Multi-parameter target classification and identification method based on photon counting laser radar

Abstract

The invention discloses a multi-parameter target classification and identification method based on a photon counting laser radar, and the method comprises the steps: collecting three-dimensional point cloud data through the photon counting laser radar, dividing a point cloud into a multi-scale grid neighborhood, and carrying out the classification and recognition of a multi-parameter target according to the divided multi-scale grid neighborhood, features such as a point cloud density feature, a three-dimensional coordinate variance feature, an upper and lower neighborhood density ratio feature and a horizontal neighborhood density ratio are extracted, all the extracted features are combined into a multi-dimensional feature vector, the length of the feature vector is the number of photon points, and each point has a plurality of corresponding features; inputting the obtained multi-dimensional feature vectors into a classifier to carry out target classification and identification to obtain an initial classification result; aiming at the initial classification result, correcting misclassification points by using a category self-adaptive grid optimization algorithm; and finally, outputting a final classification label result of each point. According to the invention, denoising and target classification can be realized at one time, and the calculation speed and precision are ensured.

Description

Technical Field

[0001] This invention relates to the field of laser detection technology, and in particular to a multi-parameter target classification and recognition method based on photon counting lidar. Background Technology

[0002] Photon-counting lidar is a highly sensitive detection system capable of detecting single photons. Compared to traditional full-waveform lidar, photon-counting lidar has advantages such as small size, high detection sensitivity, and large data sampling rate. It is particularly suitable for deployment on UAVs or satellite platforms because the equipment is lightweight, consumes little power, and can rapidly cover large-scale mapping areas.

[0003] After reconstructing 3D data from LiDAR, the resulting 3D point cloud can be used for target classification. Current point cloud classification methods mainly include: methods based on full waveform features: analyzing the morphology of the laser echo, including amplitude, width, and delay, to distinguish different land features, such as land, sea, and vegetation; classification based on geometric features of segmented segments: first, the point cloud is segmented, using methods such as region growing, plane fitting, and clustering, then the geometric and statistical features of the segmented segments are extracted, and supervised classifiers such as support vector machines and random forests are used for discrimination; and point feature-based classification: directly calculating the features of the points, and then using a classifier for discrimination.

[0004] However, in actual target classification, traditional full-waveform lidar is prone to waveform aliasing in certain specific areas, such as the land-sea interface. Existing research has attempted to combine green band and infrared intensity information to achieve land-sea segmentation. However, due to the superposition of multiple echo signals in shallow water areas, the waveform feature differences are weakened, easily leading to confusion between the sea surface and the seabed, limiting classification accuracy. Geometric feature classification based on segmented segments requires excellent segmentation algorithms. These methods are effective in ground feature point cloud classification, such as for buildings, but if the segmentation effect is poor, such as in the transition area between water and land, the segmentation is often unstable due to differences in point cloud density and signal aliasing, and the error is amplified in subsequent classifications. Classification based on point features lacks post-processing optimization mechanisms, resulting in isolated points and insufficient spatial continuity in the classification results. Especially for photon-counting lidar, the collected data point cloud has high density and many noisy points. Traditional classification methods first perform point cloud denoising; however, noise can also provide effective features for classification.

[0005] In summary, the existing technology has the following main shortcomings: 1. Traditional methods rely on full waveform characteristics, making it difficult to handle multi-echo interference in transitional regions between different media; 2. Algorithms that rely on point cloud segmentation are prone to producing unstable results in boundary regions. Photon counting lidar data is large and clustering is difficult. 3. Existing supervised classification methods fail to fully utilize the multi-channel information of photon counting lidar, and cannot effectively and quickly fuse multi-parameter features; 4. The lack of a post-processing optimization mechanism for point feature classification leads to problems such as isolated points and insufficient spatial continuity in the classification results; 5. It relies on pre-prepared denoising algorithms, making the processing flow complex. Summary of the Invention

[0006] This invention proposes a multi-parameter target classification and recognition method based on photon-counting lidar. It utilizes the three-dimensional point cloud information of photon-counting lidar to calculate the features of points and classify them. At the same time, it combines a new method of spatial continuity optimization to achieve denoising and target classification in one step, ensuring both calculation speed and accuracy.

[0007] A multi-parameter target classification and recognition method based on photon counting lidar includes the following steps: (1) Use photon counting lidar to collect three-dimensional point cloud data and divide the point cloud into multi-scale grid neighborhoods, including small cube grids, large cube grids, flat cuboid grids and vertical cuboid grids. (2) Based on the divided multi-scale grid neighborhood, extract point cloud density features, three-dimensional coordinate variance features, upper and lower neighborhood density ratio features, and horizontal neighborhood density ratio features. Combine all the extracted features into a multi-dimensional feature vector. The length of the feature vector is the number of photon points, and each point has multiple corresponding features. (3) Input the obtained multidimensional feature vector into the classifier to classify and identify the target, and obtain the initial classification result; (4) For the initial classification results, the class adaptive grid optimization algorithm is used to adaptively adjust the neighborhood discrimination and correct misclassified points by combining the spatial continuity features of different classes; finally, the final classification label result of each point is output.

[0008] In step (1), the side length of the small cube grid is 0.1~1m; the side length of the large cube grid is 3~10m; the length of the flat cuboid grid is 3~10m, the width is 1~5m, and the height is 0.1~1m; the length of the vertical cuboid grid is 0.5~1m, the width is 0.5~1m, and the height is 3~10m.

[0009] In this invention, the three-dimensional space of the point cloud is divided into small cubic meshes to reflect the characteristics at a small scale; the three-dimensional space of the point cloud is divided into large cubic meshes to reflect the characteristics at a large scale; the three-dimensional space of the point cloud is divided into flat cuboid meshes to reflect the characteristics at a horizontal scale; and the three-dimensional space of the point cloud is divided into vertical cuboid meshes to reflect the characteristics at a vertical scale.

[0010] In step (2), the point cloud density features need to be extracted from small cube meshes, large cube meshes, flat cuboid meshes, and vertical cuboid meshes. The specific process is as follows: Extract the number of all points in each grid and assign this number to all points in the corresponding grid as the point cloud density feature of each point (a total of four point cloud density features are obtained for each point), and save the grid index.

[0011] In step (2), the three-dimensional coordinate variance features only need to be extracted from a large cube mesh. The specific process is as follows: Given a large cube mesh, obtain the point cloud coordinates within it, assuming there are... If there are 1 point, then , , The variance of the direction is: ; ; ; in, , and Representing photon points , , The coordinates of the direction. This represents the variance value in a certain direction. , and Each represents all points within the grid. , , The mean of the directional coordinates, and the total variance of the point cloud within each large cube grid are: ; Among them, each large cube mesh The values ​​are assigned to all points within the grid as the 3D coordinate variance features of each point.

[0012] In step (2), the density ratio features of the upper and lower neighborhoods and the density ratio features of the horizontal neighborhood only need to be extracted using a small cube mesh. The specific process is as follows: Extract the number of all points within each small cube grid and use it as the density value for each grid. For each small cube mesh, find its adjacent small cube mesh above it, calculate the density ratio between the two, and assign the value to all points in that mesh as the upper neighborhood density ratio feature of each point. For each small cube grid, find its adjacent small cube grid below it, calculate the density ratio between the two, and assign the value to all points in that grid as the lower neighborhood density ratio feature of each point; For each small cube mesh, find its four horizontally adjacent small cube meshes, then calculate the ratio of its density to the total density of these four meshes, and assign this value to all points within that mesh as the horizontal neighborhood density ratio feature of each point.

[0013] If the photon counting lidar acquires four-channel 3D point cloud data, step (2) also requires extracting the density features and channel density ratio features of the remaining channels. The specific process is as follows: The point cloud of each channel is divided into the same multi-scale grid neighborhood, and the point cloud density value of the small cube grid in each channel is extracted. Obtain the point cloud density values ​​of the small cube mesh calculated in the second, third, and fourth channels, and assign the obtained values ​​to all points of the small cube mesh in the first channel as the density features of the remaining channels for each point; Calculate the point cloud density ratio of each small cube mesh in the first channel to the corresponding small cube mesh in the third channel. Assign the ratio to all points of the small cube mesh in the first channel, and each point obtains the density ratio feature of the first and third channels. Calculate the point cloud density ratio of each small cube mesh in the second channel to the corresponding small cube mesh in the fourth channel. Assign the ratio to all points of the small cube mesh in the first channel, and each point obtains the density ratio feature of the second and fourth channels.

[0014] In step (3), the classifier used is first pre-trained using the random forest algorithm.

[0015] In step (4), a category-adaptive mesh optimization algorithm is used, and the specific process is as follows: (4-1) Divide the grid neighborhood into optimization grids, count the number of different categories in each grid, and define the category with the most occurrences as the dominant category of that grid; (4-2) For different categories, based on their three-dimensional characteristics, determine several different neighbors as optimization references; (4-3) Combining the category statistics of itself and its neighbors, assign weights to neighbors, and add extra weights to the current dominant category to encourage continuity of the same category; (4-4) Consider the impact of optimizing the class within the neighborhood and update the dominant class for each grid; (4-5) Modify misclassified points based on the updated dominant category, and remove isolated noise points and isolated misclassified points.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The method of this invention does not rely on point cloud pre-segmentation or pre-denoising algorithms, and overcomes the problems of complex training, difficulty in classification in signal aliasing areas, inability to fully integrate multi-channel parameter features, and numerous initial classification errors in current point cloud classification algorithms. It successfully achieves accurate classification and identification of point clouds in a land-sea junction area, which has guiding significance for the practical application of photon counting lidar in cross-regional detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-parameter target classification and recognition method based on photon counting lidar according to an embodiment of the present invention.

[0019] Figure 2 This is a grid neighborhood diagram defined in an embodiment of the present invention.

[0020] Figure 3 This is a diagram showing the effect of the category-adaptive grid optimization algorithm used in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention, starting from practical needs and applications, studies the classification and identification problem of multi-channel UAV-borne photon counting lidar in integrated land and sea detection, determines parameters, extracts features, and classifies into four categories: noise, sea surface, land, and seabed.

[0023] like Figure 1 As shown, a multi-parameter target classification and recognition method based on photon counting lidar includes the following steps: Step S1: Point cloud data acquisition and preprocessing Three-dimensional point cloud data is acquired using a photon-counting lidar with four-channel polarization detection capability, and the neighborhood is divided using a multi-scale grid to construct a data structure suitable for feature calculation. For example... Figure 2As shown in (a) and (c), four types of grid neighborhood divisions were selected, with the small cube grid having a side length of 0.5m, the large cube grid having a side length of 5m, the flat cuboid grid having side lengths of 3m, 5m, and 0.5m respectively, and the vertical cuboid grid having side lengths of 1m, 1m, and 5m respectively.

[0024] S2. Multi-parameter feature extraction and fusion Point cloud density is calculated based on the four grid neighborhoods. Specifically, the number of points in each neighborhood is the density of that grid. Here, the density of small cube grids for the four channels, as well as the density of large cube grids, flat cuboid grids, and vertical cuboid grids for channel 1, are calculated.

[0025] The density ratio between neighborhoods is calculated based on the density of the small cube mesh in channel 1, such as... Figure 2 As shown in (b) and (d), the density ratios of the upper, lower, and horizontal neighborhoods were calculated. Specifically, the density ratios above and below each grid cell in the vertical direction were calculated as features, and these ratios were assigned to all points within the grid. Additionally, considering the unique characteristics of sea surface signals, the density ratio of the grid's horizontal neighborhood was calculated. This involved calculating the density of the four adjacent small cube grid cells in the horizontal direction of the grid, then calculating the density ratio, and assigning this ratio to the points within the grid.

[0026] The point cloud density of an effective signal is usually higher than that of background noise. To distinguish signals from land, sea surface, and seabed, the discreteness of the signal point cloud is also an important feature. Specifically, land is covered by vegetation and typically has significant elevation changes, while the sea surface is at a uniform height. Therefore, the variance of the 3D point cloud is also included in the feature calculation.

[0027] The main task is to calculate the variance of the 3D coordinates within each large cube grid in the point cloud data, and then assign the variance characteristics to all points within the grid. For a given grid, the point cloud coordinates are obtained, assuming the shape is... , indicating that there are within the grid If there are 1 point, then , , The variance of the direction is: ; ; ; in, , and Representing photon points , , The coordinates of the direction. This represents the variance value in a certain direction. , and Each represents all points within the grid. , , The mean of the directional coordinates, and the total variance of the point cloud within each large cube grid are: ; Among them, each large cube mesh The values ​​are assigned to all points within the grid as the 3D coordinate variance features of each point.

[0028] The experimental data acquisition device has four polarization channels. Channels 1 and 2 record the intensity information in the parallel polarization direction, and channels 3 and 4 record the intensity information in the perpendicular polarization direction. The ratio of the density between channels is used to represent the depolarization ratio.

[0029] The depolarization ratio is an important physical quantity describing the change in the polarization characteristics of electromagnetic waves (such as lasers), and is commonly used in lidar, remote sensing, and optical measurements. It reflects the change in polarization state after scattering or reflection from a target. Specifically, the depolarization ratio is defined as the ratio of the intensities of different polarization components after scattering from the target, and is usually expressed as: ; in, It is a deviation ratio. It is the intensity of the vertical polarization component of the reflected light, relative to the polarization direction of the incident light; This represents the intensity of the parallel polarization component of the reflected light. When the laser encounters the sea surface, the vertical polarization component in the returned signal is weak, while the parallel polarization component dominates, resulting in a small depolarization ratio. When the laser passes through the water and reaches the seabed, due to the effects of rough surfaces and volume scattering, the vertical polarization component in the returned signal strengthens, and the depolarization ratio increases.

[0030] Therefore, based on the grid density extracted from each channel, a corresponding index is determined for each grid. Grids with the same index value represent the same position in space. This not only enables the mapping of densities between different channels, assigning the density values ​​of the same points in channels 2, 3, and 4 to the points in channel 1, but also enables the calculation of density ratios between channels, assigning the density ratios of channel 1 and channel 3, and channel 2 and channel 4 to the points in channel 1, thus achieving feature fusion.

[0031] S3. The classifier performs point cloud classification. Random forest classifiers are ensemble learning methods based on decision trees, primarily used for classification and regression tasks. Their core idea is to classify the target by constructing multiple decision trees. Each point in a 3D point cloud dataset possesses 13 dimensions of features. Random forests, through random sampling and feature subset selection, effectively reduce the risk of overfitting, especially performing well with high-dimensional features. Furthermore, random forests provide an assessment of feature importance; by calculating the contribution of each feature to the split, the importance ranking of features can be derived, which is also beneficial for feature selection. Therefore, this invention selects random forest as the classifier and trains it using a divided training set.

[0032] S5, Optimization Algorithm Since this invention uses direct point classification, there will be points that are misclassified after the initial classification using a classifier. Most of these misclassified points are isolated. Therefore, to address this problem, a class-adaptive optimization algorithm is used to correct these erroneous points.

[0033] The optimization algorithm is based on the observation that different categories of objects exhibit varying degrees of spatial continuity. For example, trees on land extend continuously primarily vertically, while water surfaces extend continuously primarily horizontally. First, we tailor different neighbor selection strategies for different categories of meshes, enabling the optimization process to better adapt to the natural spatial distribution characteristics of point clouds for each category.

[0034] The category-adaptive grid optimization algorithm, after determining the neighbors of a grid, considers the neighbor categories, statistically analyzes the category distribution of neighboring grids, and assigns weights to neighbors based on the category statistics of the current dominant category. It also adds extra weight to the currently dominant category to encourage continuity within the same category. Thus, if all neighbors of a certain category are also of the same category, no modification is made. If the neighbors of a certain category are of a different category—for example, if most of the horizontal neighbors of a sea surface point are not sea surface points—it is verified whether it is an abnormal error point. Then, its vertical neighbors are considered; if most of the neighbors are now vegetation points, it is modified to increase continuity within the same category. Of course, if most of the points in a grid are classified as noise points, no further modification is made to avoid modification errors.

[0035] To address the shortcomings of existing point cloud classification methods, such as insufficient classification accuracy in land-sea interface areas and shallow water areas, over-reliance on point cloud segmentation, and failure to fully utilize the multi-channel features of photon-counting lidar, the method of this invention achieves high-precision classification and identification of land, sea surface, seabed, and noise points by fusing multi-channel polarization density features and depolarization ratio features, and combining them with a category adaptive optimization algorithm.

[0036] After dividing the training and test sets, a random forest classifier was trained using 8,908,914 labeled points. The trained model was then used to classify 2,227,229 points in the test set. The points were divided into four classes: noise, land, sea surface, and seabed. The corresponding confusion matrices are shown in Table 1.

[0037] Table 1 Based on the confusion matrix results in Table 1, the random forest classifier demonstrates good overall performance on the test set. Both classification precision and recall are high, with land classification achieving precision and recall of 99.51% and 99.45% respectively, indicating extremely high accuracy and stability in classifying land points. Precision for noise and sea surface categories is 95.89% and 97.08%, with recalls of 96.31% and 96.92%, showing robustness in these categories as well. However, seabed classification performs slightly worse, achieving a precision of 96.83% but a recall of 96.17%, slightly lower than other categories. Notably, there is some confusion between noise and sea surface points, manifested in a high number of noise points misclassified as sea surface (17886). Furthermore, misclassification of seabed points is mainly concentrated in noise points (9607), suggesting that further optimization of feature extraction or model parameters may improve classification performance. Overall, this method demonstrates strong classification capabilities and is suitable for classification tasks involving large-scale point cloud data. Experimental results show that the classification accuracy of the method of the present invention can reach 96%, and the Kappa coefficient reaches 96.26%. In particular, it performs significantly better than traditional methods in the transition zone between land and sea and shallow water areas, and has broad engineering application prospects.

[0038] Since this method directly classifies meshed point clouds, some misclassified points will exist after the initial classification. Figure 3 The image shows the classification results before and after using the optimized algorithm. It shows the point cloud classification results of a land-sea boundary area. On the basis of correct classification, the algorithm also performs noise reduction processing. The outline of the island can be seen, and the seabed in the nearshore area is identified and classified.

[0039] It should be understood that the method described in this invention is not limited to the specific embodiments disclosed in the specification and drawings. Without departing from the spirit and essence of this invention, those skilled in the art can make various substitutions or improvements to the feature parameters, algorithm models, feature extraction methods, and classifier types. For example, the classifier is not limited to random forests; support vector machines, deep neural networks, or other supervised learning methods can also be used. The neighborhood partitioning method can use multi-dimensional grids; the density ratio between multiple channels can also be modified according to the actual photon-counting lidar parameters; the optimization algorithm, in addition to the category-adaptive neighborhood selection strategy, can also be replaced with a specific adaptive neighborhood for a specific category. Any modifications, supplements, and equivalent substitutions made within the scope of the principles of this invention should be considered as included within the protection scope of this invention.

Claims

1. A multi-parameter target classification and recognition method based on photon counting lidar, characterized in that, Includes the following steps: (1) Use photon counting lidar to collect three-dimensional point cloud data and divide the point cloud into multi-scale grid neighborhoods, including small cube grids, large cube grids, flat cuboid grids and vertical cuboid grids. (2) Based on the divided multi-scale grid neighborhood, extract point cloud density features, three-dimensional coordinate variance features, upper and lower neighborhood density ratio features, and horizontal neighborhood density ratio features. Combine all the extracted features into a multi-dimensional feature vector. The length of the feature vector is the number of photon points, and each point has multiple corresponding features. (3) Input the obtained multidimensional feature vector into the classifier to classify and identify the target, and obtain the initial classification result; (4) For the initial classification results, the class adaptive grid optimization algorithm is used to adaptively adjust the neighborhood discrimination and correct misclassification points by combining the spatial continuity features of different classes; Finally, output the final classification label result for each point.

2. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, In step (1), the side length of the small cube grid is 0.1~1m; the side length of the large cube grid is 3~10m; the length of the flat cuboid grid is 3~10m, the width is 1~5m, and the height is 0.1~1m; the length of the vertical cuboid grid is 0.5~1m, the width is 0.5~1m, and the height is 3~10m.

3. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, In step (2), the point cloud density features need to be extracted from small cube meshes, large cube meshes, flat cuboid meshes, and vertical cuboid meshes. The specific process is as follows: Extract the number of all points in each grid and assign this number to all points in the corresponding grid as the point cloud density feature for each point.

4. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, In step (2), the three-dimensional coordinate variance features only need to be extracted from a large cube mesh. The specific process is as follows: Given a large cube mesh, obtain the point cloud coordinates within it, assuming there are... If there are 1 point, then , , The variance of the direction is: ； ； ； in, , and Representing photon points , , The coordinates of the direction. This represents the variance value in a certain direction. , and Representing all points within the grid , , The mean of the directional coordinates, and the total variance of the point cloud within each large cube grid are: ； Among them, each large cube mesh The values ​​are assigned to all points within the grid as the 3D coordinate variance features of each point.

5. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, In step (2), the density ratio features of the upper and lower neighborhoods and the density ratio features of the horizontal neighborhood only need to be extracted using a small cube mesh. The specific process is as follows: Extract the number of all points within each small cube grid and use it as the density value for each grid. For each small cube mesh, find its adjacent small cube mesh above it, calculate the density ratio between the two, and assign the value to all points in that mesh as the upper neighborhood density ratio feature of each point. For each small cube grid, find its adjacent small cube grid below it, calculate the density ratio between the two, and assign the value to all points in that grid as the lower neighborhood density ratio feature of each point; For each small cube mesh, find its four horizontally adjacent small cube meshes, then calculate the ratio of its density to the total density of these four meshes, and assign this value to all points within that mesh as the horizontal neighborhood density ratio feature of each point.

6. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, If the photon counting lidar acquires four-channel 3D point cloud data, step (2) also requires extracting the density features and channel density ratio features of the remaining channels. The specific process is as follows: The point cloud of each channel is divided into the same multi-scale grid neighborhood, and the point cloud density value of the small cube grid in each channel is extracted. Obtain the point cloud density values ​​of the small cube mesh calculated in the second, third, and fourth channels, and assign the obtained values ​​to all points of the small cube mesh in the first channel as the density features of the remaining channels for each point; Calculate the point cloud density ratio of each small cube grid in the first channel to the corresponding small cube grid in the third channel. Assign the ratio to all points of the small cube grid in the first channel, and each point obtains a density ratio feature of the first and third channels. Calculate the point cloud density ratio of each small cube grid in the second channel to the corresponding small cube grid in the fourth channel. Assign the ratio to all points of the small cube grid in the first channel, and each point obtains the density ratio feature of the second and fourth channels.

7. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, In step (3), the classifier used is first pre-trained using the random forest algorithm.

8. The multi-parameter target classification and recognition method based on photon counting lidar according to claim 1, characterized in that, In step (4), a category-adaptive mesh optimization algorithm is used, and the specific process is as follows: (4-1) Divide the grid neighborhood into optimization grids, count the number of different categories in each grid, and define the category with the most occurrences as the dominant category of that grid; (4-2) For different categories, based on their three-dimensional characteristics, Several distinct neighbors are identified as optimization references; (4-3) Combining the category statistics of itself and its neighbors, assign weights to neighbors, and add extra weights to the current dominant category to encourage continuity of the same category; (4-4) Consider the impact of optimizing the class within the neighborhood and update the dominant class for each grid; (4-5) Modify misclassified points based on the updated dominant category, and remove isolated noise points and isolated misclassified points.

Images