A clustering denoising algorithm based on key frame extraction

Abstract

The application relates to a clustering denoising algorithm for laser data in positioning technology, in particular to a clustering denoising algorithm based on key frame extraction, which comprises a key frame extraction process and a clustering denoising process. The key frame extraction process comprises the following steps: constructing a histogram by using the angle value of laser points in space, calculating the similarity of the histogram by using a PCC coefficient, judging and analyzing the key frame of the laser points by using the similarity, and discarding the laser points of non-key frames. The clustering denoising process comprises the following steps: a region segmentation processing step of carrying out region segmentation on the laser points, a clustering step of clustering the laser points in the same region after the region segmentation processing, a point cloud block merging step of analyzing and processing whether to merge the laser points after the clustering is completed, and a filtering cluster step of analyzing and processing whether the point cloud block is a normal point. The algorithm solves the problem of processing redundant data faced by a SLAM system and can reduce the overall computing pressure of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robot positioning technology, in particular to a clustering denoising algorithm for laser data in positioning technology. BACKGROUND

[0002] In an indoor laser SLAM system, due to the influence of the laser radar system itself and external factors such as dust and multi-path return caused by diffuse reflection, there are likely to be more abnormal points in the scan, in addition, the indoor laser SLAM system also faces the problem of processing redundant data, that is, the robot still needs to process a large number of similar scans when the robot is moving at a very low speed or even stops, which causes a waste of certain computing power.

[0003] To solve the above problems, the data of the radar redundancy is often processed, and the commonly used clustering denoising algorithm is mainly based on traditional clustering-based denoising algorithm and density-based noise application space clustering. Among them, the clustering-based denoising algorithm is mainly based on K-means, and the algorithm idea of K-means algorithm is roughly as follows: first, randomly select kk samples from the sample set as cluster centers, and calculate the distance between all samples and the kk cluster centers, for each sample, divide it into the cluster where the nearest cluster center is located, and calculate the new cluster center of each cluster. The density-based noise application space clustering is mainly based on DBSCAN, DBSCAN can find all dense regions of sample points and regard these dense regions as a cluster of clusters, and takes the neighborhood radius R and the minimum number of points minpoints as algorithm parameters to describe a dense region, and the concept of dense region is that when the number of points in the neighborhood radius R is greater than the minimum number of points minpoints, it is dense, the number of sample points in the neighborhood radius R is greater than or equal to minpoints. The point is called a core point, and the point which is not a core point but in the neighborhood of a core point is called a boundary point, and the point which is neither a core point nor a boundary point is a noise point.

[0004] Since the traditional clustering-based denoising algorithm such as K-means and the density-based noise application space clustering DBSCAN often have high time complexity, the indoor mobile robot with limited computing resources cannot guarantee the real-time performance of these methods. SUMMARY

[0005] The purpose of the present application is to provide a clustering denoising algorithm based on key frame extraction which solves the problem of processing redundant data faced by the SLAM system and can reduce the overall computing pressure of the system.

[0006] To achieve the above object, the technical scheme of the present application is: a clustering denoising algorithm based on key frame extraction, comprising a key frame extraction process of performing key frame extraction of laser points and a clustering denoising process of performing clustering denoising based on the key frames extracted in the key frame extraction process;

[0007] The key frame extraction process comprises constructing a histogram using angle values of the laser points in space, calculating similarity of the histogram using PCC coefficients, obtaining key frames of the laser points through similarity judgment analysis, and discarding laser points of non-key frames;

[0008] The clustering denoising process comprises a region segmentation processing step of performing region segmentation on the laser points, a clustering step of performing clustering on the laser points in the same region after the region segmentation processing, a point cloud block merging step of performing whether to merge analysis processing on the laser points after the clustering is completed, and a filtering cluster step of performing whether to be normal point analysis processing on the point cloud block.

[0009] The histogram construction method is to divide the laser points into several angle intervals according to the size of the angle values of the laser points, the angle difference in each angle interval is equal, and different angle intervals are used as the first dimension of the two-dimensional histogram, and the second dimension of the histogram is used to record the average distance of the laser points in the angle interval as Δx;

[0010] The similarity calculation method is as follows:

[0011] The histograms in the same sliding window are regarded as random variables with the same distribution, and the PCC coefficients are used to calculate the similarity;

[0012] The PCC coefficient calculation formula between histograms X and Y is:

[0013]

[0014] Where n represents the number of columns of the histogram, x i and y i represent the values of each column of the histograms X and Y, P X，Y takes a value between [-1, 1], and the value of P X，Y is closer to 1, the stronger the correlation between the two is, and the greater the similarity is;

[0015] The similarity threshold between two frames of scanning is denoted as P pair , the similarity threshold of the discarded target frame is denoted as P t res old , the past key frames saved in the sliding window are called reference frames, the latest scanning is called a target frame, P couple is used to represent the similarity between the target frame and a single reference frame, P common is used to represent the similarity between the target frame and all reference frames in the window, m is used to represent the size of the sliding window, and Vrobot representing the moving speed,

[0016] If P common < P t res old , the target frame will be reserved as a key frame, otherwise the target frame will be discarded, P pair is related to P t res old , P couple is related to P common , and the relationship is:

[0017]

[0018]

[0019] The window size m is directly related to the average distance Δx between two scans, and the definition of Δx is:

[0020]

[0021] where V robot is in centimeters, f is the scanning frequency of the laser radar, and the quantitative relationship between Δx and m is set as 18 2

[0022] The area segmentation processing step is to divide the area according to the principle that the ring areas of different areas are equal in the polar coordinate system, and the area of each area must meet:

[0023]

[0024] where r i represents the radius of the i-th circle, i = 1, 2…t, t represents the number of areas to be segmented,

[0025]

[0026] where r1 is obtained according to the distance r max of the farthest point of the laser radar, r max is directly obtained in the laser radar data, and the relationship between r1 and r max is:

[0027] In the clustering step, the angle difference between adjacent points is calculated to compare with the preset angle threshold θ threshold , if the angle difference is greater than the angle threshold θ threshold , then the two points are considered as the end of one point cloud block and the beginning of another point cloud block;

[0028] The point cloud block merging step method is to calculate the angle difference of the two points with the smallest angle of two laser point cloud blocks to compare with the preset angle difference threshold θ​threshold and the Euclidean distance threshold d threshold The angle difference threshold θ threshold The Euclidean distance threshold d is judged after the comparison condition threshold The angle difference threshold θ threshold The two point cloud blocks are directly considered as unable to be merged if the angle threshold θ threshold and the Euclidean distance threshold d are both met threshold The two laser point cloud blocks are merged into one if the comparison condition is met

[0029] The Euclidean distance formula is where R i and R i+1 represent the measured distances of the laser points i and i+1, and Δθ represents the angle difference between them.

[0030] The Euclidean distance threshold d threshold is adaptively changed with the change of the region, and the distance threshold calculation formula between the regions i and i+1 is where r i represents the radius of the circle where the i-th region is located, and θ threshold is directly related to the angle resolution θ resolution of the laser point. resolution is directly obtained in the original laser radar data.

[0031] The method of the filtering cluster step is to regard the obtained point cloud block as a cluster, and when the number of laser points in a cluster is less than a preset number threshold m threshold , the laser points in the cluster are considered as abnormal points, the abnormal points are removed, otherwise, the laser points in the cluster are considered as normal points, and the cluster is retained. The point set composed of the retained clusters is the laser point set obtained after noise reduction.

[0032] By adopting the above technical solution, the present application has the beneficial effects that the clustering noise reduction algorithm based on key frame extraction of the present application refers to the key frame concept in visual SLAM, utilizes histograms to perform dimension reduction processing on data, compares the similarity between each histogram, and the scan with low similarity is the key frame. Through the extraction of the key frame, the scan not containing new environmental features is discarded, and the extracted key frame is subjected to clustering noise reduction processing. In order to reduce the time complexity of the noise reduction method, on one hand, the present application uses a region segmentation method when clustering, traverses the points in the same region, and greatly reduces the dimension of traversal, and on the other hand, uses a key frame extraction method to screen the scan frames, and only the extracted key frames can enter the clustering noise reduction link.

[0033] In summary, the algorithm of the key frame and the clustering method of the application solves the problem of processing redundant data faced by the SLAM system on one hand, and reduces the overall computing pressure of the system after the clustering and noise reduction algorithm is added, so as to achieve the above-mentioned purpose of the application, and is particularly suitable for the clustering and noise reduction based on key frame extraction in an unstructured environment. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a flow chart of a clustering and noise reduction algorithm based on key frame extraction related to the application.

[0035] Figure 2 is a process diagram of a key frame extraction step in a clustering and noise reduction algorithm based on key frame extraction related to the application.

[0036] Figure 3 is a diagram of a region segmentation step in a clustering and noise reduction algorithm based on key frame extraction related to the application.

[0037] Figure 4 is a diagram of a clustering step in a clustering and noise reduction algorithm based on key frame extraction related to the application.

[0038] Figure 5 is a diagram of a point cloud block merging satisfying a condition in a clustering and noise reduction algorithm based on key frame extraction related to the application.

[0039] Figure 6 and Figure 7 is a diagram of two point cloud blocks not satisfying a condition in a clustering and noise reduction algorithm based on key frame extraction related to the application. DETAILED DESCRIPTION

[0040] In order to further explain the technical scheme of the application, the application will be described in detail below through specific embodiments.

[0041] The clustering and noise reduction algorithm based on key frame extraction disclosed by the application has an algorithm process as shown in Figure 1 , which includes a key frame extraction process of key frame extraction of laser points and a clustering and noise reduction process of clustering and noise reduction based on the key frames extracted in the key frame extraction process; the key frame extraction process includes constructing a histogram using the angle value of the laser points in space, calculating the similarity of the histogram using the PCC coefficient, obtaining the key frame of the laser points through similarity judgment analysis, and discarding the laser points of non-key frames, as shown in Figure 2The clustering denoising process includes a region segmentation processing step of region segmentation of laser points, a clustering step of clustering laser points in the same region after region segmentation processing, a point cloud block merging step of whether to merge analysis processing of laser points after clustering is completed, and a filtering cluster step of whether to be a normal point analysis processing of point cloud blocks, as shown in Figures 3-7 The specific calculation processing process of the present application is described in detail below.

[0042] (I) Key frame extraction process

[0043] 1. Constructing a histogram

[0044] Compared with the original laser radar data with thousands of points, the histogram has only tens of columns of data, and the dimension of the data is greatly reduced. In addition, the histogram is more flexible, and parameters such as the number of columns and the size of the columns can be easily changed when used. In order to facilitate the calculation of the similarity between each histogram, the algorithm of the present application regards the histogram as a random variable with unknown distribution and statistical characteristics.

[0045] Two-dimensional laser radar data is generally in polar coordinates according to the angle value from small to large, and the distance and angle of each laser point in the polar coordinate system can be obtained. Obviously, the characteristics of the regular change of the angle value of the laser point in space can be used to construct the histogram. For a frame of laser points, according to the size of the angle value of the laser point, they are divided into several angle intervals, and the angle difference in each angle interval is equal. The first dimension of the two-dimensional histogram is used as the different angle interval. There are two ideas for the second dimension of the histogram: the first is to use the number of laser points in the angle interval, and the other is to use the average distance of the laser points in the angle interval. Since the number of effective laser points between adjacent scans is relatively close, the number of laser points in each angle interval is also very close. Therefore, using the first idea will make the similarity between the target and the reference scan very high, which will easily lead to the excessive decline of the scan. Therefore, the algorithm of the present application adopts the second idea, that is, the second dimension of the histogram uses the average distance of the laser points in the angle interval, denoted as Δx. The advantages are that on the one hand, using the average distance of the laser points can reduce the influence of noise on the construction of the histogram, and on the other hand, the average distance can also reflect the change of the environmental characteristics.

[0046] 2. Similarity calculation

[0047] Since the scans between adjacent scans are usually similar, the method algorithm considers the histograms in the same sliding window as random variables with the same distribution, calculates the similarity using the PCC coefficient, which is a statistical indicator for measuring the strength and direction of the linear relationship between two random variables, is very sensitive to the fluctuation of numerical value, and can capture the small changes of variables, which is very helpful for solving the similarity of histograms, and most importantly, the PCC time complexity under the sliding window is only O(n*m), where n is the number of columns of the histogram, and m is the window size, and calculating PCC does not bring obvious computational burden to the SLAM system.

[0048] The PCC coefficient calculation formula between histograms X and Y is:

[0049]

[0050] Where x i and y i represent the values of each column of histograms X and Y, P X，Y takes a value between [-1, 1], and the value of P X，Y is closer to 1, the stronger the correlation between the two, the greater the similarity, and for this case of the algorithm, the value of P X，Y should be close to 1.

[0051] There are mainly four parameters to be selected in the similarity calculation, which are the number of histogram columns n, the size of the sliding window m, the similarity threshold between two frames P pair , and the similarity threshold for discarding target frames P t res old . For n, the value is the number of angle intervals of the radar resolution, for a frame of data, the larger the value of n, the fewer the laser points contained in each column, and the more details contained, two extreme cases can be considered, the first case is that the value of n is set to the number of laser points, that is, a column only contains one laser point, at this time, the similarity between histograms is very sensitive to each laser point and even noise, and the similarity will be greatly reduced, which cannot play the role of extracting key frames, the second case is to set the value of n to 1, in this case, all scans will become similar, resulting in discarding too many scans.

[0052] It can be seen that m, P pair and P t res old are closely related, the similarity threshold between two scans P pair and the similarity threshold for discarding target frames P t res old are preset, the algorithm of the application refers to the past key frames saved in the sliding window as reference frames, and the latest scan as a target frame, uses P couple to represent the similarity between the target frame and a single reference frame, and uses P commonto represent the similarity of the target frame and all reference frames within the window, if P common <P t res old , then the target frame will be retained as a key frame, otherwise the target frame will be discarded, P pair and P t res old , P couple and P common , the relationship is:

[0053]

[0054]

[0055] According to the above formula, it can be concluded that the smaller the value of m, the larger the value of P common , the easier it is to discard the target scan, therefore, the value of m will directly affect the extraction effect of key frames, generally, if the moving speed V robot of the mobile robot increases, the obtained P couple will decrease, if the value of m remains unchanged at this time, the value of P common will also decrease, at this time, in order to ensure that enough scans can be discarded, it is necessary to appropriately reduce the value of m, according to this point, the present application can obtain an approximate relationship between m and V robot , that is, the larger V robot , the smaller m needs to be, a more accurate description is that the window size m is directly related to the average distance Δx between two scans, the definition of Δx is: where V robot is in units of centimeters, f is the scanning frequency of the laser radar, and the quantitative relationship between Δx and m is set to 18 2 ·m<30.

[0056] The laser points confirmed as key frames are obtained through the above process, and the algorithm process is shown in Figure 2 .

[0057] (II) Clustering and denoising process

[0058] In order to improve the clustering effect and speed up the clustering speed, a region segmentation method is used in the clustering process. The laser points in the scan are divided into different regions by using the region segmentation method, and only the laser points in the same region are clustered. The same object may be divided into different regions during segmentation, which may eventually lead to the false removal of normal points. In order to prevent this situation, the obtained laser point cloud blocks need to be merged. The specific method steps are as follows.

[0059] 1. Region segmentation step

[0060] Region segmentation is generally used to process large-scale point clouds, and regions are divided in polar coordinate system according to the principle that the ring areas of different regions are equal, for example,Figure 3 As shown in FIG. 2, wherein the dot represents the laser point, the annulus represents the 3 regions divided, the annulus first represents the boundary of the three regions, the laser radar is located at the center of the region, and the area of each region must meet the following conditions when the region is divided:

[0061]

[0062] Wherein r i represents the radius of the i-th circle, i = 1, 2…t, t represents the number of regions to be divided, and t = 5 is better in an indoor environment. The above formula can be simplified as Wherein r1 is obtained according to the distance r max of the farthest point of the laser radar, and r max is directly obtained from the laser radar data, and the relationship between r1 and r max is as follows:

[0063]

[0064] 2. Clustering step

[0065] In the clustering step, the method of calculating the angle difference between adjacent points is used for clustering in the same region after the region is divided, which is different from the traditional clustering method based on Euclidean distance. Since the maximum and minimum values of the laser point distance in the same region are determined, the angle difference can be equivalent to the Euclidean distance. In addition, compared with the dimension that needs to be traversed in the traditional clustering, the dimension is greatly reduced after the region is divided. As can be seen from the analysis, the time complexity of the clustering denoising method is O(n*a), wherein a represents the time of comparing and calculating the angle difference between laser points, and n represents the dimension that needs to be traversed, that is, the number of laser points in a region. The time complexity of DBSCAN is O(m 2 *d), wherein m is the number of laser points, that is, the dimension that needs to be traversed, and d is the time of calculating the Euclidean distance. It can be known that m > n and d > > a, and the time complexity of the method proposed in the present application is superior to the commonly used DBSCAN method.

[0066] However, the algorithm of the present application calculates the angle difference between adjacent points to compare with the preset angle threshold θ threshold . If the angle difference is greater than the angle threshold θ threshold , then the two points are considered as the end of one point cloud block and the beginning of another point cloud block, as shown in FIG. 3. Figure 4 In the clustering, different regions do not affect each other, wherein the angle difference between the laser points i and i+1 is greater than the threshold θ threshold , so that the points i and i+1 are clustered into different blocks, i is the end of block 1, and i+1 is the beginning of block 2.

[0067] 3. Merging of point cloud blocks

[0068] To address the problem that a single object may be segmented into different regions, thus forming different point cloud blocks, this invention proposes a method for merging laser point cloud blocks within adjacent regions. It's important to note that it's unnecessary to determine whether laser point cloud blocks from non-adjacent regions can be merged. This is because, firstly, blocks across regions are usually too far apart to be merged, and secondly, if blocks across regions can be merged, then there would be mergeable blocks in the regions between them. Therefore, the point cloud block merging step in this invention's algorithm involves calculating the angle difference between the two closest points of two laser point cloud blocks and comparing it to a pre-set angle difference threshold θ. threshold Euclidean distance threshold d threshold The comparison conditions are used for judgment. When the angle difference threshold θ is met... threshold Only after comparing the conditions is it determined whether the Euclidean distance threshold d is met. threshold If the comparison condition is not met, the angle difference threshold θ will be considered. threshold If the comparison conditions are met, it is directly assumed that the two point cloud blocks cannot be merged, such as... Figure 6 The laser points i and i+1 shown satisfy the angle difference threshold, while the Euclidean distance d does not satisfy the distance threshold. Figure 7 The laser point i and i+1 do not meet the angle threshold, so we will no longer check whether d meets the distance threshold. In both of these cases, point cloud blocks 1 and 2 cannot be merged. When both laser point cloud blocks simultaneously meet the angle threshold θ, ... threshold Euclidean distance threshold d threshold When comparing conditions, two laser point cloud blocks are merged into one, such as... Figure 5 As shown, d represents the Euclidean distance between points i and i+1. At this point, both the angle difference and the distance between points i and i+1 satisfy the threshold, and point cloud blocks 1 and 2 can be merged. Since the scanned data is in polar coordinates, the Euclidean distance formula is:

[0069]

[0070] Where R i and R i+1 Let d represent the measured distance between laser points i and i+1, and Δθ represent the angular difference between them. Euclidean distance threshold d. threshold It will adapt to changes in the region, and the distance threshold between region i and i+1 is calculated using the following formula: Where r i θ represents the radius of the circle containing the i-th region. threshold Angular resolution θ of the laser point resolution Directly related, θ resolution Obtained directly from the raw lidar data.

[0071] 4. Filtering Cluster Steps

[0072] The filtering cluster step is to regard the obtained point cloud block as a cluster, when the number of laser points in a cluster is less than a preset number threshold m threshold , the laser points in the cluster are considered as abnormal points, the abnormal points are removed, otherwise, the laser points in the cluster are considered as normal points, the cluster is reserved, and the point set composed of the reserved clusters is the laser point set obtained after noise reduction.

[0073] In summary, the clustering noise reduction method based on key frame extraction can reduce the overall system computing pressure and improve the trajectory accuracy of the robot, compared with the traditional clustering-based noise reduction algorithm such as K-means and the density-based noise application space clustering DBSCAN, the key frame extraction of the application reduces the laser radar information to a histogram, uses the histogram to calculate the similarity, and the scanning with too high similarity will be discarded, the main contribution of the key frame extraction part is to reduce the computing pressure of the system, and some unnecessary calculations are avoided. Secondly, in order to speed up the clustering process, the application proposes a region segmentation method suitable for two-dimensional laser radar data, in order to avoid the misremoval of the normal point set, the application proposes a point cloud block merging method based on region segmentation. The overall algorithm can reduce the overall system computing pressure and improve the positioning accuracy of the system.

[0074] The above embodiments and drawings do not limit the product form and style of the application, and any appropriate changes or modifications made by those skilled in the art shall be considered as not departing from the patent scope of the application.

Claims

1. A clustering noise reduction algorithm based on keyframe extraction, characterized in that, This includes a keyframe extraction process for extracting keyframes of laser points and a clustering noise reduction process based on the keyframes extracted during the keyframe extraction process. The keyframe extraction process includes constructing a histogram using the angle values ​​of laser points in space, calculating the similarity of the histogram using PCC coefficients, analyzing the similarity to determine the keyframes of the laser points, and discarding non-keyframe laser points. The clustering noise reduction process includes a region segmentation process for dividing laser points into regions, a clustering process for dividing laser points in the same region after region segmentation, a point cloud block merging process for analyzing whether the clustered laser points are merged, and a filtering clustering process for analyzing whether the point cloud blocks are normal points. The method for constructing the histogram involves dividing the laser points into several angle intervals based on their angle values. The angle difference within each interval is equal. These different angle intervals form the first dimension of the two-dimensional histogram, and the second dimension is denoted by the average distance between the laser points within each angle interval. ; The similarity calculation method is as follows: Treat histograms within the same sliding window as random variables with the same distribution, and use the PCC coefficient to calculate similarity; Formula for calculating the PCC coefficient between histogram X and Y: ,in This indicates the number of columns in the histogram. and This represents the values ​​in each column of the histogram X and Y. The value is between [-1, 1]. The closer the value is to 1, the stronger the correlation and the greater the similarity between the two. The pre-set similarity threshold between two scan frames is denoted as The preset similarity threshold for discarding target frames is denoted as... The past keyframes stored in the sliding window are called reference frames, and the most recent scan is called the target frame. To represent the similarity between the target frame and a single reference frame, use To represent the similarity between the target frame and all reference frames within the window, use... To indicate the size of the sliding window, use Indicates movement speed. if < If the target frame is not found, it will be retained as a keyframe; otherwise, it will be discarded. and , and Relationship: ， 。 2. The clustering noise reduction algorithm based on keyframe extraction as described in claim 1, characterized in that, Window size The average distance between two scans Directly related The defining formula is: in In centimeters It is the scanning frequency of the lidar. and The quantitative relationship between them is set at 18. .

3. A clustering noise reduction algorithm based on keyframe extraction as described in claim 1 or 2, characterized in that, The region segmentation process involves dividing the region in a polar coordinate system according to the principle that the annular areas of different regions are equal. During region segmentation, the area of ​​each region must meet the following requirements: ， in It represents the first The radius of the circle, , Indicates the number of regions to be divided into. , in Based on the distance of the farthest point of the lidar get, Obtained directly from lidar data. and The relation is: ; The clustering step uses the angle difference between adjacent points to compare with a preset angle threshold. Compare the angles; if the angle difference is greater than the angle threshold... If these two points are considered to be the end of one point cloud block and the beginning of another, then they are considered to be the end of the point cloud block and the beginning of the other point cloud block.

4. The clustering noise reduction algorithm based on keyframe extraction as described in claim 3, characterized in that, The method for merging point cloud blocks involves calculating the angle difference between the two closest points of two laser point cloud blocks and comparing it with a preset angle difference threshold. Euclidean distance threshold The comparison conditions are used for judgment, and the angle difference threshold is met. Only after comparing the conditions is it determined whether the Euclidean distance threshold is met. If the comparison conditions are not met, the angle difference threshold will not be met. If the comparison conditions are met, it is directly assumed that the two point cloud blocks cannot be merged. This applies when both laser point cloud blocks simultaneously satisfy the angle threshold. Euclidean distance threshold When comparing conditions, two laser point cloud blocks are merged into one.

5. The clustering noise reduction algorithm based on keyframe extraction as described in claim 4, characterized in that, The Euclidean distance formula is , in and Indicates laser point and The measured distance, It represents the angular difference between the two; Euclidean distance threshold It will adapt to changes in the region. and The distance threshold between them is calculated using the following formula: ,in It indicates the first The radius of the circle containing each region. Angular resolution with respect to laser point Directly related Obtained directly from the raw lidar data.

6. The clustering noise reduction algorithm based on keyframe extraction as described in claim 3, characterized in that, The method for the filtering cluster step is to treat the obtained point cloud blocks as clusters, and when the number of laser points in a cluster is less than a preset threshold, the cluster is filtered out. If the laser points in a cluster are not normal, they are considered abnormal and removed. Otherwise, the cluster is considered normal and retained. The set of points formed by the retained clusters is the set of laser points obtained after noise reduction.

7. The clustering noise reduction algorithm based on keyframe extraction as described in claim 4, characterized in that, The method for the filtering cluster step is to treat the obtained point cloud blocks as clusters, and when the number of laser points in a cluster is less than a preset threshold, the cluster is filtered out. If the laser points in a cluster are not normal, they are considered abnormal and removed. Otherwise, the cluster is considered normal and retained. The set of points formed by the retained clusters is the set of laser points obtained after noise reduction.

8. The clustering noise reduction algorithm based on keyframe extraction as described in claim 5, characterized in that, The method for the filtering cluster step is to treat the obtained point cloud blocks as clusters, and when the number of laser points in a cluster is less than a preset threshold, the cluster is filtered out. If the laser points in a cluster are not normal, they are considered abnormal and removed. Otherwise, the cluster is considered normal and retained. The set of points formed by the retained clusters is the set of laser points obtained after noise reduction.

Images