Multi-step clustering point cloud denoising method for disinfection robot in water mist environment

Abstract

The invention discloses a multi-step clustering point cloud denoising method for a disinfection robot in a water mist environment, and belongs to the field of disinfection robot radar point cloud processing. Firstly, primary noise separation is carried out based on point set density difference, and a density distribution model is constructed to realize initial point set division; a dynamic compensation mechanism is introduced on the basis, and accurate re-classification of misclassification point sets is realized through a shared nearest neighbor clustering algorithm; and finally, an improved I-DBSCAN algorithm fused with a local density adaptive parameter adjustment strategy is adopted to complete secondary filtering of residual noise. According to the method, noise and targets can be adaptively distinguished through a multi-stage cooperative processing mechanism, the integrity of effective point clouds is guaranteed while water mist interference and miscellaneous point shielding are efficiently removed, the influence of water mist generated when the disinfection robot works and complex environment data on correct navigation information is greatly weakened, and the navigation accuracy is improved. The limitation that laser radar cannot be used for accurate navigation when the robot sprays disinfectant fluid is solved.

Description

Technical Field

[0001] This invention relates to the field of robot control technology and autonomous navigation, specifically to a multi-step clustering point cloud denoising method for disinfection robots in water mist environments. Background Technology

[0002] With the development of intelligent farming technology, lidar is increasingly widely used in pigsty environmental monitoring. It can accurately perceive pig behavior and environmental obstacles through point cloud data. However, complex interference factors exist in the pigsty environment, such as water mist generated by spray disinfection and cooling, as well as equipment obstruction, resulting in a large amount of noise (such as water mist reflection points and obstruction noise) in the point cloud data collected by the lidar. This noise severely reduces the accuracy and reliability of target detection, affecting the decision-making efficiency of the automated system.

[0003] Traditional point cloud denoising methods (such as statistical filtering and radius filtering) typically remove noise based on global density or distance thresholds, but they have significant limitations in the dynamic environment of pigsties: First, water mist noise and real obstacle point clouds have similar spatial distributions, making it difficult to effectively distinguish them with a single threshold; second, discrete noise points generated by equipment occlusion locally overlap with the target point cloud, and conventional clustering algorithms (such as DBSCAN) are prone to missegmentation due to fixed parameter settings; third, when multiple noise sources are coupled, existing methods struggle to balance the comprehensiveness of noise removal with the integrity of the target point cloud. For example, traditional density clustering algorithms are sensitive to parameters and in dynamic water mist environments, they are prone to misclassifying sparse real fence points as noise or failing to effectively recover sparse target points that have been incorrectly removed.

[0004] In recent years, multi-step processing frameworks have been proposed for point cloud denoising in complex environments, but most methods have not fully considered the special characteristics of agricultural scenarios. For example, deep learning-based denoising methods rely on large amounts of labeled data and lack generalization ability in the context of variable pigsty environments; while geometric feature-based methods lack adaptive adjustment capabilities for local density changes caused by water mist. In addition, existing technologies often employ direct filtering strategies for outliers caused by occlusion, which may result in the loss of effective information, especially in areas with complex structures such as medicine boxes and fences, where the problem of false rejection is particularly prominent.

[0005] Therefore, there is an urgent need for a point cloud denoising method for pigsty water mist environments. This method should be able to adaptively distinguish between noise and targets through a multi-stage collaborative processing mechanism. It should efficiently remove water mist interference and occlusion noise while ensuring the integrity of the effective point cloud, thus providing a high-precision data foundation for subsequent pigsty environment modeling and target detection. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention proposes a multi-step clustering point cloud denoising method for disinfection robots in water mist environments. The method progressively denoises based on the characteristics of different point sets in a frame of point cloud data. First, the y-axis distance constraint is introduced into the relative density formula based on the point set characteristics to achieve the separation of the first layer of noise. Next, the shared nearest neighbor clustering method is used to re-acquire misclassified point sets. Finally, the I-DBSCAN clustering method is used to process the remaining point sets to achieve the filtering of the second layer of noise, resulting in more accurate navigation information. This invention significantly reduces the impact of water mist generated during the operation of the pigsty disinfection robot and the complex environmental data of the pigsty on accurate navigation information, solving the limitation that the robot cannot use lidar for precise navigation when spraying disinfectant.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A multi-step clustering point cloud denoising method for disinfection robots in water mist environments includes the following steps:

[0009] S1, Real-time acquisition of environmental point cloud data: A high-performance DTOF lidar sensor is used to acquire two-dimensional point cloud data streams in real time, and the polar coordinate parameters of each point are recorded synchronously, including radial distance and azimuth angle.

[0010] S2, Point cloud data preprocessing: Calculate the visible range of the lidar based on its location and the obstruction angle of the medicine box. Perform preliminary screening of the original point cloud based on angle constraints to suppress erroneous reflections caused by self-occlusion and improve the robustness of subsequent algorithms. Then, transform the screened point cloud data from polar coordinates to Cartesian coordinates.

[0011] S3, peripheral noise coarse filtering: The point cloud data is divided into a noise set and a point of interest set by a relative density constraint with Euclidean distance limit, which quickly removes peripheral irrelevant noise and improves processing efficiency;

[0012] S4, Misclassified Point Cloud Correction: For the true points of interest that have been misclassified into the noisy point set, the Shared Nearest Neighbor Clustering (SNNC) method is introduced to restore the misclassified points and ensure the integrity of the target.

[0013] S5, Water Mist Noise Removal: An improved adaptive density clustering I-DBSCAN algorithm is used to process the set of points of interest. Through a dynamic parameter optimization mechanism, water mist noise features are separated and accurately removed, and a high-precision fence point set is finally output.

[0014] Furthermore, in S3, the relative density constraint formula with Euclidean distance limitation is expressed as follows:

[0015]

[0016] In the formula, k represents the total number of samples in the point set, i represents the label of the i-th sample, and p i Let rd represent the i-th point. i Let yi represent the relative density of the i-th point in the point set, and y0 be the y-axis distance constraint constant. i p is the y-coordinate of the i-th point, n is the iteration variable for summation, and p n It is the nth point during the summation iteration. This represents the Euclidean distance between the i-th point and the n-th point.

[0017] Furthermore, in step S4, the process of processing the noisy set using the shared nearest neighbor clustering algorithm includes the following sub-steps:

[0018] S4-1, find the k nearest neighbors of each of the m points in the noise set, and use k... n [i] represents, where k is a preset constant, if k n [i] and k n If [j] have an intersection, then point p will be... i and p j Save to k c [i] in;

[0019] S4-2, Determine k c [i] and k c If [j] has an intersection, then let k = [j]. c [i] and k c The union of [j] is saved to k. c [i] and clear k c [j];

[0020] S4-3, Calculate all non-empty k c [i] contains the mean of the number of points in the cloud;

[0021] S4-4, The quantity of k is less than the average. c The point cloud in [i] is extracted and categorized into the set of points of interest.

[0022] Furthermore, in step S5, the processing of the set of points of interest using the I-DBSCAN algorithm includes the following sub-steps:

[0023] S5-1, Generate the distance distribution matrix of the set of points of interest D:

[0024] D n×n ={Dist(i,j)∣1≤i≤n,1≤j≤n}

[0025] Among them, D n×nIt is an n×n matrix, where n represents the total number of data objects in the dataset, i and j represent the i-th or j-th data point, and Dist(i,j) represents the distance from data point i to data point j.

[0026] S5-2, using the Gaussian smoothing formula to weight each data point, smooths the distance distribution matrix:

[0027]

[0028] Where G(x,y) represents the Gaussian weight value at position (x,y), x and y represent the positions of the pixel or data point in the horizontal and vertical directions, respectively, and σ represents the standard deviation of the Gaussian function, which determines the shape and smoothness of the Gaussian function;

[0029] S5-3: The distance distribution matrix after Gaussian smoothing is sorted in ascending order. In this matrix, the first column element of each row represents the distance from the data object to itself, with a value of 0. The data elements located in the k-th column of the distance distribution matrix constitute the k-nearest neighbor vector D. k ;

[0030] S5-4, for each D k The k-mean nearest neighbor distance is calculated by averaging the values ​​and then placed into a set as candidate Eps parameters in order. This results in a set of candidate Eps parameters for all k-mean nearest neighbor distances. The number of data point objects in the neighborhood corresponding to each Eps parameter is calculated, and then the expected value of the Eps neighborhood data objects for all objects in dataset D is calculated. This generates the MinPts parameter set for D, as shown in the formula:

[0031]

[0032] Where n represents the total number of objects in the dataset, i is the iteration variable when summing, and Pi represents the number of objects contained in the Eps neighborhood of the i-th object;

[0033] S5-5: Select the corresponding Eps parameter from the Eps parameter set and the corresponding MinPts parameter from the MinPts parameter set. Input these parameters into the I-DBSCAN algorithm and observe the clustering results. Obtain the cluster number results under different k values. If the generated cluster number results are the same for four consecutive times, it can be determined that the clustering has become stable. The resulting cluster number p is considered to be the optimal cluster number. Continue execution. When the generated cluster number is no longer equal to the optimal cluster number p, select the maximum k value corresponding to the cluster number equal to p as the optimal k value. The Eps parameter corresponding to the optimal k value is considered to be the optimal Eps parameter, the corresponding MinPts parameter is considered to be the optimal MinPts parameter, and the corresponding clustering result is considered to be the optimal result.

[0034] S5-6, the fence point set is extracted according to the distance feature of the two farthest points in the cluster to achieve the final denoising effect.

[0035] The main advantages of this invention are as follows: It progressively denoises based on the characteristics of different point sets in a frame of point cloud data. By introducing relative density with Euclidean distance constraints, it achieves the separation of the first layer of noise. Then, it uses shared nearest neighbor clustering to re-acquire misclassified point sets. Finally, it uses I-DBSCAN clustering to process the remaining point sets, achieving the filtering of the second layer of noise and obtaining more accurate navigation information. This invention significantly reduces the impact of water mist generated during the operation of the pigsty disinfection robot and the complex environmental data of the pigsty on accurate navigation information, overcoming the limitation that the robot cannot use lidar for precise navigation when spraying disinfectant. Attached Figure Description

[0036] Figure 1 This is a block diagram illustrating the implementation of the method of the present invention.

[0037] Figure 2 This is a noise-free point cloud map of an embodiment of the present invention.

[0038] Figure 3 This is a point cloud map of an environment with noise, as described in an embodiment of the present invention.

[0039] Figure 4 This is a diagram illustrating the verification effect of the method in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to specific embodiments.

[0041] Reference Figures 1-4 In this embodiment, an intelligent disinfection robot designed and manufactured by the research group was selected. Environmental point clouds scanned and saved using the SLAMTEC S2 lidar mounted on the robot at a pig farm in Changshan County, Quzhou City, were used as test data to verify the method proposed in this invention. During the robot's navigation process, the extracted ideal fence point cloud is shown below. Figure 2 As shown, a very accurate navigation channel can be obtained. Figure 3 The two images shown represent real point cloud data containing only peripheral noise and real point cloud data containing both types of noise, respectively. First, in step S1, a complete frame of point cloud data containing distance and angle information is obtained. Then, in step S2, radar angle constraints are set to remove erroneous data caused by occlusion. In step S3, the relative density of each point is calculated, and the point cloud is divided into two parts using a pre-set threshold parameter. The relative density formula incorporates the point cloud's y-axis distance constraint. Next, in step S4, the k-nearest neighbor set of each point in the low relative density set is calculated. By determining whether there is intersection between these sets, the nearest neighbor set is shared, thereby restoring misclassified points to the high relative density set. In step S5, a distance distribution matrix is ​​first generated based on the set of interest. Then, the distance distribution matrix is ​​sorted in ascending order. The k-mean nearest neighbor method is used to calculate the k-nearest neighbor distance between each data point in the dataset and its k-th nearest neighbor. The mean distance is then calculated, and the remaining k values ​​are calculated in the same way. Finally, the set of parameters to be selected for the Eps parameter is obtained. Then, the expected value method is used to analyze and process the Eps parameter, and finally, the set of parameters to be selected for the MinPts parameter is generated. Finally, the Eps set and the MinPts set are input into the DBSCAN algorithm for clustering. Based on the clustering results, when the number of clusters generated is the same for four consecutive times, the clustering result is considered stable, and the number of clusters is recorded as the optimal number of clusters. The parameter corresponding to the maximum k value of the optimal number of clusters is selected as the optimal parameter. Finally, based on the characteristics of the fence points, only the two longest clusters in the optimal clusters are retained to improve the fence line extraction rate. The test results are as follows: Figure 4 As shown, the fence points obtained by our method can effectively preserve the straight-line features of the fence, thus providing a precise navigation path for the robot. The experiment used a mobile intelligent disinfection robot for pigsties equipped with an Intel Core i7-5500U CPU industrial computer and a SLAMTEC S2 LiDAR.

[0042] A multi-step clustering point cloud denoising method for disinfection robots in water mist environments, the method comprising the following steps:

[0043] S1, Real-time acquisition of environmental point cloud data: A high-performance DTOF lidar sensor is used to acquire two-dimensional point cloud data streams in real time, and the polar coordinate parameters of each point are recorded synchronously, including radial distance and azimuth angle.

[0044] S2, Point cloud data preprocessing: Calculate the visible range of the lidar based on its location and the obstruction angle of the medicine box. Perform preliminary screening of the original point cloud based on angle constraints to suppress erroneous reflections caused by self-occlusion and improve the robustness of subsequent algorithms. Then, transform the screened point cloud data from polar coordinates to Cartesian coordinates.

[0045] In S2, the visible range of the lidar is calculated based on the lidar position and the obstruction angle of the medicine box, and corresponding angle constraints are added to filter out unnecessary point clouds. In this example, the point cloud filtering threshold is set to 30° to the left and right of the rear (i.e., the required point cloud azimuth angle θ is in the range of 30° to 330°).

[0046] The relative density constraint formula with Euclidean distance restrictions is expressed as follows:

[0047]

[0048] In the formula, k represents the total number of samples in the point set, i represents the label of the i-th sample, and p i Let rd represent the i-th point. i Let yi represent the relative density of the i-th point in the point set, and y0 be the y-axis distance constraint constant. i Here, is the y-coordinate of the i-th point, and n is the iteration variable for summation. This represents the Euclidean distance between the i-th and n-th points; in this example, the constraint constant y0 = 120cm is set.

[0049] S3, peripheral noise coarse filtering: The point cloud data is divided into a noise set and a point of interest set by a relative density constraint with Euclidean distance limit, which quickly removes peripheral irrelevant noise and improves processing efficiency;

[0050] S4, Misclassified Point Cloud Correction: For the true points of interest that have been misclassified into the noisy point set, the Shared Nearest Neighbor Clustering (SNNC) method is introduced to restore the misclassified points and ensure the integrity of the target.

[0051] In a preferred embodiment of the present invention, in step S4, the shared nearest neighbor clustering (SNNC) algorithm is used to correct misclassified points in the noisy point set, including the following sub-steps:

[0052] S4-1, find the k nearest neighbors of each of the m points in the noise set, and use k... n [i] represents, where k is a preset constant, if k n [i] and k n If [j] have an intersection, then point p will be... i and p j Save to k c In [i]; in this example, k = 12;

[0053] S4-2, Determine k c [i] and k c If [j] has an intersection, then let k = [j]. c [i] and k c The union of [j] is saved to k. c [i] and clear k c [j]

[0054] S4-3, Calculate all non-empty k c [i] contains the mean of the number of points in the cloud;

[0055] S4-4, The quantity of k is less than the average. c The point cloud in [i] is extracted and categorized into the set of points of interest;

[0056] S5, Water Mist Noise Removal: An improved adaptive density clustering (I-DBSCAN) algorithm is used to process the set of points of interest. Through a dynamic parameter optimization mechanism, water mist noise features are separated and accurately removed, and a high-precision fence point set is finally output.

[0057] In a preferred embodiment of the present invention, in step S5, an improved adaptive density clustering (I-DBSCAN) algorithm is used to refine the set of points of interest, including the following sub-steps:

[0058] S5-1, Generate the distance distribution matrix of the set of points of interest D:

[0059] D n×n ={Dist(i,j)∣1≤i≤n,1≤j≤n}

[0060] Among them, D n×n It is an n×n matrix, where n represents the total number of data objects in the dataset, and Dist(i,j) represents the distance from data point i to data point j;

[0061] S5-2, using the Gaussian smoothing formula to weight each data point, smooths the distance distribution matrix:

[0062]

[0063] Where G(x,y) represents the Gaussian weight value at position (x,y), x and y represent the positions of the pixel or data point in the horizontal and vertical directions, respectively, and σ represents the standard deviation of the Gaussian function, which determines the shape and smoothness of the Gaussian function.

[0064] S5-3: Sort each row of the distance distribution matrix after Gaussian smoothing in ascending order. In this distance distribution matrix, the first column element of each row represents the distance from the data object to itself, with a value of 0. The data elements located in the k-th column of the distance distribution matrix constitute the k-nearest neighbor vector D. k .

[0065] S5-4, for each D kThe k-mean nearest neighbor distance is calculated by averaging the values ​​and then placed into a set as candidate Eps parameters in order. This results in a set of candidate Eps parameters for all k-mean nearest neighbor distances. The number of data point objects in the neighborhood corresponding to each Eps parameter is calculated, and then the expected value of the Eps neighborhood data objects for all objects in dataset D is calculated. This generates the MinPts parameter set for D, as shown in the formula:

[0066]

[0067] Where n represents the total number of objects in the dataset, P i This represents the number of objects contained in the Eps neighborhood of the i-th object;

[0068] S5-5: Select the corresponding Eps parameter from the Eps parameter set and the corresponding MinPts parameter from the MinPts parameter set. Input these parameters into the I-DBSCAN algorithm and observe the clustering results. Obtain the cluster number results for different k values. If the generated cluster number results are the same for four consecutive times, it can be determined that the clustering has become stable. The resulting cluster number p is considered the optimal cluster number. Continue execution. When the generated cluster number is no longer equal to the optimal cluster number o, select the maximum k value corresponding to the cluster number equal to o as the optimal k value. The Eps parameter corresponding to the optimal k value is considered the optimal Eps parameter, the corresponding MinPts parameter is considered the optimal MinPts parameter, and the corresponding clustering result is considered the optimal result.

[0069] S5-6, the fence point set is extracted according to the distance feature of the two farthest points in the cluster to achieve the final denoising effect.

[0070] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A multi-step clustering point cloud denoising method for disinfection robots in water mist environments, characterized in that, The method includes the following steps: S1, Real-time acquisition of environmental point cloud data: A high-performance DTOF lidar sensor is used to acquire two-dimensional point cloud data streams in real time, and the polar coordinate parameters of each point are recorded synchronously, including radial distance and azimuth angle. S2, Point cloud data preprocessing: Calculate the visible range of the lidar based on its location and the obstruction angle of the medicine box. Perform preliminary screening of the original point cloud based on angle constraints to suppress erroneous reflections caused by self-occlusion and improve the robustness of subsequent algorithms. Then, transform the screened point cloud data from polar coordinates to Cartesian coordinates. S3, peripheral noise coarse filtering: The point cloud data is divided into a noise set and a point of interest set by a relative density constraint with Euclidean distance limit, which quickly removes peripheral irrelevant noise and improves processing efficiency; S4, Misclassified Point Cloud Correction: For the true points of interest that have been misclassified into the noisy point set, the Shared Nearest Neighbor Clustering (SNNC) method is introduced to restore the misclassified points and ensure the integrity of the target. S5, Water Mist Noise Removal: An improved adaptive density clustering I-DBSCAN algorithm is used to process the set of points of interest. Through a dynamic parameter optimization mechanism, water mist noise features are separated and accurately removed, and a high-precision fence point set is finally output.

2. The multi-step clustering point cloud denoising method for a disinfection robot in a water mist environment according to claim 1, characterized in that, In step S3, the relative density constraint formula with Euclidean distance limitation is expressed as follows: In the formula, k represents the total number of samples in the point set, i represents the label of the i-th sample, and p i Let rd represent the i-th point. i Let yi represent the relative density of the i-th point in the point set, and y0 be the y-axis distance constraint constant. i p is the y-coordinate of the i-th point, n is the iteration variable for summation, and p n It is the nth point during the summation iteration. This represents the Euclidean distance between the i-th point and the n-th point.

3. A multi-step clustering point cloud denoising method for a disinfection robot in a water mist environment according to claim 1 or 2, characterized in that, In step S4, the shared nearest neighbor clustering (SNNC) algorithm is used to correct misclassified points in the noisy point set, including the following sub-steps: S4-1, find the k nearest neighbors of each of the m points in the noise set, and use k... n [i] represents, where k is a preset constant, if k n [i] and k n If [j] have an intersection, then point p will be... i and p j Save to k c [i] in; S4-2, Determine k c [i] and k c If [j] has an intersection, then let k = [j]. c [i] and k c The union of [j] is saved to k. c [i] and clear k c [j]; S4-3, Calculate all non-empty k c [i] contains the mean of the number of points in the cloud; S4-4, The quantity of k is less than the average. c The point cloud in [i] is extracted and categorized into the set of points of interest.

4. A multi-step clustering point cloud denoising method for a disinfection robot in a water mist environment according to claim 1 or 2, characterized in that, In step S5, the improved adaptive density clustering I-DBSCAN algorithm is used to refine the set of points of interest, including the following sub-steps: S5-1, Generate the distance distribution matrix of the set of points of interest D: D n×n ={Dist(i,j)∣1≤i≤n,1≤j≤n} Among them, D n×n It is an n×n matrix, where n represents the total number of data objects in the dataset, i and j represent the i-th or j-th data point, and Dist(i,j) represents the distance from data point i to data point j. S5-2, using the Gaussian smoothing formula to weight each data point, smooths the distance distribution matrix: Where G(x,y) represents the Gaussian weight value at position (x,y), x and y represent the positions of the pixel or data point in the horizontal and vertical directions, respectively, and σ represents the standard deviation of the Gaussian function, which determines the shape and smoothness of the Gaussian function; S5-3: Sort each row of the distance distribution matrix after Gaussian smoothing in ascending order. In this distance distribution matrix, the first column element of each row represents the distance from the data object to itself, with a value of 0. The data elements located in the k-th column of the distance distribution matrix constitute the k-nearest neighbor vector D. k ; S5-4, for each D k The k-mean nearest neighbor distance is calculated by averaging the values ​​and then placed into a set as candidate Eps parameters in order. This results in a set of candidate Eps parameters for all k-mean nearest neighbor distances. The number of data point objects in the neighborhood corresponding to each Eps parameter is calculated, and then the expected value of the Eps neighborhood data objects for all objects in dataset D is calculated. This generates the MinPts parameter set for D, as shown in the formula: Where n represents the total number of objects in the dataset, i is the iteration variable for summation, and P i This represents the number of objects contained in the Eps neighborhood of the i-th object; S5-5: Select the corresponding Eps parameter from the Eps parameter set and the corresponding MinPts parameter from the MinPts parameter set. Input these parameters into the I-DBSCAN algorithm and observe the clustering results. Obtain the cluster number results under different k values. If the generated cluster number results are the same for four consecutive times, it can be determined that the clustering has become stable. The resulting cluster number p is considered to be the optimal cluster number. Continue execution. When the generated cluster number is no longer equal to the optimal cluster number p, select the maximum k value corresponding to the cluster number equal to p as the optimal k value. The Eps parameter corresponding to the optimal k value is considered to be the optimal Eps parameter, the corresponding MinPts parameter is considered to be the optimal MinPts parameter, and the corresponding clustering result is considered to be the optimal result. S5-6, the fence point set is extracted according to the distance feature of the two farthest points in the cluster to achieve the final denoising effect.

Images