A deformable point cloud neighborhood selection method and its use

Abstract

The present invention relates to a deformable point cloud neighborhood selection method and its use. By compressing the point cloud in a specific direction, the shape of the neighborhood selected by the K-nearest neighbor algorithm is controlled. This method can directly replace the K-nearest neighbor algorithm, enabling the neighborhood selection of the deep neural network to be deformable. The beneficial effects mainly include: the neighborhood shape can be transformed as needed, effectively overcoming the weakness of the traditional K-nearest neighbor algorithm in having a fixed neighborhood shape. Since the neighborhood space is changed through a transformation matrix, the calculation of complex irregular neighborhood shapes is avoided, effectively reducing the computational complexity. This method can replace the process of using the K-nearest neighbor to select the neighborhood of feature points and can be used to optimize all existing point cloud segmentation methods that use the K-nearest neighbor algorithm.

Description

Technical Field

[0001] The present invention belongs to the application of deep neural networks in 3D computer vision, and relates to a deformable point cloud neighborhood selection method and its use. Background Art

[0002] With the rapid development of 3D point cloud sensors, point clouds have been widely used in fields such as autonomous driving and remote sensing. Point cloud recognition and segmentation algorithms based on deep neural networks have gradually become a popular research direction. The vast majority of existing deep neural network methods use the K-nearest neighbor algorithm to select the range for local feature extraction, and use methods such as multi-layer perceptrons with shared weights and attention modules to calculate local features. The semantic features of each point are extracted through an encoder-decoder network structure, and finally the full connection layer is used to achieve the recognition and segmentation of point clouds.

[0003] Dai Lu, Wang Junliang, etc. ("A Non-Equivalent Point Cloud Segmentation Method Based on Convolutional Neural Networks", Journal of Donghua University (Natural Science Edition), 2019, 45(6): 862-868) proposed a point cloud segmentation neural network based on CNN for the non-equivalence in point cloud segmentation. On the basis of designing a network random sampling layer and a max pooling layer to solve the variable data volume and order of point clouds, a distance matrix after the action of a penalty function was introduced to weight the classification errors of each point, and the model training method was optimized. However, due to the inaccuracy in the design of the artificially designed penalty function, this method cannot ensure the optimal model training effect.

[0004] The vast majority of existing deep neural network methods use the K-nearest neighbor algorithm to select the range for local feature extraction. The shape of the neighborhood selected by this method is relatively fixed. Through the analysis of existing algorithms, it is found that methods with deformable neighborhood shapes usually have higher accuracy than methods using the K-nearest neighbor algorithm. Summary of the Invention

[0005] Technical Problems to be Solved

[0006] In order to avoid the deficiencies of the prior art, the present invention proposes a deformable point cloud neighborhood selection method and its use, which is applicable to the vast majority of existing deep neural network methods. Without changing the feature calculation method, the K-nearest neighbor algorithm is replaced to make the neighborhood shape of the existing method variable, so as to improve the accuracy of the existing method.

[0007] Technical Solution

[0008] A deformable point cloud neighborhood selection method, characterized by the following steps:

[0009] Step 1: Randomly select a feature point p in the cloud key , the point cloud contains N points, p iRepresents the point cloud The three-dimensional coordinates of the i-th point in

[0010] Step 2: Multiply each point in the point cloud by the transformation matrix T to obtain the transformed point cloud

[0011]

[0012] The transformation matrix T = diag(t x , t y , t z ), where t x , t y , t z are the scale transformation parameters of the point cloud on the x-axis, y-axis, and z-axis respectively;

[0013] Step 3: Calculate the transformed feature point p key ′ = p key T;

[0014] Step 4: Use the K-nearest neighbor algorithm to calculate the K nearest points of the feature point p key ′ in the point cloud and record the numbers of these K points. Select the corresponding points in the original point cloud to form the neighborhood of the feature point p key .

[0015] The value range of the scale transformation parameter is t x , t y , t z ∈(0, 1].

[0016] An application of the deformable point cloud neighborhood selection method described above, characterized in that: replacing the process of using the K-nearest neighbor to select the feature point neighborhood, and optimizing all existing point cloud segmentation methods using the K-nearest neighbor algorithm.

[0017] Beneficial effects

[0018] A deformable point cloud neighborhood selection method and its application proposed by the present invention controls the shape of the neighborhood selected by the K-nearest neighbor algorithm by compressing the point cloud in a specific direction. This method can directly replace the K-nearest neighbor algorithm, making the neighborhood selection of the deep neural network deformable.

[0019] The beneficial effects of adopting the method of the present invention mainly include:

[0020] (1) The point cloud neighborhood selection method proposed by the present invention can transform the neighborhood shape according to needs, effectively overcoming the weakness of the traditional K-nearest neighbor algorithm in selecting a fixed neighborhood shape.

[0021] (2) By changing the neighborhood space through the transformation matrix, the complex calculation of irregular neighborhood shapes is avoided, effectively reducing the computational complexity.

[0022] (3) This method can replace the process of selecting the neighborhood of feature points using K-nearest neighbors and can be used to optimize all existing point cloud segmentation methods that use the K-nearest neighbor algorithm. Description of the Drawings

[0023] Figure 1 is the flowchart of the deformable neighborhood selection method

[0024] Figure 2 is the test result diagram Detailed Implementation Manner

[0025] The present invention will be further described in conjunction with embodiments and drawings:

[0026] The present invention is a deformable point cloud neighborhood selection method, and the process is as Figure 1 shown. Taking the PointTransformer point cloud segmentation network as an example, the specific implementation manner of the present invention is described, but the technical content of the present invention is not limited to the described scope. The specific implementation manner includes the following steps:

[0027] Step 1: Analyze the network structure. The PointTransformer point cloud segmentation neural network includes 5 feature encoding layers, 5 feature decoding layers, and 3 classification layers. Only the 5 feature encoding layers are operated here. Each feature encoding layer is composed of a neighborhood selection module and a feature extraction module connected in series. The dimensionalities of the feature vectors output by the 5 feature selection modules are 32, 64, 128, 256, and 512 respectively.

[0028] Step 2: Modify the network structure using the method of the present invention. For the first encoding layer, set four transformation matrices T 1 、T 2 、T 3 、T 4 , which are respectively:

[0029] T 1 =diag(1,1,1)

[0030] T 2 =diag(0.5,1,1)

[0031] T 3 =diag(1,0.5,1)

[0032] T 4 =diag(1,1,0.5)

[0033] Feature points are selected from the input point cloud. After transformation using four transformation matrices, four sets of neighborhood points are obtained using the K-nearest neighbor algorithm respectively. Four new feature extraction modules are constructed, with the same structure as the original feature extraction module, but only changing the output vector dimension to one-fourth of the original, that is, 8. Each module corresponds to a set of neighborhood points. The feature vectors of each set of neighborhood points are calculated using the feature extraction module. Finally, the four feature vectors are concatenated to obtain the final output feature of this layer.

[0034] For the remaining feature encoding layers, the same modifications are made.

[0035] Step 3: Train the neural network. Use the STPLS3D data and the backpropagation algorithm to train the neural network parameters until the model error converges.

[0036] Step 4: Use the point cloud segmentation network model trained in the above steps to segment the test samples, and calculate the class average intersection over union, which reaches 55.96%, an improvement of 8.86%. The test results are as Figure 2 shown.

Claims

1. A deformable point cloud neighborhood selection method, characterized in that the steps are as follows: Step 1: Randomly select a feature point p from the point cloud key . The point cloud contains N points, and p i represents the three-dimensional coordinates of the i-th point in the point cloud . Step 2: Multiply each point in the point cloud by the transformation matrix T to obtain the transformed point cloud The transformation matrix T = diag(t x , t y , t z ), where t x , t y , t z are the scale transformation parameters of the point cloud on the x-axis, y-axis, and z-axis, respectively; Step 3: Calculate the transformed feature point p key ′ = p key T; Step 4: Use the K-nearest neighbor algorithm to calculate the K nearest points of the feature point p key ' in the point cloud and record the serial numbers of these K points. Select the corresponding points in the original point cloud according to the serial numbers, which form the neighborhood of the feature point p key .

2. The deformable point cloud neighborhood selection method according to claim 1, characterized in that: The value range of the scale transformation parameter is t x , t y , t z ∈(0, 1].

3. Use of the deformable point cloud neighborhood selection method according to claim 1 or 2, characterized in that: Replace the process of using K-nearest neighbors to select the feature point neighborhood, and optimize all existing point cloud segmentation methods that use the K-nearest neighbor algorithm.

Images