A point cloud data processing method, system and computer device based on global outlier feature suppression

Abstract

This invention provides a point cloud data processing method, system, and computer device based on global outlier feature suppression. The method first acquires point cloud data, then uses M point cloud sampling channels to perform convolutional coding feature extraction, obtaining M feature output information. Next, it calculates the Euclidean distance between each point cloud and other point clouds, obtaining distance information for each point cloud. Based on this distance information, it calculates the estimated local contribution of each point cloud, and performs multi-layer convolutional layer fitting to obtain a point cloud weight score for measuring outlier information. Based on the point cloud weight score, it fuses the M feature output information to obtain the first feature suppression result. This invention makes feature vectors far from the center point approach 0, while feature vectors near the center point remain almost unchanged, achieving suppression of specific point cloud features based on global outliers, improving classification accuracy, and compensating for the lack of spatial awareness.

Description

Technical Field

[0001] This invention belongs to the field of intelligent point cloud data analysis technology, and in particular relates to a point cloud data processing method, system and computer equipment based on global outlier feature suppression. Background Technology

[0002] With the development of computer and electronic technologies, computer vision technology has been widely applied in many fields. For example, in autonomous driving technology, computer vision is often used to analyze the road environment and then drive the vehicle based on the analysis results.

[0003] In the field of autonomous driving, point cloud data is needed to collect and analyze road environment information. Point cloud classification can help autonomous driving systems identify and classify various obstacles on the road, such as vehicles, pedestrians, and bicycles. However, in the process of point cloud data classification, the accuracy of point cloud classification is low due to the influence of sensor measurement accuracy and noise, as well as the potential imbalance in the number of samples of different categories.

[0004] In existing technologies, the PointNet++ model is generally used to process point cloud data and determine object categories from the overall point cloud data. Specifically, such as... Figure 1 As shown, it includes the following steps:

[0005] (1) Input the point cloud data A to be processed, and obtain the center point by sampling the farthest point;

[0006] (2) Input the whole frame point cloud data A into the single-channel KNN (K nearest neighbor) sampling module S1 to perform point cloud group sampling (sampling 32 points) to obtain the sampling neighborhood C of each center point;

[0007] (3) Input all the sampling neighborhoods C into the first layer PointNet (point cloud feature extraction network) P1, perform convolutional encoding to increase the dimensionality of the point cloud to achieve feature extraction, and leave the most significant feature R after Maxpooling filtering;

[0008] (4) Input the output feature R of Maxpooling into the single-channel KNN sampling module S2 of the second layer. The second layer PointNet (P2) will extract features from the sampled point cloud of S2.

[0009] It is evident that while existing methods utilize neighborhood information of point clouds, they fail to consider outlier characteristics. Outlier characteristics reflect the distance distribution of points within a point cloud dataset, providing local and global information about the shape and structure of the point cloud. Point cloud classification, in particular, focuses on the overall characteristics of the point cloud. Therefore, current point cloud data processing methods cannot achieve efficient classification, resulting in a loss of spatial information. Summary of the Invention

[0010] In order to overcome the shortcomings of the prior art, the present invention aims to provide a point cloud data processing method, system and computer device based on global outlier feature suppression, which is mainly used to solve the defects of low classification accuracy and missing spatial information caused by the failure to consider the outlier features of point cloud when processing point cloud data in the prior art.

[0011] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0012] In a first aspect, the present invention provides a point cloud data processing method based on global outlier feature suppression, comprising:

[0013] Acquire point cloud data, and use M point cloud sampling channels to perform convolutional coding feature extraction on the point cloud data to obtain M feature output information;

[0014] Calculate the Euclidean distance between each point cloud and other point clouds in the point cloud data to obtain the distance information for each point cloud. Based on the distance information, calculate the local contribution estimate of the point cloud. After fitting through multiple convolutional layers, obtain the point cloud weight score used to measure the outlier information of the point cloud.

[0015] Based on the point cloud weight scores, the output information of M features is fused to obtain the first feature suppression result.

[0016] In some embodiments, calculating the local contribution estimate of the point cloud based on the distance information includes:

[0017] Based on the distance information, a kernel function is placed within the distance information of each point cloud. Using the distance information and the kernel function, the correlation between the current point cloud and other point clouds is estimated through the kernel function, thereby obtaining the local contribution estimate of the point cloud.

[0018] In some embodiments, the fitting process after multiple convolutional layers includes:

[0019] For point clouds whose local contribution estimates are greater than the first threshold, the weight values ​​are amplified by multilayer perceptron fitting to obtain the first point cloud weight score.

[0020] For point clouds whose local contribution estimates are less than the first threshold, their weight values ​​are reduced by multilayer perceptron fitting to obtain the second point cloud weight score.

[0021] The weight score of the first point cloud is greater than the weight score of the second point cloud.

[0022] In some embodiments, when calculating the Euclidean distance between each point cloud and other point clouds in the point cloud data to obtain the distance information corresponding to each point cloud, the process includes:

[0023] Define n point clouds d1, d2, ..., dn. Calculate the distance parameters of each point cloud to all other point clouds. Record the distance parameters of each point cloud to all other point clouds using a distance matrix.

[0024] Based on the distance matrix, the average Euclidean distance of each point cloud is calculated, and the reciprocal is taken to obtain the distance information corresponding to each point cloud.

[0025] In some embodiments, after obtaining the distance information corresponding to each point cloud, the information is input into a Gaussian kernel function. In the diagram, dis represents the distance information corresponding to each point cloud, b is the standard deviation, and e is the natural constant.

[0026] Gaussian kernel function After processing, the local contribution estimate is output and then fitted through four one-dimensional convolutional layers. In each convolutional layer, the local contribution estimate is used to extract features and then batch normalized to adjust it to a standard normal distribution with a mean of 0 and a variance of 1. Then, the ReLU activation function is applied to set the output of negative input values ​​to 0 and keep the output of positive input values ​​unchanged.

[0027] Finally, the local contribution estimates are reduced to one dimension, and after feature extraction, they are normalized using the sigmoid function to obtain the point cloud weight scores.

[0028] In some embodiments, after obtaining the point cloud weight score, the point cloud weight score is multiplied by the feature output information, and after convolutional coding, the data from different sampling channels are uniformly distributed to obtain the first feature suppression result.

[0029] In some embodiments, the first feature suppression result is passed through three one-dimensional convolutional layers for dimensionality reduction to obtain the second feature suppression result.

[0030] The result of the second feature suppression is passed through BatchNormal, ReLU and Dropout layers in sequence, and then dimensionality reduction is performed using a one-dimensional convolutional layer. The result is input into the Log Softmax function to obtain the probability score of the current frame point cloud belonging to each category. The loss value is obtained by using Negative Log-Likelihood Loss, and the model weight parameters are updated by backpropagation using the loss value.

[0031] In some embodiments, after dimensionality reduction using a one-dimensional convolutional layer, the number of channels and the number of categories in the resulting product are the same.

[0032] Secondly, the present invention provides a point cloud data processing system based on global outlier feature suppression, which applies the point cloud data processing method based on global outlier feature suppression as described above, including:

[0033] The data acquisition module is used to acquire point cloud data;

[0034] The data sampling module is used to perform convolutional coding feature extraction on the point cloud data using M point cloud sampling channels to obtain M feature output information;

[0035] The feature suppression module is used to calculate the Euclidean distance between each point cloud and other point clouds in the point cloud data, obtain the distance information of each point cloud, calculate the local contribution estimate of the point cloud based on the distance information, and obtain the point cloud weight score used to measure the outlier information of the point cloud after fitting processing through multiple convolutional layers.

[0036] The feature fusion module is used to fuse the output information of M features based on the point cloud weight scores to obtain the first feature suppression result.

[0037] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0038] Compared with the prior art, the present invention has at least the following beneficial effects:

[0039] Using the distance information between point clouds as the initial features of the point clouds, and after being fitted by the multi-layer convolutional layer of the input MLP, the point cloud weight score of the outlier degree of each point cloud is predicted by the Gaussian kernel function. Based on the point cloud weight score, feature fusion is performed so that the feature vector far from the center point approaches 0, while the feature vector close to the center point remains almost unchanged. This achieves the suppression of specific point cloud features based on global outliers, improves classification accuracy, and makes up for the lack of spatial sense.

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0041] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0042] Figure 1 This is a schematic diagram of the point cloud data processing flow of the PointNet++ model in the existing technology.

[0043] Figure 2 This is a flowchart illustrating the point cloud data processing method provided in this embodiment.

[0044] Figure 3 This is a schematic diagram of the feature suppression module in this embodiment.

[0045] Figure 4 This is a schematic diagram of the point cloud data processing system provided in this embodiment. Detailed Implementation

[0046] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0047] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0048] In the description of this invention, when a specific device is described as being located between a first device and a second device, an intermediary device may or may not be present between the specific device and the first or second device. When a specific device is described as being connected to other devices, the specific device may be directly connected to the other devices without an intermediary device, or it may not be directly connected to the other devices but may have an intermediary device.

[0049] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0050] The inventor discovered:

[0051] When using point cloud data for classification, two main challenges are generally encountered:

[0052] 1. Noise: Point cloud data is often affected by sensor measurement accuracy and noise. For example, limitations in sensor manufacturing processes and inaccurate measurements can lead to noise in point cloud data. This makes point cloud classification tasks more difficult.

[0053] 2. Imbalanced Data: In point cloud data, the number of samples in different categories may be imbalanced. For example, some categories may have fewer samples, while other categories may have more. This imbalanced data can negatively impact classification accuracy when training a classification model.

[0054] However, when performing point cloud classification, the point cloud density characteristics of a compact object and a sparse object are significantly different. For example, a car and a tree have very different point cloud density characteristics due to the unavoidable noise and data imbalance in point cloud data. Moreover, when classifying point cloud data, the more obvious or highlighting the outline of an object, the fewer outliers there are in the point cloud. Currently, there is no relevant existing technology to process point cloud data based on these characteristics.

[0055] In view of this, in order to solve the above-mentioned existing problems, firstly, referring to Figure 2 This invention provides a point cloud data processing method based on global outlier feature suppression, comprising:

[0056] The point cloud data is acquired, and convolutional coding is performed on the point cloud data using M point cloud sampling channels to extract M feature output information. The point cloud sampling channels are KNN sampling channels, and each channel can be set with different sampling attributes, such as cars, trees, pedestrians, etc.

[0057] Based on point cloud data, the Euclidean distance between each point cloud and other point clouds is calculated to obtain the distance information of each point cloud. Based on the distance information, the local contribution estimate of the point cloud is calculated. After fitting through multiple convolutional layers, the point cloud weight score used to measure the outlier information of the point cloud is obtained.

[0058] Based on the point cloud weight scores, the output information of M features is fused to obtain the first feature suppression result.

[0059] It should be noted that, optionally, M equals 3. Since point cloud data includes point clouds of various objects, both compact and sparse, point clouds on the same object will exhibit different senses of distance and space compared to point clouds on other objects. However, this sense of distance and space is difficult to perfectly represent with a single parameter. Therefore, in this embodiment, distance information between point clouds is used as the initial feature of the point cloud. This distance information is then processed by a Gaussian kernel function, and the result is input into multiple convolutional layers of an MLP for fitting. This is equivalent to using a Gaussian kernel function in conjunction with an MLP to fit a kernel function that conforms to the distribution of the point cloud data. This is because directly using the kernel density function to estimate the outlier information of the current point cloud within the entire point cloud is problematic due to the complex spatial distribution of point cloud data, which includes not only distance factors but also spatial factors. Since it is impossible to estimate the outlier degree of the current point cloud in the entire point cloud using a precise kernel density function expression, this embodiment uses a multilayer perceptron (MLP) to fit a Gaussian kernel function to obtain a kernel density function that conforms to the distribution of point cloud data in order to obtain global outlier information. Only point cloud data representing the overall spatial information is used as input, and the final output is a point cloud weight score used to measure the outlier information of the point cloud. The point cloud weight score represents the global outlier degree of different point clouds. Through the configuration of the point cloud weight scores, some point clouds are suppressed and some are preserved. Then, feature fusion is performed based on the point cloud weight scores, so that the feature vector far from the center point approaches 0, while the feature vector close to the center point remains almost unchanged. This achieves the suppression of specific point cloud features based on global outliers, improves classification accuracy, and makes up for the lack of spatial sense.

[0060] As one implementation method, when calculating the local contribution estimate of a point cloud based on distance information, the method includes:

[0061] Based on distance information, each point cloud is treated as a probabilistic point mass. A kernel function is placed around each point cloud; optionally, a Gaussian kernel function is placed within the distance information of each point cloud. The Gaussian kernel function is used to represent the local contribution of that point cloud. Therefore, using the distance information and the Gaussian kernel function, the correlation between the current point cloud and other point clouds is estimated through the kernel function, thus obtaining an estimate of the local contribution of the point cloud. Since a point cloud is essentially a point mass in space, its relationship with other point clouds can be estimated using a kernel density function.

[0062] As one implementation method, the fitting process after multiple convolutional layers includes:

[0063] Since the local contribution estimates of different point clouds vary significantly, and these different local contribution estimates are treated differently in the subsequent multilayer perceptron fitting process, the details are as follows:

[0064] For point clouds whose local contribution estimates are greater than the first threshold, the weight values ​​are amplified by multilayer perceptron fitting to obtain the first point cloud weight score.

[0065] For point clouds whose local contribution estimates are less than the first threshold, their weight values ​​are reduced by multilayer perceptron fitting to obtain the second point cloud weight score.

[0066] The weight score of the first cloud point is greater than the weight score of the second cloud point.

[0067] Among them, the first threshold is a preset parameter. The local contribution estimate obtained from different point clouds is compared with it. Based on different comparison results, different fitting measures are taken to achieve targeted amplification or reduction of weight values, so that the weight score of point clouds with local contribution estimates greater than the first threshold is greater than the weight score of point clouds with local contribution estimates less than the first threshold.

[0068] Furthermore, different thresholds can be set for different object categories. For example, there is a threshold A for cars and a threshold B for trees. Since the point cloud density features of cars are different from those of trees, cars have fewer outliers and trees have more outliers. Therefore, the comparison standards for cars and trees are different. Threshold A needs to be greater than threshold B in order to better suppress the weight of outliers and improve classification accuracy.

[0069] Example 1:

[0070] Combination Figure 3 In this embodiment, when calculating the Euclidean distance between each point cloud and other point clouds in the point cloud data to obtain the distance information corresponding to each point cloud, the process includes:

[0071] Define n point clouds d1, d2, ..., dn. Calculate the distance parameters of each point cloud to all other point clouds. Record the distance parameters of each point cloud to all other point clouds using a distance matrix.

[0072] Based on the distance matrix, the average Euclidean distance of each point cloud is calculated, and the reciprocal is taken to obtain the distance information corresponding to each point cloud.

[0073] After obtaining the distance information corresponding to each point cloud, it is input into the Gaussian kernel function. In the diagram, dis represents the distance information corresponding to each point cloud, b is the standard deviation, and e is the natural constant.

[0074] Gaussian kernel function After processing, the local contribution estimate is output and then fitted through four one-dimensional convolutional layers. In each convolutional layer, the local contribution estimate is used for feature extraction and batch normalization to adjust it to a standard normal distribution with a mean of 0 and a variance of 1. This helps to avoid the gradient vanishing or gradient exploding problems in deep networks, making the network easier to train and accelerating the training convergence speed. Then, the ReLU activation function is applied to set the output of negative input values ​​to 0 and keep the output of positive input values ​​unchanged, realizing non-linear mapping and achieving data sparsity.

[0075] Finally, the local contribution estimates are reduced to one dimension, and after feature extraction, they are normalized using the sigmoid function to obtain the point cloud weight scores.

[0076] Specifically, the first convolutional layer J is (B,8,S,N), the second convolutional layer K is (B,16,S,N), the third convolutional layer L is (B,8,S,N), and the fourth convolutional layer P is (B,1,S,N). The stride of the first convolutional layer J is 1, which expands dis from 1 dimension to 8 dimensions to extract features. After processing by the batch normalization (BN) layer and the ReLU activation function, the result is input into the second convolutional layer K, and so on, until the fourth convolutional layer P, which has a stride of 1, reduces dis from 8 dimensions to 1 dimension to extract features, and obtains the weight score of the current point cloud. Then, the point cloud weight score is normalized by the sigmoid function to obtain the final point cloud weight score.

[0077] After obtaining the point cloud weight score, the point cloud weight score is multiplied with the feature output information. After convolutional encoding, the data from different sampling channels are uniformly distributed to obtain the first feature suppression result.

[0078] The first feature suppression result is then passed through three one-dimensional convolutional layers for dimensionality reduction to obtain the second feature suppression result.

[0079] The result of the second feature suppression is passed through BatchNormal, ReLU and Dropout layers in sequence, and then dimensionality reduction is performed using a one-dimensional convolutional layer. The result is input into the Log Softmax function to obtain the probability score of the current frame point cloud belonging to each category. The loss value is obtained by using Negative Log-Likelihood Loss, and the model weight parameters are updated by backpropagation using the loss value.

[0080] It's worth noting that using the Log Softmax function can increase the discriminative power of the probability values ​​of the entire point cloud belonging to different categories, making it easier for the decision-making layer to determine the category of the entire point cloud. Using Negative Log-LikelihoodLoss to guide model training allows the model weights to fit the complex data distribution of the point cloud data.

[0081] Preferably, the number of points in the point cloud data is greater than the preset number of center points, so that M point cloud sampling channels can be used to perform convolutional coding feature extraction on the point cloud data; wherein, each point cloud sampling channel samples 16, 32, and 128 points in sequence.

[0082] Preferably, after the Dropout layer outputs the results, a one-dimensional convolutional layer is used for dimensionality reduction, resulting in a result with the same number of channels and categories.

[0083] Example 2:

[0084] On the ModelNet40 dataset, the original algorithm achieved a global class accuracy of 91.9% and an average class accuracy of 89.6%.

[0085] By introducing a global outlier feature suppression implementation based on the method in Example 1, point cloud data representing overall spatial information is used as input, and the final output is a point cloud weight score used to measure outlier information in the point cloud. Feature fusion is then performed based on the point cloud weight score. The first feature suppression result is then processed by three one-dimensional convolutional layers to obtain the second feature suppression result, which is then processed by BatchNormal, ReLU, and Dropout layers, and then processed by a one-dimensional convolutional layer. The result is input into the Log Softmax function to obtain the probability score of the current frame point cloud belonging to each category. The loss value is obtained using Negative Log-Likelihood Loss, and the model weight parameters are updated by backpropagation using the loss value. After training the model, based on the ModelNet40 dataset, the global category classification accuracy can be improved to 93.3%, and the average category accuracy can be improved to 91.5%.

[0086] The specific data comparison is shown in Table 1 below:

[0087] Table 1 Comparison of Classification Accuracy

[0088] Model Global category accuracy Average category accuracy PointNet++ 0.919 0.896 PointConv 0.923 0.890 PointWeb 0.923 0.894 This embodiment 0.933 0.915

[0089] Global class accuracy is calculated as the number of correctly predicted samples in the test set divided by the total number of samples in the test set. Average class accuracy is the sum of the classification accuracies for each class, followed by the average.

[0090] Reference Figure 4 Secondly, embodiments of the present invention provide a point cloud data processing system based on global outlier feature suppression, applying the point cloud data processing method based on global outlier feature suppression as described above, including:

[0091] The data acquisition module is used to acquire point cloud data;

[0092] The data sampling module is used to extract convolutional coding features from point cloud data using M point cloud sampling channels to obtain M feature output information.

[0093] The feature suppression module is used to calculate the Euclidean distance between each point cloud and other point clouds in the point cloud data, obtain the distance information of each point cloud, calculate the local contribution estimate of the point cloud based on the distance information, and obtain the point cloud weight score used to measure the outlier information of the point cloud after fitting through multiple convolutional layers.

[0094] The feature fusion module is used to fuse M feature output information based on point cloud weight scores to obtain the first feature suppression result. Specifically, after obtaining the point cloud weight scores, the point cloud weight scores are multiplied with the feature output information, and after convolutional coding, the data from different sampling channels are uniformly distributed to obtain the first feature suppression result.

[0095] The feature suppression module includes a distance matrix, a Gaussian kernel function, and an MLP (Multilayer Perceptron) with four one-dimensional convolutional layers. Record the distance parameters of each point cloud to other point clouds. Based on the distance matrix, calculate the average Euclidean distance of each point cloud. Take the reciprocal to obtain the distance information corresponding to each point cloud. Input the distance information into the Gaussian kernel function. In the middle, through the Gaussian kernel function After processing, the output local contribution estimates are sequentially fitted through four one-dimensional convolutional layers. In each convolutional layer, features are extracted from the local contribution estimates, followed by batch normalization (BN) layers and ReLU activation. After feature extraction, the features are normalized using the sigmoid function to obtain the point cloud weight scores. This process removes noise and preserves as much of the original point cloud features as possible. Simultaneously, it allows the model to remember the point cloud density characteristics of different objects, improving not only the accuracy of point cloud classification but also mitigating the impact of data imbalance on the model's prediction accuracy.

[0096] As one implementation method, this point cloud data processing system also includes:

[0097] The one-dimensional convolutional dimensionality reduction module is used to process the first feature suppression result through three one-dimensional convolutional layers to obtain the second feature suppression result.

[0098] The data processing module is used to process the second feature suppression result through BatchNormal, ReLU and Dropout layers in sequence, and then use a one-dimensional convolutional layer for dimensionality reduction. The result is then input into the Log Softmax function to obtain the probability score of the current frame point cloud belonging to each category. The loss value is obtained by using Negative Log-Likelihood Loss, and the loss value is used for backpropagation to update the model weight parameters.

[0099] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0100] Fourthly, embodiments of the present invention provide a computer-readable storage medium. This computer-readable storage medium may be a non-volatile computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the point cloud data processing method described above.

[0101] Those skilled in the art will readily understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the foregoing description. 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 invention.

[0102] In summary, compared with the prior art, the above embodiments provide a point cloud data processing method, system, and computer device based on global outlier feature suppression. Utilizing the distance information between point clouds as the initial features, and after Gaussian kernel calculation and fitting through multiple convolutional layers of an MLP, the outlier degree of each point cloud is predicted as a point cloud weight score. Feature fusion is then performed based on the point cloud weight scores, causing feature vectors far from the center point to approach 0, while feature vectors near the center point remain almost unchanged. This achieves suppression of specific point cloud features based on global outliers, improves classification accuracy, and compensates for the lack of spatial awareness.

[0103] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A point cloud data processing method based on global outlier feature suppression, characterized in that, include: Acquire point cloud data, and use M point cloud sampling channels to perform convolutional coding feature extraction on the point cloud data to obtain M feature output information; Calculate the Euclidean distance between each point cloud and other point clouds in the point cloud data to obtain the distance information for each point cloud. Based on the distance information, calculate the local contribution estimate of the point cloud. After fitting through multiple convolutional layers, obtain the point cloud weight score used to measure the outlier information of the point cloud. Based on the point cloud weight scores, the output information of M features is fused to obtain the first feature suppression result.

2. The method as described in claim 1, characterized in that, When calculating the local contribution estimate of the point cloud based on the distance information, the following steps are included: Based on the distance information, a kernel function is placed within the distance information of each point cloud. Using the distance information and the kernel function, the correlation between the current point cloud and other point clouds is estimated through the kernel function, thereby obtaining the local contribution estimate of the point cloud.

3. The method as described in claim 2, characterized in that, During the fitting process through multiple convolutional layers, the following is included: For point clouds whose local contribution estimates are greater than the first threshold, the weight values ​​are amplified by multilayer perceptron fitting to obtain the first point cloud weight score. For point clouds whose local contribution estimates are less than the first threshold, their weight values ​​are reduced by multilayer perceptron fitting to obtain the second point cloud weight score. The weight score of the first point cloud is greater than the weight score of the second point cloud.

4. The method as described in claim 3, characterized in that, When calculating the Euclidean distance between each point cloud and other point clouds in point cloud data to obtain the distance information for each point cloud, the following steps are taken: Define d1, d2, ..., d n Given n point clouds, calculate the distance parameters between each point cloud and all other point clouds. Record these distance parameters using a distance matrix. Based on the distance matrix, the average Euclidean distance of each point cloud is calculated, and the reciprocal is taken to obtain the distance information corresponding to each point cloud.

5. The method as described in claim 4, characterized in that, After obtaining the distance information corresponding to each point cloud, it is input into the Gaussian kernel function. In the diagram, dis represents the distance information corresponding to each point cloud, b is the standard deviation, and e is the natural constant. Gaussian kernel function After processing, the local contribution estimate is output and then fitted through four one-dimensional convolutional layers. In each convolutional layer, the local contribution estimate is used to extract features and then batch normalized to adjust it to a standard normal distribution with a mean of 0 and a variance of 1. Then, the ReLU activation function is applied to set the output of negative input values ​​to 0 and keep the output of positive input values ​​unchanged. Finally, the local contribution estimates are reduced to one dimension, and after feature extraction, they are normalized using the sigmoid function to obtain the point cloud weight scores.

6. The method as described in claim 5, characterized in that, After obtaining the point cloud weight score, the point cloud weight score is multiplied with the feature output information. After convolutional encoding, the data from different sampling channels are uniformly distributed to obtain the first feature suppression result.

7. The method according to any one of claims 1 to 6, characterized in that, The first feature suppression result is passed through three layers of one-dimensional convolutional layers for dimensionality reduction to obtain the second feature suppression result. The result of the second feature suppression is passed through BatchNormal, ReLU and Dropout layers in sequence, and then dimensionality reduction is performed using a one-dimensional convolutional layer. The result is input into the Log Softmax function to obtain the probability score of the current frame point cloud belonging to each category. The loss value is obtained by using Negative Log-Likelihood Loss, and the model weight parameters are updated by backpropagation using the loss value.

8. The method as described in claim 7, characterized in that, After dimensionality reduction using a one-dimensional convolutional layer, the number of channels and the number of categories in the resulting product are the same.

9. A point cloud data processing system based on global outlier feature suppression, employing the point cloud data processing method based on global outlier feature suppression as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire point cloud data; The data sampling module is used to perform convolutional coding feature extraction on the point cloud data using M point cloud sampling channels to obtain M feature output information; The feature suppression module is used to calculate the Euclidean distance between each point cloud and other point clouds in the point cloud data, obtain the distance information of each point cloud, calculate the local contribution estimate of the point cloud based on the distance information, and obtain the point cloud weight score used to measure the outlier information of the point cloud after fitting processing through multiple convolutional layers. The feature fusion module is used to fuse the output information of M features based on the point cloud weight scores to obtain the first feature suppression result.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images