithm; S4: The set of cluster centers of all clusters is output as a typical scene set; The calculation formula of the density value is: Among them, d ij Represents the Euclidean distance between two scene data; d c Represents the cutoff distance taking all d ij The 95% quantile, cutoff distance represents the calculated density value ρ i How large a range of data points should be considered when calculating density? zero_one(t) represents the selection function. When calculating density, zero_one(t) selects the point i whose distance is less than d. c point; The calculation formula for zero_one(t) is: The step S3 comprises: First, set the cluster merging difficulty adjustment parameter W, and then repeat the cluster generation process until any scene data is assigned to at least one cluster, then merge all mergeable clusters according to the cluster merging rule, and finally adjust the cluster merging difficulty adjustment parameter W, and repeat the above generation process until the best result is obtained; In the cluster generation process, the scene data with the largest density value is first selected from the unassigned scene data as the cluster center, and then the scene data with decreasing adjacent density values ​​in all the remaining scene data are found from the cluster center and included in the cluster. The definition of scene data with decreasing adjacent density values ​​is: If there exists an ordered sequence consisting of multiple data points in the scene dataset P (P i , P x1 , P x2 , …P xn , P j ), satisfying d ix1 ≤d c And i ≥ρ x1 ;d xnj ≤d c And xn ≥ρ j ; 1≤k<n, there is and Then the data point P j is the data point P i The adjacent density decreasing point.

2. A scene extraction method based on improved density peak clustering as claimed in claim 1, characterized in that: The first scene data set includes multiple feature values, including traffic environment feature values ​​and scene subject feature values, the traffic environment feature values ​​include at least one of weather type, lighting conditions, road type, and traffic density, and the scene subject feature values ​​include vehicle type and vehicle position of the tested vehicle and the target vehicle.

3. A scene extraction method based on improved density peak clustering as claimed in claim 2, characterized in that: The step S1 comprises: Normalizing the traffic environment characteristic values ​​and the scene subject characteristic values, and then mapping the normalized traffic environment characteristic values ​​and the scene subject characteristic values ​​to a range of 0 to 1 to obtain a second scene data set including a plurality of scene data; The normalization formula is: Wherein, x′ represents the normalized result value of the eigenvalue, x represents the eigenvalue, and μ represents the standard deviation of the eigenvalue.

4. The scene extraction method based on improved density peak clustering according to claim 1, characterized in that: The value range of the cluster merging difficulty adjustment parameter W is {0, 0.1, 0.2…1.0}, and the initial value is 1.0; for any cluster C x With cluster C y The merging rule is: if there is a data point P i , there is P i ∈C x And P i ∈C y ; Data point P i The density value is greater than that of cluster C x W*|C x | data points, and is larger than cluster C y W*|C y | data points, cluster C x With cluster C y merge.

5. A scene extraction method based on improved density peak clustering as claimed in claim 4, characterized in that: The step S3 further comprises: The strategy for adjusting the cluster merging difficulty adjustment parameter W is: if the number of generated clusters is less than expected, increase the cluster merging difficulty adjustment parameter W; if the number of generated clusters is more than expected, decrease the cluster merging difficulty adjustment parameter W.

Images