Low-altitude point cloud key point extraction method fusing feature engineering and deep learning

Abstract

The invention relates to a low-altitude point cloud key point extraction method fusing feature engineering and deep learning, and the method comprises the following steps: obtaining sample low-altitude point cloud data, and carrying out the denoising processing; based on the denoised sample low-altitude point cloud data, establishing neighborhood information to construct a neighborhood shape projection space, extracting and optimizing a neighborhood shape descriptor, and obtaining a spatial neighborhood descriptor; performing feature fusion based on the spatial neighborhood descriptors of the sample point clouds to obtain fusion features of the sample point clouds; based on the fusion features of the sample point clouds, key point coarse extraction is carried out, then key point fine extraction is carried out in combination with a non-maximum suppression strategy, and a key point extraction result is obtained; dividing a positive sample, a neutral sample and a negative sample through a key point extraction result of the sample low-altitude point cloud data, constructing a training sample set, and training a deep learning model through the training sample set; and performing key point extraction of the target low-altitude point cloud data by using the trained deep learning model.

Description

Technical Field

[0001] This invention relates to a method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning, belonging to the field of point cloud data processing technology. Background Technology

[0002] In typical low-altitude scenarios characterized by urban buildings, forests, and intertwined cables, the primary task for safe autonomous flight is to achieve three-dimensional perception and understanding of the environment. LiDAR, as a crucial sensor for this task, acquires three-dimensional point clouds that comprehensively record the geometric information of the environment. However, raw point clouds are characterized by disorder, large density variations, and strong noise interference, making them unsuitable for direct machine understanding and requiring processing steps such as feature extraction. Extracting discriminative and stable key points from large-scale low-altitude point cloud data is a critical step in this process and a vital foundation for supporting subsequent high-level tasks such as target recognition and semantic segmentation.

[0003] Currently, point cloud keypoint extraction methods are mainly divided into two categories. The first is traditional methods based on handcrafted features, such as ISS and Harris-3D. These methods select points with scale invariance and rotation invariance as keypoints based on local spatial characteristics. Although these methods have strong interpretability and fast computation speed, they fail to fully explore the deep information of the local space, resulting in low keypoint extraction accuracy. The second is supervised methods based on deep learning, such as USIP and D3Feat. These methods learn deep features to explore the characteristics of the local space and obtain better keypoints. However, most deep models rely on limited semantically labeled data for training, ignoring the inherent features of massive amounts of unlabeled data. Furthermore, relying solely on global semantic features obtained through deep learning is insufficient to overcome the complexity of ground features and data gaps in real-world urban low-altitude point cloud data. For example: Chinese patent CN120726326A, entitled "A Semantic Segmentation Method for Indoor Point Clouds to Enhance the Geometric Structural Features of Objects," proposes a point vector convolutional network (PVCNN) based on point vector feature decomposition (PVConv) and a three-layer PVConv to enhance the extraction of local features from point clouds. However, this method relies on a large amount of manually labeled data for training, limiting its applicability in scenarios with high labeling costs or scarce data. Furthermore, this patent primarily targets structured indoor scenes, and its ability to handle overlapping and occlusion of objects in complex scenes is limited.

[0004] Chinese patent CN120563857A, entitled "Robust Extraction Method of 3D Geometric Features of Precast Beams Based on Point Cloud Edge Topology Reorganization," involves projecting a 3D point cloud onto a 2D plane, extracting and ordering edge points, then performing directional clustering and line fitting. Finally, it calculates the intersection of adjacent edge lines to achieve robust extraction of feature corner points. However, this method relies on the accuracy of point cloud plane fitting and 2D projection. In scenarios with complex curved surfaces or large-scale occlusion, the ordering of edge points and directional clustering are easily interfered with, leading to failure or decreased accuracy in feature corner point extraction.

[0005] Chinese patent CN115239779A, entitled "A 3D Point Cloud Registration Method, Apparatus, Medium, and Electronic Equipment," proposes a method for storing point cloud feature points based on an octree data structure and introduces reinforcement learning to optimize the feature points in the octree, aiming to overcome the problem that traditional methods can only obtain local optima. However, this method relies on the correspondence between the source and target point clouds and lacks the ability to independently discriminate feature learning on a single frame of point cloud, making it impossible to directly extract key points from individual data.

[0006] Chinese patent CN113807366A, entitled "A Point Cloud Keypoint Extraction Method Based on Deep Learning," proposes a point cloud keypoint extraction model framework combining an attention mechanism and a Siamese network. This method extracts keypoints through point cloud preprocessing, local data sampling, and rotation transformation, utilizing a Siamese network. Although data augmentation reduces the reliance on semantic annotation, model training still requires pre-training on manually annotated point cloud datasets and loss constraints, failing to completely eliminate the need for manual annotation. Furthermore, the reliance on the complex Siamese framework and multi-objective loss function leads to slow model convergence and low training efficiency.

[0007] Chinese patent CN104143210A, entitled "A Multi-Scale Normal Feature Point Cloud Registration Method," extracts key points based on a multi-scale curvature consistency criterion. It calculates the curvature of each point in the point cloud within three different radius neighborhoods and selects points whose curvature changes at multiple scales are all less than a set threshold as key points. This method exhibits good geometric stability and repeatability, but its extraction process relies entirely on preset scale radii and curvature thresholds, resulting in weak adaptability under different point cloud densities and geometric characteristics.

[0008] Therefore, how to solve the problems of relying on manual prior knowledge, being sensitive to data quality, and having weak generalization ability in complex scenarios in existing technologies, and how to help intelligent interpretation of low-altitude point clouds, has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention proposes a method for extracting key points from low-altitude point clouds that integrates feature engineering and deep learning. The aim is to comprehensively utilize unsupervised manual features and supervised deep features to effectively extract key points from low-altitude point clouds, thereby improving the efficiency and accuracy of downstream tasks.

[0010] The technical solution of the present invention is as follows: On the one hand, this invention provides a method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning, comprising the following steps: Acquire sample low-altitude point cloud data and perform noise reduction processing on the sample low-altitude point cloud data; Based on the denoised low-altitude point cloud data, neighborhood information is established to construct a neighborhood shape projection space, and neighborhood shape descriptors are extracted. Based on the spatial relationship of the point cloud neighborhood, spatial distribution weights are constructed to optimize the neighborhood shape descriptors and obtain spatial neighborhood descriptors. Based on the spatial neighborhood descriptors of each sample point cloud, feature fusion is performed with the neighborhood points to obtain the fused features of each sample point cloud. Based on the fusion features of the point clouds of each sample, coarse key point extraction is performed, and then fine key point extraction is performed by combining a non-maximum suppression strategy to obtain the key point extraction results. The key point extraction results of the sample low-altitude point cloud data are divided into positive samples, neutral samples and negative samples to construct a training sample set, and a deep learning model is trained using the training sample set. Key points are extracted from target low-altitude point cloud data using a trained deep learning model. In a preferred embodiment, the step of denoising the sample low-altitude point cloud data includes: Establish a point cloud data index structure based on KD tree, and use the point cloud data index structure to search the neighborhood point set of each point in the sample low-altitude point cloud data with a preset radius; Calculate the density of the searched neighborhood point set, divide the dense region into dense region and discrete region according to the density and the preset density threshold, and filter out the points in the discrete region. A multi-influence factor coupled denoising model is constructed. The sample low-altitude point cloud data after filtering out points in discrete regions is input into the multi-influence factor coupled denoising model. The model filters out noise points in the sample low-altitude point cloud data through density factor and distance factor.

[0011] In a preferred embodiment, the step of extracting the neighborhood shape descriptor includes: Based on the denoised low-altitude point cloud data, a point cloud data indexing structure based on KD tree is used to obtain the neighborhood point set of each point with a preset radius. Based on the neighborhood point set of each point, a neighborhood matrix is ​​constructed. By performing singular value decomposition on the neighborhood matrix and reconstructing the eigenvector of the left singular matrix, a neighborhood shape projection space is established. The neighborhood shape descriptor is obtained by performing feature transformation on the original features of the corresponding point cloud through the neighborhood shape projection space.

[0012] In a preferred embodiment, the step of feature fusion based on the spatial neighborhood descriptors of each sample point cloud and the neighborhood points uses an aggregation function for feature fusion, the specific formula of which is: ; in, This represents the number of neighborhood points of the corresponding sample point cloud. It is a multidimensional aggregation function. For spatial neighborhood descriptors, For the first The fusion features of individual sample point clouds.

[0013] As a preferred embodiment, the steps of performing coarse key point extraction based on the fusion features of each sample point cloud, and then performing fine key point extraction by combining a non-maximum suppression strategy to obtain the key point extraction results are as follows: Based on the fusion features of the point clouds of each sample, the following formula is used for coarse extraction of key points: ; in, The label indicates whether it is a key point. This is a manually set threshold, where N is the total number of sample point clouds; Based on the key points extracted from the coarse extraction, a non-maximum suppression strategy is adopted for fine extraction, specifically including: Mark all coarsely extracted keypoints as unassigned; Randomly select an initial point from the key points marked as unassigned; Using the initial point as the center of the sphere, compare the fusion features of all neighborhood points within the preset radius, mark the point with the largest fusion feature as assigned, and discard the remaining neighborhood points; Repeat the above steps until all coarsely extracted keypoints are marked as assigned or discarded; The output points marked as assigned are used as key points after fine-tuning.

[0014] In a preferred embodiment, the step of dividing positive samples, neutral samples, and negative samples based on the key point extraction results of the sample low-altitude point cloud data includes: Based on the key point extraction results, the key points are used as positive samples, and the remaining points are used as initial negative samples. The probability distribution of the fusion features of each initial negative sample is statistically analyzed, and the value of the fusion feature of each initial negative sample is converted into a discrete label in the range of 0 to 1 using the cumulative distribution function. Based on the discrete label value range of each initial negative sample, the initial negative samples are divided into neutral samples and secondary negative samples. The second negative sample is used as the final negative sample.

[0015] In a preferred embodiment, the step of training the deep learning model using a training sample set includes: Based on the sample segmentation results, different sample point clouds are assigned corresponding labels; Establish a contrastive learning model and construct a three-class loss function based on the weighted cross-entropy function, with the three classes corresponding to the labels of positive, neutral, and negative samples, respectively; input the training sample set into the contrastive learning model for iterative training until the model reaches the preset accuracy or the iteration stops.

[0016] On the other hand, the present invention also provides a semantic segmentation method based on automatic key point extraction, comprising the following steps: The low-altitude point cloud key point extraction method based on any embodiment of the present invention, which integrates feature engineering and deep learning, automatically extracts key points from target low-altitude point cloud data. The autoencoder is pre-trained using extracted key points to learn the global geometric features of the target. Semantic segmentation of target low-altitude point cloud data based on pre-trained autoencoders.

[0017] In another aspect, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the low-altitude point cloud key point extraction method that integrates feature engineering and deep learning as described in any embodiment of the present invention.

[0018] Furthermore, the present invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the low-altitude point cloud keypoint extraction method that integrates feature engineering and deep learning as described in any embodiment of the present invention.

[0019] The technical effects of this invention are as follows: This invention extracts point cloud features and filters key point samples based on low-altitude point cloud data, enabling automatic key point extraction via a deep learning model. The method primarily employs a multi-influence factor coupled denoising model for point cloud denoising, addressing the challenge of handling complex noise in real-world point cloud scenarios where single denoising methods are insufficient. Furthermore, based on the probability distribution of point cloud features, the continuous feature distribution is converted into discrete numerical labels, flattening the original uneven probability distribution into a uniform distribution within the [0,1] interval. This constructs a data-independent scale to classify the criticality of samples, reducing the impact of sample imbalance on deep learning model training and improving the accuracy of automatic key point extraction.

[0020] Additional aspects and advantages of the invention will be set forth in the following description, and some of them will be obvious from the description, or may be learned by practicing the invention. Furthermore, various aspects and advantages of the invention may be realized and obtained by means of method steps and combinations particularly pointed out in the appended claims. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall method flow of Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the point cloud denoising process in an embodiment of the present invention; Figure 3 This is a schematic diagram of the spatial distribution of point cloud neighborhood shape features in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the process of coarse extraction and fine processing of point cloud data to obtain key points in an embodiment of the present invention. Figure 5 This is a schematic diagram showing the results of key point extraction in an embodiment of the present invention; Figure 6 This is a schematic diagram illustrating the training of a deep learning model by constructing samples from key points in an embodiment of the present invention. Figure 7 This is a schematic diagram of the sample partitioning process in an embodiment of the present invention; Figure 8 This is a schematic diagram of the method flow of Embodiment 2 of the present invention; Figure 9 This is an example of semantic segmentation in an embodiment of the present invention. Detailed Implementation

[0022] 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.

[0023] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0024] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0026] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0027] Example 1: See Figure 1 This embodiment provides a method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning, including the following steps: S100. Acquire sample low-altitude point cloud data and perform noise reduction processing on the sample low-altitude point cloud data. In one implementation, the denoising process in step S100 specifically employs a statistical filtering algorithm. A radius search range and a point count threshold are set, and each point in the sample low-altitude point cloud data is traversed. The number of neighboring points within the specified radius range is counted. If the number of neighboring point clouds is less than the set threshold, the point is determined to be a noise point and is removed to eliminate isolated noise points introduced by factors such as sensor noise and atmospheric scattering in the low-altitude environment, thereby improving the purity of the point cloud data.

[0028] Because real-world point cloud scenes contain multiple complex noises, single denoising methods and key point extraction algorithms are insufficient. Therefore, in another implementation of this embodiment, the denoising process in step S100 specifically employs a set of nearest neighbors based on the entire scene. A mathematical model is established by coupling multiple noise influencing factors to overcome the multiple noises in the scene. See details. Figure 2The noise reduction steps in this implementation are as follows: Step S110: Using KD-tree as a sample, low-altitude point cloud data Establish a spatial index structure, starting from any point. With the central node as the center, within a preset radius Search within its spherical neighborhood The nearest neighbor points eventually form the neighborhood point set. ,in Represents the three-dimensional coordinates of a point cloud.

[0029] Step S121: Count the number of points contained in the nearest neighbor set of each point. It is used as the density influence factor, and the calculation formula is as follows: ; in, In a point cloud scenario, The set of neighborhood points of the center point density, Represents the set of neighboring points The first in One point, This represents counting points within a space; Step S122: Based on the neighborhood point set of the entire scene, calculate the average distance within each neighborhood point set. The calculation formula is as follows: ; in, In a point cloud scenario, The set of neighborhood points of the center point average distance, Represents the set of neighboring points The first in One point, Indicated by The set of neighborhood points of the center point The number of points included; Step S123: Based on the density of the nearest neighbor set Average distance to neighboring points The noise points are identified and filtered out using the following formula: ; ; in, This is the noise threshold. For hyperparameters, It is a manually set noise threshold, when When it is considered noise, it is discarded. In this embodiment, it is set... , , .

[0030] Step S130: To eliminate the influence of dimensions between different features of the point cloud, this step involves processing the denoised sample point cloud data from steps S121 to S123. Normalization is performed. This point cloud data contains spatial coordinates (XYZ), color information (RGB), and intensity (I), among other features. Its total dimension is determined by the data itself. The normalization function is shown below. After normalization, the standardized point cloud data is obtained. : ; in, This represents the modulo operation on a vector. This is the standardized point cloud data.

[0031] S200. Based on the denoised sample low-altitude point cloud data, establish neighborhood information to construct a neighborhood shape projection space and extract neighborhood shape descriptors; optimize the neighborhood shape descriptors by constructing spatial distribution weights based on the point cloud neighborhood spatial relationship to obtain spatial neighborhood descriptors.

[0032] In one embodiment of this example, step S200 can first construct a local neighborhood for each point in the sample low-altitude point cloud data using KNN or spherical neighborhood, construct a projection space such as a three-dimensional orthogonal basis / local coordinate system based on the neighborhood point set, and extract shape descriptors such as PFH / FPFH / SHOT; then calculate the spatial distribution features such as the density and Euclidean distance of the neighborhood points, design spatial weights with distance decay / density weighting, and optimize the shape descriptors to obtain spatial neighborhood descriptors.

[0033] To enhance spatial features, please refer to the following for details. Figure 3 In another embodiment of this implementation, step S200 specifically includes the following steps: S201, Based on denoised sample low-altitude point cloud data A spatial index relationship is constructed using a KD-tree, and with Extract a spherical neighborhood for each point with a radius, forming a point set containing k neighborhood points. .

[0034] S202. The key points extracted in this embodiment are robust feature points with rotation and scale invariance, which remain consistent after the point cloud undergoes translation, rotation, or scaling transformations. Therefore, based on the neighborhood point set of each point... Constructing a neighborhood matrix representing local features And establish the neighborhood shape projection space through its singular value decomposition (SVD) matrix. The neighborhood matrix The calculation formula is as follows: ; ; in, Indicated by The point set centered The Middle Features of neighboring points of a given point It is a right singular value matrix. It is a left singular value matrix. It is a diagonal matrix.

[0035] S203, the right singular matrix obtained in step S202 The eigenvectors are reconstructed to establish a neighborhood shape projection space that is rotation-invariant and scale-invariant. The calculation formula is as follows: ; in, , Right singular value matrices The first and second eigenvectors; Point cloud features point by point Projecting space through neighborhood shape Projecting the neighborhood shape descriptor of each point yields the shape descriptor. The calculation formula is as follows: ; in, This is the spatial projection function, which in this embodiment is matrix multiplication.

[0036] S204. Considering the differences in the spatial distribution of point clouds, local point sets... Points at different locations within the center point have varying degrees of influence. Therefore, this embodiment constructs a spatial distribution weight based on the spatial location of each point. And use the weights to apply to the aforementioned neighborhood shape descriptors The updated spatial neighborhood descriptor is obtained. The aforementioned and The calculation method is as follows: ; ; in, For distance calculations in Euclidean space. Indicated by The point set centered The Middle The neighboring points of each point, For point set The variance and mean of the distribution of distances from all points within the range to the center point.

[0037] S300. Based on the spatial neighborhood descriptors of each sample point cloud, feature fusion is performed with neighboring points to obtain the fused features of each sample point cloud; specifically as follows: By using an aggregation function, the spatial neighborhood descriptors of the sample point cloud are aggregated, and point-by-point fused features are output. : ; in, Represents local neighborhood Number of dots Spatial neighborhood descriptors the number of rows, For spatial neighborhood descriptors The element in the i-th row and j-th column.

[0038] S400. Based on the fusion features of the point clouds of each sample, coarse key point extraction is performed, followed by fine key point extraction using a non-maximum suppression strategy, to obtain the key point extraction results; the specific steps are as follows: S410, First, based on fusion features An adaptive decision function is used to coarsely extract key points from the point cloud. The decision formula is as follows: ; in, The label indicates whether it is a key point. It is a manually set threshold, in this example Set as a fusion feature The sum of the peak value and variance of the probability density function.

[0039] S420. To address the clustered distribution and redundancy issues in the coarse extraction of key points, a non-maximum suppression method is employed, based on its fusion characteristics. The process involves fine-tuning. The specific steps are as follows: S421. Mark all points extracted in the coarse extraction as unassigned. S422. Randomly select an initial point from the key points marked as unassigned. S423. Using the initial point as the center of the sphere, compare the aggregated spatial neighborhood descriptors of all points within the spherical neighborhood with a radius of sphere. Mark the point corresponding to the largest aggregated spatial neighborhood descriptor as assigned and discard the remaining unmarked points in the neighborhood. S424. Repeat steps S422~S423 until all points are marked as assigned; S425. Output all assigned points as key points after fine-tuning.

[0040] Based on the above steps, the phased extraction process of key points in point clouds and the final extraction results are as follows: Figure 4 and Figure 5 As shown.

[0041] S500, see details Figure 6 Based on the keypoints obtained, suitable positive and negative samples are selected for use in the deep learning network, ultimately serving downstream tasks. First, the keypoint extraction results from the sample low-altitude point cloud data are used to divide the data into positive, neutral, and negative samples to construct a training sample set. The deep learning model is then trained using this training sample set. The specific steps are as follows: S510. Based on the key point extraction results, the key points extracted after fine processing are used as positive samples, and the remaining points are used as initial negative samples. S520. To resolve the imbalance between positive and negative samples, the negative samples are re-divided. The specific steps are as follows: S521. Calculate the probability distribution of the fusion features of each initial negative sample. The fusion feature data of each initial negative sample will form a continuous, smooth probability density distribution on the number axis. This distribution is usually non-uniform (e.g., a large number of points are concentrated near certain common values). The cumulative distribution function is used to convert the value of the fusion feature of each initial negative sample into a discrete label in the interval [0,1]; that is, mapping the original feature value of each point to its percentile rank in the entire sample. This transformation "flattens" the original non-uniform distribution into a uniform distribution in the interval [0,1], thus constructing a fair "difficulty scale" independent of the data. The construction process of this function is as follows: Assuming the original continuous fusion features are The converted discrete labels are ,at the same time , There exists a transformation function: ; and Satisfying At the same time, an inverse transform function also exists: ; Also satisfies exist It is monotonically increasing within the range, and .

[0042] Assumption for The probability density function, for The probability density function of the distribution function, according to the definition of the distribution function, is:

[0043] in For continuous variables The distribution function. Since the probability density function is the derivative of the distribution function, then: ; From the above formula, we can derive that by transformation It can change the original continuous probability density function Thus, the discrete probability density function is obtained. .

[0044] From the previous step, we can conclude that: ; And because It is a uniform discrete distribution, therefore Then it exists: ; Integrating it, we can obtain: ; Therefore, there exists a transformation formula such that... .

[0045] S522. Neutral samples and secondary negative samples are divided by the discrete label value range of each initial negative sample (i.e., uniformly distributed values ​​in the [0,1] interval). The lowest-ranked point has the least critical features and is the easiest to distinguish from the positive samples, and is classified as a secondary negative sample; while the middle-ranked points (such as those between 0.4 and 0.6) have ambiguous features and are the easiest to be confused by the model. They are the points that the model needs to focus on learning to distinguish, and are classified as neutral samples.

[0046] The role of neutral samples is to fill the discriminative gap between positive and negative samples. They represent the class of negative samples that are most similar to positive samples and are most easily misclassified, often referred to as "hard negative samples" or "boundary samples." By incorporating neutral samples into contrastive learning and assigning them appropriate weights, the model can perform reinforcement learning on these "most difficult-to-distinguish samples" during the pre-training stage, thereby enhancing the downstream semantic segmentation model's ability to handle complex scenes and detailed boundaries.

[0047] S530. The secondary negative samples are used as the final negative samples, thus obtaining the final positive, neutral, and negative samples. Differential labels are then assigned to samples of different categories. The process of reconstructing these sample labels is as follows: Figure 7 As shown, the original labels were only binary positive and negative samples of 0 and 1, resulting in extreme sample imbalance. After sample label reconstruction, the labels were replaced with three-class labels of 0, 1, and 2, where 2 is a neutral sample label, which significantly reduced the impact of sample imbalance on model training.

[0048] S600: Use a trained deep learning model to extract key points from target low-altitude point cloud data. In one embodiment of this example, step S500, which involves training a deep learning model using a training sample set, specifically includes: Based on the sample segmentation results, different sample point clouds are assigned corresponding labels; Establish a contrastive learning model and construct a three-class loss function based on the weighted cross-entropy function, with the three classes corresponding to the labels of positive, neutral, and negative samples, respectively; input the training sample set into the contrastive learning model for iterative training until the model reaches the preset accuracy or the iteration stops.

[0049] Specifically, the loss function is formulated as follows:

[0050] in, is the number of samples of class i that appear during training (corresponding to positive, neutral, and negative samples); N is the total number of samples in the training data; c represents the number of sample classes. , Let represent the true label vector and the model-predicted label vector of the i-th sample, respectively.

[0051] Example 2: See details Figure 8 This embodiment provides a semantic segmentation method based on automatic key point extraction, including the following steps: The low-altitude point cloud key point extraction method based on the fusion of feature engineering and deep learning described in Embodiment 1 above automatically extracts key points from target low-altitude point cloud data. The autoencoder is pre-trained using extracted key points to learn the global geometric features of the target. Semantic segmentation of target low-altitude point cloud data based on pre-trained autoencoders.

[0052] like Figure 9 As shown, the segmentation results in three different scenarios verify the effectiveness of the semantic segmentation method based on key point automatic extraction presented in this embodiment.

[0053] Example 3: This embodiment proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the low-altitude point cloud key point extraction method that integrates feature engineering and deep learning as described in any embodiment of the present invention.

[0054] Example 4: This embodiment proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a low-altitude point cloud keypoint extraction method that integrates feature engineering and deep learning as described in any embodiment of the present invention.

[0055] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0056] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0057] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0058] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning, characterized in that, Includes the following steps: Acquire sample low-altitude point cloud data and denoise the data. Based on the denoised low-altitude point cloud data, establish neighborhood information to construct a neighborhood shape projection space and extract neighborhood shape descriptors. Optimize the neighborhood shape descriptors by constructing spatial distribution weights based on the spatial relationships of the point cloud neighborhoods to obtain spatial neighborhood descriptors. Based on the spatial neighborhood descriptors of each sample point cloud, perform feature fusion with neighboring points to obtain the fused features of each sample point cloud. Based on the fusion features of each sample point cloud, key points are coarsely extracted, and then refined by combining a non-maximum suppression strategy to obtain key point extraction results. The key point extraction results of the sample low-altitude point cloud data are divided into positive samples, neutral samples and negative samples to construct a training sample set. A deep learning model is trained using the training sample set. The trained deep learning model is then used to extract key points from the target low-altitude point cloud data.

2. The method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning according to claim 1, characterized in that, The steps for denoising the sample low-altitude point cloud data include: Establish a point cloud data index structure based on KD tree, and use the point cloud data index structure to search the neighborhood point set of each point in the sample low-altitude point cloud data with a preset radius; Calculate the density of the searched neighborhood point set, divide the dense region into dense region and discrete region according to the density and the preset density threshold, and filter out the points in the discrete region. A multi-influence factor coupled denoising model is constructed. The sample low-altitude point cloud data after filtering out points in discrete regions is input into the multi-influence factor coupled denoising model. The model filters out noise points in the sample low-altitude point cloud data through density factor and distance factor.

3. The method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning according to claim 1, characterized in that, The step of extracting neighborhood shape descriptors includes: Based on the denoised low-altitude point cloud data, a point cloud data indexing structure based on KD tree is used to obtain the neighborhood point set of each point with a preset radius. Based on the neighborhood point set of each point, a neighborhood matrix is ​​constructed. By performing singular value decomposition on the neighborhood matrix and reconstructing the eigenvector of the left singular matrix, a neighborhood shape projection space is established. The neighborhood shape descriptor is obtained by performing feature transformation on the original features of the corresponding point cloud through the neighborhood shape projection space.

4. The method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning according to claim 1, characterized in that, In the step of feature fusion based on the spatial neighborhood descriptors of each sample point cloud and the neighboring points, an aggregation function is used for feature fusion, and the specific formula is as follows: ； in, This represents the number of neighborhood points of the corresponding sample point cloud. It is a multidimensional aggregation function. For spatial neighborhood descriptors, For the first The fusion features of individual sample point clouds.

5. The method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning according to claim 1, characterized in that, The steps for coarsely extracting key points based on the fusion features of each sample point cloud, and then finely extracting key points by combining a non-maximum suppression strategy, to obtain the key point extraction results are as follows: Based on the fusion features of the point clouds of each sample, the following formula is used for coarse extraction of key points: ； in, The label indicates whether it is a key point. This is a manually set threshold, where N is the total number of sample point clouds; Based on the key points extracted from the coarse extraction, a non-maximum suppression strategy is adopted for fine extraction, specifically including: Mark all coarsely extracted keypoints as unassigned; Randomly select an initial point from the key points marked as unassigned; Using the initial point as the center of the sphere, compare the fusion features of all neighborhood points within the preset radius, mark the point with the largest fusion feature as assigned, and discard the remaining neighborhood points; Repeat the above steps until all coarsely extracted keypoints are marked as assigned or discarded; The output points marked as assigned are used as key points after fine-tuning.

6. The method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning according to claim 1, characterized in that, The steps for classifying positive, neutral, and negative samples based on key point extraction results from low-altitude point cloud data include: Based on the key point extraction results, the key points are used as positive samples, and the remaining points are used as initial negative samples. The probability distribution of the fusion features of each initial negative sample is statistically analyzed, and the value of the fusion feature of each initial negative sample is converted into a discrete label in the range of 0 to 1 using the cumulative distribution function. Based on the discrete label value range of each initial negative sample, the initial negative samples are divided into neutral samples and secondary negative samples. The second negative sample is used as the final negative sample.

7. The method for extracting key points from low-altitude point clouds by integrating feature engineering and deep learning according to claim 6, characterized in that, The step of training a deep learning model using a training sample set includes: Based on the sample segmentation results, different sample point clouds are assigned corresponding labels; Establish a contrastive learning model and construct a three-class loss function based on the weighted cross-entropy function, with the three classes corresponding to the labels of positive, neutral, and negative samples, respectively; input the training sample set into the contrastive learning model for iterative training until the model reaches the preset accuracy or the iteration stops.

8. A semantic segmentation method based on automatic keypoint extraction, characterized in that, Includes the following steps: The low-altitude point cloud key point extraction method based on the fusion of feature engineering and deep learning as described in any one of claims 1 to 7 automatically extracts key points from target low-altitude point cloud data. The autoencoder is pre-trained using extracted key points to learn the global geometric features of the target. Semantic segmentation of target low-altitude point cloud data based on pre-trained autoencoders.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the low-altitude point cloud key point extraction method that integrates feature engineering and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the low-altitude point cloud keypoint extraction method that integrates feature engineering and deep learning as described in any one of claims 1 to 7.

Images