A pseudo-supervoxel point cloud preprocessing method for point cloud target detection

Abstract

This invention discloses a pseudo-supervoxel point cloud preprocessing method for point cloud target detection, which solves the problems of noise handling, loss of local details, and time-consuming supervoxel segmentation in point cloud data preprocessing. The method achieves the following: obtaining a pseudo-image feature map, acquiring a category segmentation mask, constructing a local mean pseudo-image that preserves spatial details, and generating a pseudo-image feature map after pseudo-supervoxel preprocessing. It uses a superpixel algorithm to extract supervoxel semantic information, achieving better speed and computational complexity than the supervoxel algorithm. By utilizing the correspondence between the category segmentation mask and the pseudo-image feature map, local meanization is performed to reduce the impact of random noise. Furthermore, a cross-channel attention vector and a supervoxel factor matrix are constructed to adjust the feature weight distribution, enhancing the semantic representation of spatial information and improving detection accuracy. This method can be applied to point cloud target detection feature maps of different resolutions without fundamentally modifying the algorithm architecture, making it suitable for scenarios such as autonomous driving and robot navigation.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and mainly relates to point cloud data processing. Specifically, it is a pseudo-supervoxel point cloud preprocessing method for point cloud target detection, which can be used to enhance the accuracy of point cloud target detection and can be applied to application scenarios such as autonomous driving and robot navigation. Background Technology

[0002] Point cloud data, as an important data form for 3D reconstruction and environmental perception, is widely used in fields such as autonomous driving, UAV mapping, and virtual reality. However, point cloud data often suffers from large data volume, high noise levels, and uneven density, posing challenges to subsequent processing and analysis in point cloud target detection tasks. Voxelization is a commonly used point cloud data preprocessing method: by dividing the point cloud into voxels of equal volume for subsequent processing, the data volume can be reduced and the processing speed can be accelerated. However, voxelization is a simple and crude method that cannot effectively remove noisy data. In areas rich in detail, voxelization may lead to blurring or loss of details, thus affecting subsequent analysis and decision-making. When using point cloud downsampling methods for point cloud data preprocessing, the data volume and computational complexity are usually reduced by decreasing the number of points in the point cloud. However, this reduction in the number of points may also lead to the loss of detailed information, affecting the accuracy of subsequent processing. Outlier filtering is another commonly used point cloud data preprocessing method: using statistical methods or neighborhood-based feature analysis, outliers in the point cloud data are removed. Although this point cloud preprocessing method can effectively remove noise, it has high computational complexity, may lead to data loss, and has poor performance when processing large-scale point cloud data.

[0003] Superpixels are a form of image segmentation that divides an image into several irregular regions with certain semantic information. Pixels within these regions are relatively consistent in features such as color, texture, and brightness. Pixels located close to each other and sharing certain common characteristics are considered similar, and these similarities are used to group pixels together to form superpixels. Regions, on the other hand, exhibit significant differences. In image processing tasks, this superpixel method "aggregates" independent, scattered small pixels into larger basic units, obtaining more representative image information. This not only simplifies subsequent image processing tasks but also preserves image edges and other important features, significantly reducing the workload and difficulty of subsequent image processing tasks. It is commonly used in tasks such as image simplification, image compression, image segmentation, and feature extraction.

[0004] Supervoxels, similar to superpixels, are a concept used in 3D point cloud data. They are an extension of the voxel concept, used to achieve finer-grained segmentation in 3D space. A voxel is the smallest unit of volume in 3D space, similar to a pixel in a 2D image. Supervoxels are essentially a voxel clustering process, grouping voxels with similar characteristics such as color, texture, shape, and normal direction into a larger region, i.e., a supervoxel. Performing supervoxels helps reduce the computational load of subsequent operations and can serve as a preprocessing stage for many tasks in 3D voxel space. The concept of supervoxels is mainly used for point cloud data, especially in fields such as 3D reconstruction, medical image processing, and robot vision.

[0005] Patent document CN2024102888711 discloses "A point cloud segmentation method, apparatus, and device based on voxel and point set fusion." This method preprocesses the original point cloud data to obtain supervoxel data; it then performs clustering segmentation on the supervoxel data to obtain clustered supervoxel data, further obtaining supervoxel features; finally, it fuses the supervoxel features with the point cloud features to obtain fused features, and then performs semantic classification segmentation on each point in the original point cloud data based on the fused features to obtain the point cloud segmentation result. The shortcomings of this method are that, since point cloud data typically contains a large number of points, directly performing supervoxel processing on the point cloud data during preprocessing results in high computational complexity and is very time-consuming; furthermore, the distribution of point cloud data may be uneven, and direct supervoxel processing may lead to excessively large supervoxels in some areas, lacking sufficient detail and affecting the accuracy of the processing results.

[0006] In summary, existing point cloud data preprocessing techniques and supervoxel segmentation techniques have the following main shortcomings: Most point cloud data preprocessing methods do not perform noise reduction on large-scale point cloud data. Furthermore, since point clouds are typically sparse, simplifying or compressing point cloud data often involves reducing the number of points, which may lead to the loss of key points, resulting in the loss of local details and features, affecting subsequent processing and analysis. Supervoxel segmentation algorithms involve complex mathematical and geometric calculations, such as distance metrics, feature extraction, and clustering. These calculations become even more complex and time-consuming in high-dimensional spaces. Therefore, directly performing supervoxel segmentation on voxels is extremely time-consuming and cannot be applied to real-time scenarios with extremely high real-time requirements. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of the prior art by proposing a pseudo-supervoxel point cloud preprocessing method for point cloud target detection that preserves more spatial detail features and reduces the time consumption of supervoxels.

[0008] This invention is a pseudo-supervoxel point cloud preprocessing method for point cloud target detection. Its key feature is that it utilizes the correspondence between the category segmentation mask generated by the superpixel segmentation algorithm and the pseudo-image feature map to perform local averaging to generate a local mean pseudo-image. The method then learns and fuses the pseudo-image and the local mean pseudo-image, adjusting the weight distribution of features in the pseudo-image feature map. The method includes the following steps:

[0009] (1) Obtaining the pseudo-image feature map: Obtain the original point cloud to be preprocessed, perform voxelization, divide the point cloud space into regular columnar voxels in the xy direction, use a simplified PointNet network and max pooling to encode each columnar voxel to obtain voxel features, arrange the voxel features according to their positions in three-dimensional space to obtain the pseudo-image feature map F. pi ;

[0010] (2) Obtain the category segmentation mask: For the pseudo-image feature map F pi Using a pseudo-supervoxel segmentation module, the Simple Linear Iterative Clustering (SLIC) algorithm is employed for pseudo-supervoxel segmentation, focusing only on the pseudo-image feature map F. pi The first three channels are used for pseudo-supervoxel segmentation, and the number of superpixel labels is determined according to the desired number of superpixel regions to generate a class segmentation mask T. mask Category segmentation mask T mask The value of each element in the table represents the superpixel label to which the corresponding pixel belongs;

[0011] (3) Constructing a local mean pseudo-image that preserves spatial details: For the pseudo-image feature map F pi and category segmentation mask T mask Using a local meanization module, the feature map F of the pseudo-image is targeted. pi The mean value of the feature values ​​of pixels belonging to the same category is calculated and used as a local mean pseudo-image F. lm The feature value at the corresponding pixel location, i.e., using the category segmentation mask T mask With pseudo-image feature map F pi The correspondence is used to construct a local mean pseudo-image F that preserves spatial detail features. lm ;

[0012] (4) Generate pseudo-image feature maps after pseudo-supervoxel point cloud preprocessing: Use the factor extraction and aggregation module to process the pseudo-image feature maps F pi and the local mean pseudo-image F lm Learn and integrate, adjust pseudo-image feature map F pi The weight distribution of features in the middle is analyzed, and semantic information in the supervoxel space is fused to generate a pseudo-image feature map F' after preprocessing of the pseudo-supervoxel point cloud. piThis completes the pseudo-supervoxel point cloud preprocessing task for point cloud target detection in the original point cloud to be preprocessed.

[0013] This invention addresses the technical problems that most point cloud data preprocessing methods do not perform noise reduction on point cloud data, and that simplifying or compressing point cloud data can easily lead to the loss of local details and features, as well as the time-consuming nature of directly performing super-voxel segmentation on voxels.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] To preserve more spatial detail features of the original point cloud and improve target detection accuracy, this invention designs a pseudo-supervoxel segmentation module. This module uses the Simple Linear Iterative Clustering (SLIC) algorithm to group regions with similar features together, preserving the main structure and edge information of the original point cloud. A factor extraction and aggregation module is also designed to aggregate the two pseudo-image feature maps, generating a supervoxel factor matrix. This matrix represents the weight distribution of each voxel in the input pseudo-image after pseudo-supervoxel segmentation. The input pseudo-image is multiplied by the supervoxel factor matrix to generate a pseudo-image feature map after pseudo-supervoxel point cloud preprocessing. This pseudo-image feature map, after pseudo-supervoxel point cloud preprocessing, integrates supervoxel semantic spatial information, preserving the spatial detail features of the original point cloud. This helps the target detection model to more accurately and effectively identify targets, improving target detection accuracy.

[0016] Less time-consuming than supervoxel segmentation: This invention designs a pseudo-supervoxel segmentation module, which uses the Simple Linear Iterative Clustering (SLIC) algorithm to perform pseudo-supervoxel segmentation on the pseudo-image feature map generated from the original point cloud voxels. This achieves the effect of supervoxel processing of voxels and alleviates the time-consuming situation when directly performing supervoxel processing on voxels.

[0017] To mitigate random noise and reduce its impact, this invention designs a pseudo-supervoxel segmentation module that uses the Simple Linear Iterative Clustering (SLIC) algorithm to aggregate regions with similar features, smoothing out random variations caused by noise and effectively reducing its impact. A local meanization module is also designed to average the feature values ​​of regions with similar features, further smoothing the feature map and weakening random noise.

[0018] Expanding the scope of application: This invention can be directly applied to feature maps of different resolutions in any point cloud object detection method without fundamentally modifying the algorithm architecture. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the present invention;

[0020] Figure 2This is a schematic diagram illustrating the effect of the pseudo-superpixel segmentation module used in this invention for single-channel pseudo-image feature maps;

[0021] Figure 3 This is a flowchart of the local averaging module of the present invention;

[0022] Figure 4 This is a partial structural diagram of the simulation experiment of this invention;

[0023] Figure 5 The change in average detection accuracy after inserting the present invention into feature maps of different resolutions is shown.

[0024] The technical solution and effects of the present invention will be described in detail below with reference to the accompanying drawings. Detailed Implementation

[0025] Example 1:

[0026] In common point cloud data preprocessing methods, voxelization, which divides the point cloud into voxels of equal volume for subsequent processing, is a frequently used method. However, voxelization cannot effectively remove noise data, and in areas rich in detail, it may lead to blurring or loss of details, thus affecting the accuracy of subsequent processing. Outlier filtering is another common point cloud data preprocessing method: statistical methods or neighborhood-based feature analysis are used to remove outliers from the point cloud data. Although this point cloud preprocessing method can effectively remove noise, it has high computational complexity, may lead to data loss, and performs poorly when processing large-scale point cloud data. Another method is super-voxel segmentation of the original point cloud voxels during point cloud data preprocessing, but super-voxel segmentation algorithms involve complex mathematical and geometric calculations, which are very time-consuming and complex. To address these problems, this invention starts with the correlation between two-dimensional pseudo-image feature maps and voxels, and through exploratory experiments and research, proposes a pseudo-super-voxel point cloud preprocessing method for point cloud target detection.

[0027] This invention is a pseudo-supervoxel point cloud preprocessing method for point cloud target detection. It utilizes the correspondence between the category segmentation mask generated by the superpixel segmentation algorithm and the pseudo-image feature map to perform local averaging to generate a local mean pseudo-image. The pseudo-image and the local mean pseudo-image are then learned and fused, and the weight distribution of features in the pseudo-image feature map is adjusted. (See [link to related documentation]). Figure 1 , Figure 1 This is a flowchart of the present invention, which includes the following steps:

[0028] (1) Obtaining the pseudo-image feature map: Obtain the original point cloud to be preprocessed, perform voxelization, encoding, and feature extraction to obtain the pseudo-image feature map. Specifically, obtain the original point cloud data to be preprocessed, filter the point cloud data according to the preset three-dimensional boundary, divide the point cloud space into uniform regular columnar voxels in the xy direction according to the preset voxel grid size, and assign each point to the corresponding columnar voxel according to the x and y coordinates of the point. Set a sampling threshold D, randomly sample the points in columnar voxels with more points than the threshold, and zero-fill the points in columnar voxels with less points than the threshold; within each columnar voxel, calculate the coordinates of the center point and the relative coordinates of each point relative to the center point, generate the feature vector of the point according to the spatial coordinates, reflection intensity, relative coordinates, etc., and encode each columnar voxel using a simplified PointNet network and max pooling to obtain voxel features. Arrange the features of each columnar voxel according to its position in three-dimensional space to obtain the pseudo-image feature map F. pi .

[0029] (2) Obtaining the category segmentation mask: The pseudo-supervoxel segmentation module is used to perform pseudo-supervoxel segmentation on the pseudo-image feature map. The pseudo-supervoxel segmentation module processes the first three channels of the pseudo-image feature map using the Simple Linear Iterative Clustering (SLIC) algorithm to obtain the category segmentation mask. Specifically, for the pseudo-image feature map F pi The pseudo-supervoxel segmentation module employs the Simple Linear Iterative Clustering (SLIC) algorithm for pseudo-supervoxel segmentation. The SLIC algorithm mainly includes six steps: setting seed points, reselecting seed points, label assignment, distance measurement, iterative optimization, and enhancing connectivity. First, based on the number of pixels N in the input image and the pre-defined number of superpixels K, the input image is uniformly divided into K equal-sized regions. The center of each region is used as the seed point, i.e., the superpixel center. The superpixel size is calculated by dividing the number of pixels by the number of superpixels, i.e., N / K. The distance between adjacent seed points is approximately... To avoid seed points falling on locations with large gradients and affecting subsequent clustering results, the gradient values ​​of all pixels within a certain neighborhood of the seed point are calculated. The pixel with the smallest gradient within this neighborhood is then reselected as the seed point. To accelerate algorithm convergence, within a 2S*2S neighborhood, each seed point assigns a category label to each pixel in its neighborhood, indicating which cluster center a pixel belongs to. The distance between the searched pixels and the seed point, including the color distance d, is calculated according to the formula. c and spatial distance d s The distance calculation method is as follows:

[0030]

[0031] Where, d c Represents color distance, ds N represents spatial distance. s It is the maximum spatial distance within a class, defined as Applies to each cluster. Maximum color distance N c The distance metric varies depending on both the image and the clustering, so a fixed constant m is used instead. The value of m ranges from 1 to 40, but is typically set to 10. The final distance metric D' is as follows:

[0032]

[0033] Since each pixel is searched by multiple seed points, each pixel's distance from its surrounding seed points will have a minimum value. The seed point corresponding to this minimum value is taken as the cluster center of that pixel. By repeatedly setting seed points, reselecting seed points, assigning labels, and measuring distances, the error will tend to balance. This balance means that the cluster center of each pixel will no longer change. In the previous step, some new problems may arise, such as: a large difference between the actual and expected superpixel size, unacceptable multi-connectivity between pixels, or pixels that should belong to a certain superpixel being split into multiple discontinuous small superpixels. To avoid these problems, the solution provided by this invention is: to create a new label table where all elements are -1, and to assign problematic superpixels to neighboring superpixels in a "Z" shaped order, then assign new labels to the superpixels until all pixels are covered. See also Figure 2 , Figure 2 This is a schematic diagram illustrating the effect of the pseudo-superpixel segmentation module used in this invention for single-channel pseudo-image feature maps. The pseudo-image feature map F is taken. pi The first three channels are used as input images and fed into the Simple Linear Iterative Clustering (SLIC) algorithm. The number of superpixel labels is determined according to the desired number of superpixel regions, and a class segmentation mask T is generated. mask This completes pseudo-supervoxel segmentation. Category segmentation mask T mask The value of each element in the image represents the superpixel label to which the corresponding pixel belongs. (Pseudo-image feature map F) pi The spatial relationships between the pixels are thus established.

[0034] (3) Construct a local mean averaging module to obtain a local mean pseudo-image that preserves spatial detail features: Construct a local mean averaging module to extract and utilize the correspondence between the category segmentation mask and the pseudo-image feature map, and average the feature values ​​of pixels belonging to the same category to obtain a local mean pseudo-image that preserves spatial detail features. In order to utilize the pseudo-image feature map F piTo address the spatial relationships between pixels, the local mean averaging module constructed in this invention is sequentially connected to a receiving section, a local mean calculation and application section, and an output generation section. The receiving section takes a pseudo-image feature map and a category segmentation mask as input, generates a zero tensor `seg_img` with the same shape as the pseudo-image feature map (i.e., all values ​​in the tensor are 0), and outputs the pseudo-image feature map, the category segmentation mask, and the zero tensor `seg_img` for storing the results. The local mean calculation and application section takes a pseudo-image feature map, a category segmentation mask, and the zero tensor as input, iterates through each segmentation category, calculates the local mean of the corresponding region in the pseudo-image feature map, applies these means to the corresponding positions in the zero tensor `seg_img`, and outputs the updated tensor `seg_img` after traversing all categories. The output generation section takes the tensor `seg_img` as input, integrates the data, constructs and outputs a local mean pseudo-image. The local mean averaging module first performs a local aggregation operation on the original pseudo-image and the category segmentation mask, specifically on the pseudo-image feature map F... pi and category segmentation mask T mask Using the local meanization module, starting from category label 1, for category segmentation mask T mask Perform line-by-line scanning and record the category segmentation mask T. mask The position of the pixel with category label 1 is determined by traversing the category segmentation mask T. mask All category labels in the middle. Then, for the pseudo-image feature map F pi The mean value of the feature values ​​of pixels belonging to the same category is calculated and used as a local mean pseudo-image F. lm The feature values ​​at the corresponding pixel locations, i.e., the local homogenization module uses the category segmentation mask T mask With pseudo-image feature map F pi The correspondence is used to obtain a local mean pseudo-image F that preserves spatial detail features. lm .

[0035] (4) Design a factor extraction and aggregation module to generate a pseudo-image feature map after pseudo-supervoxel point cloud preprocessing: In order not to lose the edge information of the point cloud, it is necessary to further explore the relationship between the mean and the original value, that is, to explore the pseudo-image feature map F. pi and the local mean pseudo-image F lm The relationship between them. This invention designs a factor extraction and aggregation module, which is sequentially connected to an attention mechanism part, a learning weight part, and a weight adjustment part: Attention mechanism part: The input is a pseudo-image feature map and a local mean pseudo-image. The two input images are concatenated and global average pooling is performed. The output is a cross-feature expression attention vector I; Learning weight part: The input is the cross-feature expression attention vector I and the learning vector I. The output is a super-voxel factor matrix W, which is used to represent the weight distribution of each voxel in the pseudo-image feature map; Weight adjustment part: The input is the pseudo-image feature map Fpi The supervoxel factor matrix W is multiplied by the two inputs to obtain the output, which is the pseudo-image feature map after preprocessing the pseudo-supervoxel point cloud. The factor extraction and aggregation module then processes the pseudo-image feature map F. pi and the local mean pseudo-image F lm The algorithm learns and fuses the data, collecting expressive statistics from the two inputs, and uses this information to adjust the pseudo-image feature map F. pi The weight distribution of features in the middle is analyzed, and semantic information in the supervoxel space is fused to generate a pseudo-image feature map F' after preprocessing of the pseudo-supervoxel point cloud. pi This completes the pseudo-supervoxel point cloud preprocessing task for point cloud target detection in the original point cloud to be preprocessed.

[0036] This invention is a comprehensive technical solution addressing the technical problems of time-consuming and computationally complex super-voxel segmentation of point cloud voxels, the lack of noise handling in most point cloud data preprocessing methods, and the easy loss of local details and features during point cloud data preprocessing. The technical approach of this invention is as follows: A pseudo-super-voxel segmentation module is used to process the pseudo-image feature map obtained from voxels; that is, a two-dimensional image superpixel segmentation algorithm is used instead of a three-dimensional super-voxel segmentation algorithm to alleviate the time-consuming nature of direct super-voxel processing. Subsequently, the processed category segmentation mask is processed and a local mean-smoothing module is constructed to smooth the feature map, thus addressing the lack of noise handling. Finally, a factor extraction and aggregation module is designed to fuse the processed local mean pseudo-image and the pseudo-image feature map obtained from voxels, enhancing the spatial information representation capability of the pseudo-image feature map while smoothing random noise in the pseudo-image feature map, thereby solving the problems of easily losing detailed features and ineffective noise handling in point cloud data processing. Initially, this invention aimed to establish relationships between voxels after dividing the point cloud data to be preprocessed into voxels, and to integrate the semantic information of super-voxels into the original pseudo-image, thereby enhancing the semantic expressive power of the original pseudo-image. Since the original pseudo-image is a two-dimensional image obtained by feature extraction from point cloud voxels, using a superpixel algorithm on the feature map of the original pseudo-image can achieve a clustering effect similar to using a supervoxel algorithm on voxels, extracting the semantic information of supervoxels to enhance the spatial semantic expressive power of the original pseudo-image. Furthermore, this method is faster and less computationally complex than the supervoxel algorithm. Based on this, to handle noise in the point cloud data to be preprocessed, a mean calculation strategy is used to smooth random noise and reduce random variations caused by noise. Finally, based on design adjustments and experimental analysis, the technical solution of this invention was completed.

[0037] This invention achieves a super-voxel segmentation effect equivalent to that at the voxel feature level by using a superpixel segmentation algorithm. While preserving the original spatial information, it suppresses noise signals and improves the time consumption of the super-voxel method. Ultimately, it enhances the semantic information expression capability of the acquired target features, obtains more representative target features, and achieves more accurate target detection results.

[0038] Example 2:

[0039] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Example 1. The method involves constructing a local meanization module in step (3) of this invention to obtain a local mean pseudo-image that preserves spatial detail features. See [link to relevant documentation]. Figure 3 , Figure 3 This is a flowchart of the local averaging module of the present invention, including:

[0040] 3.1) Scanning the category segmentation mask: For the category segmentation mask T mask Starting with category 1, segment the category using the mask T. mask The elements in the array are scanned row by row.

[0041] 3.2) Calculate the mean for regions of category 1: This is done while scanning the category segmentation mask T. mask During the process, the positions of pixels of category 1 are recorded, and pseudo-image feature maps F are extracted. pi Find the feature values ​​at the corresponding positions in the graph and calculate the mean of the extracted feature values.

[0042] 3.3) For the local mean pseudo-image F lm Assigning values ​​to pixels belonging to category 1: Based on the positions of pixels belonging to category 1 and the calculated mean, construct a local mean pseudo-image F. lm The mean obtained is used as the local mean pseudo-image F lm The feature value at the corresponding position in the middle.

[0043] 3.4) Construct a local mean pseudo-image F that preserves spatial detail features. lm Compare the currently scanned category with the set maximum value for each category to determine whether to traverse the category segmentation mask T. mask If the current category value is less than the set maximum category value, then all categories have not been traversed. In this case, the category value is incremented by one, and steps (3.1)-(3.3) are repeated according to the new category value until the category segmentation mask T is traversed. mask All categories, i.e., the current category value equals the set maximum category value; through statistical category segmentation mask T. maskThe locations of each category are marked, clarifying the number and location of voxels covered by each supervoxel after pseudo-supervoxel segmentation. The mean features corresponding to all categories are calculated, i.e., local mean calculation is performed. Local mean calculation achieves a certain degree of noise reduction within local regions and effectively represents the spatial meaning of supervoxels. The mean features are then categorized by category in the pseudo-image feature map F. pi The arrangement of positions on the image ultimately constructs a local mean pseudo-image F that preserves spatial detail features. lm .

[0044] This invention constructs a local meanization module to optimize the category segmentation mask T. mask Perform a scan and statistically analyze the category segmentation mask T. mask By identifying the position of each category label in the pseudo-supervoxel segmentation, we obtain the positions of each category region and the positions of the pixels contained within each category region, i.e., the pseudo-image feature map F. pi The size of each supervoxel and the position of the pixels contained in each supervoxel are determined; noise in the same category region is smoothed by calculating the feature mean of the same category region to reduce the random influence of noise points; the feature mean of each category is then applied to the pseudo-image feature map F. pi Arrange the corresponding positions on the graph to construct a local mean pseudo-image F. lm To further extract the pseudo-image feature map F pi The spatial information contained within effectively characterizes the spatial significance of supervoxels, enhancing the ability to express spatial detail features. It solves most of the technical problems of detail loss and insufficient noise processing in point cloud preprocessing.

[0045] Example 3:

[0046] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Examples 1-2, except that the record category segmentation mask T described in step 3.2) of this invention... mask The location of category 1 is identified, and the pseudo-image feature map F corresponding to category 1 is extracted based on the recorded location. pi The values ​​of the middle pixels are summed and the mean is calculated, including:

[0047] 3.2.1) Generate a Boolean tensor mask, indicating whether the corresponding element belongs to the specified category 1: for the category segmentation mask T mask For each element in the matrix, if the value of the element is equal to 1, the result is True; otherwise, it is False. This applies to the category segmentation mask T. mask The elements in the array are scanned row by row to ensure that the scan category segmentation mask T is used. mask Each location is visited no more than once. Each time a location is visited, its category value is set to 0. After the scan is complete, a segmentation mask T corresponding to the category is generated. maskA Boolean tensor mask of the same shape, where True indicates that the pixel at that position belongs to category 1, and False indicates that it does not belong to category 1.

[0048] 3.2.2) Extracting the pseudo-image F using the root Boolean tensor mask pi All pixels of category 1 in the pseudo-image F: According to the Boolean tensor mask, if a pixel value is 1 at a certain position in the Boolean tensor mask, then from the pseudo-image F pi Extract the feature value of the pixel corresponding to the given position; if the pixel value at a certain position in the Boolean tensor mask is 0, then do not modify the pseudo-image feature map F. pi The feature value of the corresponding pixel at that position is used for operation.

[0049] 3.2.3) Calculate the average value of the feature values ​​extracted from the pixels.

[0050] This invention uses only True and False to represent whether each pixel belongs to category 1, and generates a corresponding Boolean tensor mask. The Boolean tensor mask is simple and easy to understand, has a fast calculation speed, and can accurately and quickly locate the pseudo-image feature map F. pi Pixels of category 1 in the middle; by calculating the pseudo-image feature map F pi The mean feature value of pixels of category 1 is used to refine the category segmentation results and enhance the pseudo-image feature map F. pi The consistency of supervoxel semantic space features in regions of category 1 further improves the stability and accuracy of supervoxel semantic space detail feature representation. This solves the technical problem of spatial detail loss during most point cloud preprocessing.

[0051] Example 4:

[0052] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Examples 1-3: The design factor extraction and aggregation module described in step (4) of this invention generates a pseudo-image feature map after pseudo-supervoxel point cloud preprocessing, including:

[0053] 4.1) Obtain the pseudo-image F that can be collected. pi and the local mean pseudo-image F lmThe cross-channel expression attention vector I of expressive statistical information: To avoid misleading information propagation during subsequent re-expression of pseudo-image features and supervoxel semantic space, i.e., to prevent noise signals from confusing with effective signals and reducing detection accuracy, it is necessary to filter noise signals in local spatial regions. This invention designs a factor extraction and aggregation module, which first completes this task using high-confidence activation technology in point cloud streams. "High-confidence activation technology in point cloud streams" mainly involves identifying and activating high-confidence points or regions in point cloud data streams to ensure the high accuracy and reliability of the extracted information. To achieve this, global spatial information of the two feature expressions is first embedded and compressed to obtain a cross-feature expression attention vector. Based on the pseudo-image feature map F... pi and the local mean pseudo-image F lm Generate cross-channel attention vector I:

[0054] I = F gp (F pi ||F lm )

[0055] Where || denotes concatenating the two pseudo-images along the channel dimension, F gp This indicates a global average pooling operation.

[0056] 4.2) Generate a pseudo-image feature map F. pi The supervoxel factor matrix W, representing the weight distribution of each voxel after pseudo-supervoxel segmentation, is obtained by learning the supervoxel factor matrix W based on the cross-channel attention vector I.

[0057] W=σ(F mlp (I))

[0058] Among them, F mlp Let σ represent an MLP network, and σ represent the ReLU activation function.

[0059] 4.3) Adjust the pseudo-image feature map F according to the supervoxel factor matrix W. pi The feature weight distribution of each element is used to generate a pseudo-image feature map F' after preprocessing the pseudo-supervoxel point cloud. pi : Pseudo-image feature map F pi Multiplying with the supervoxel factor matrix W, the semantic information of the supervoxel space is fused, and noise signals in the point cloud stream are effectively suppressed to generate the pseudo-image feature map F' after the pseudo-supervoxel point cloud preprocessing. pi :

[0060] F' pi =W*F pi

[0061] Among them, the pseudo-image feature map F' after pseudo-super-voxel point cloud preprocessing piWith the pseudo-image feature map F pi They have the same shape.

[0062] This invention designs a factor extraction and aggregation module, which first extracts the pseudo-image feature map F' pi With the local mean pseudo-image F lm By concatenating along the channel dimension and performing global average pooling to generate channel-representation attention vectors I, the pseudo-image feature map F is effectively processed. pi With the local mean pseudo-image F lm This study provides a cross-modal global description of these two different feature representations, collecting expressive statistics of the two feature maps. Furthermore, it obtains the pseudo-image feature map F by learning the channel expression attention vector I through an MLP network. pi With the local mean pseudo-image F lm The image contains rich spatial and semantic information; the ReLU activation function is used to transform the learned information into a feature weight distribution, generating a supervoxel factor matrix W; finally, the pseudo-image feature map F is processed. pi Multiplying with the supervoxel factor matrix W achieves the processing of the pseudo-image feature map F. pi Adjusting the feature weight distribution in the middle makes the pseudo-image feature map F' generated after preprocessing the pseudo-supervoxel point cloud more accurate. pi It integrates supervoxel semantic space information without losing edge information, while smoothing random noise and reducing noise interference. It solves the technical problems of long time consumption and high computational complexity when directly using supervoxel segmentation algorithms to obtain supervoxel semantic information in point cloud preprocessing, as well as the technical problems of easy loss of spatial detail features and insufficient noise processing in most point cloud preprocessing processes.

[0063] Example 5:

[0064] A pseudo-supervoxel point cloud preprocessing method for point cloud object detection is similar to Examples 1-4: a plug-and-play pseudo-supervoxel point cloud preprocessing method that can be directly applied to feature maps of different resolutions in any point cloud object detection method without fundamentally modifying the algorithm architecture, including:

[0065] a) In the PointPillars backbone network, a pseudo-supervoxel point cloud preprocessing method is used to process feature maps of different resolutions, and then the original upsampling operation in the PointPillars backbone network is continued.

[0066] b) In the feature learning network of VoxelNet and SECOND algorithms, a pseudo-supervoxel point cloud preprocessing method is used for the 4D tensor obtained by the feature learning network, and then the established operations of VoxelNet and SECOND algorithms are continued.

[0067] This invention employs a modular design, allowing for adjustments and optimizations to modules as needed. This enables effective processing of feature maps at different resolutions, enhancing feature representation capabilities, better preserving and extracting key features, thereby improving the accuracy and reliability of object detection. Furthermore, it can be directly and easily integrated into various existing systems, reducing implementation complexity and cost. Simultaneously, it is applicable to multiple mainstream point cloud object detection algorithms, including PointPillars, VoxelNet, and SECOND. Regardless of the algorithm used, performance can be improved by incorporating this invention into the feature learning network or backbone network.

[0068] The following is a more detailed example to further illustrate the invention:

[0069] Example 6:

[0070] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Examples 1-5:

[0071] Step 1: Obtain the pseudo-image feature map after voxelization of the original point cloud:

[0072] In the PointPillars point cloud object detection method, the point cloud is processed by a cylinder feature network to generate a pseudo-image feature map. The KITTI dataset is one of the most authoritative international computer vision evaluation datasets, containing real images and point cloud data collected from urban, rural, and highway scenes. In this embodiment, the internationally published KITTI dataset is used as input, with its point cloud representation R1(x,y,z,r). The PointPillars algorithm is used as the basic framework to perform voxelization preprocessing on the point cloud data, generating a pseudo-image feature map F. pi The specific implementation is as follows:

[0073] 1a) Define the range of point cloud data: In this embodiment, the range of point cloud data on the x-axis is 0 to 69.12, on the y-axis is -39.68 to 39.68, and on the z-axis is -3 to 1. Points outside this range will be ignored.

[0074] 1b) In the xy direction, the point cloud is divided into regular columnar sub-point clouds with a base size of 0.16m × 0.16m. In this embodiment, the point cloud space is divided into 496 × 432 regular columnar sub-point clouds.

[0075] 1c) Set a sampling threshold D, randomly sample the columnar sub-point cloud with a number of points greater than the threshold, and fill the columnar sub-point cloud with zeros with a number of points less than the threshold. In this embodiment, D = 32. If the number of points in the columnar sub-point cloud is greater than 32, then 32 points are randomly sampled; if it is less than 32, then the missing points are filled with 0.

[0076] 1d) Calculate the arithmetic mean (xi) of all points within each columnar sub-point cloud. c ,y c ,z c ) and offset (x) p ,y p The enhanced point cloud (x,y,z,r,x) is obtained. c ,y c ,z c ,x p ,y p ).

[0077] 1e) Utilizing max pooling and a simplified PointNet network to augment point clouds (x,y,z,r,x) c ,y c ,z c ,x p ,y p Encode the image to generate a pseudo-image feature map F. pi .

[0078] In this embodiment, the pseudo image F pi The size is 496×432, and the number of channels is 64.

[0079] Step 2: Use the pseudo-supervoxel segmentation module to perform pseudo-supervoxel segmentation on the pseudo image and obtain the category segmentation mask.

[0080] Directly processing voxels using supervoxel methods is computationally complex and very time-consuming. Therefore, this step involves processing the pseudo-image F obtained from voxels. pi Pseudo-supervoxel segmentation is performed to achieve an effect similar to supervoxel segmentation while reducing computational complexity; and the category segmentation mask T is obtained. mask To construct the spatial relationships between pixels in the pseudo-image.

[0081] This embodiment uses the pseudo-image F pi The region was segmented into 20 superpixel regions using Simple Linear Iterative Clustering (SLIC), and the category segmentation mask T was obtained. mask Category segmentation mask T mask The dimensions are [496, 432].

[0082] Step 3: Use the local meanization module to construct a local mean pseudo-image.

[0083] During voxelization, point clouds lose some spatial information, which reduces the accuracy of target detection. Furthermore, the presence of a large amount of noisy data in point clouds also reduces accuracy without effective processing. Therefore, it is necessary to process the pseudo-image F... piThe spatial information is enhanced and noise is reduced. This step smooths the feature map using local mean operation to achieve a certain degree of noise reduction, and enhances the spatial information representation of the point cloud by utilizing the spatial relationships between pixels in the pseudo-image, ultimately generating a local mean pseudo-image F. lm .

[0084] The specific implementation of the steps is as follows:

[0085] 3a) Starting with category 1, apply the category segmentation mask T mask Perform a scan.

[0086] 3b) For the category segmentation mask T mask For each element in the array, if the value of the element is equal to 1, the result is True; otherwise, it is False. After the scan is completed, a class segmentation mask T is generated. mask A Boolean tensor mask of the same shape, where True indicates that the pixel at that location belongs to category 1, and False indicates that it does not belong to category 1.

[0087] 3c) Based on the Boolean tensor mask, from the pseudo-image F pi Extract all pixels belonging to category 1, calculate the average of the feature values ​​of these pixels, and set this average as the new feature value of these pixels.

[0088] 3d) Repeat steps 3a-3c until the category segmentation mask T is traversed. mask Collect all categories and calculate the corresponding average features to construct a local mean pseudo-image F. lm Local mean pseudo-image F lm With pseudo-image F pi They have the same shape.

[0089] In this embodiment, the category value is 1-20, and the local mean pseudo-image F lm The shape is (496, 432, 64).

[0090] Step 4: Use the factor extraction and aggregation module to generate a pseudo-image feature map after pseudo-supervoxel point cloud preprocessing.

[0091] This step involves fusing the local mean pseudo-image F. lm With pseudo-image F pi Smooth random noise; further obtain the weight distribution of each voxel after super-voxel segmentation and adjust the pseudo-image F. pi While preserving the original spatial information, it fuses supervoxel semantic information without losing edge information, generating a new pseudo-image F'. pi The specific implementation is as follows:

[0092] 4a) Based on the pseudo-image F piand the local mean pseudo-image F lm Generate cross-channel attention vector I:

[0093] I = F gp (F pi ||F lm )

[0094] Where || denotes concatenating the two pseudo-images along the channel dimension, F gp This indicates a global average pooling operation.

[0095] 4b) The supervoxel factor matrix W is obtained by learning the cross-channel attention vector I:

[0096] W=σ(F mlp (I))

[0097] Among them, F mlp Let σ represent an MLP network, and σ represent the ReLU activation function.

[0098] 4c) Transfer the pseudo image F pi Multiplying with the supervoxel factor matrix W yields the pseudo-image F' after pseudo-supervoxel point cloud preprocessing. pi .

[0099] F' pi =W*F pi

[0100] The supervoxel factor matrix W represents the pseudo-image feature map F. pi The weight distribution of each voxel after pseudo-super-voxel segmentation, and the pseudo-image feature map F pi Multiply by the supervoxel factor matrix W to adjust the pseudo-image feature map F. pi The feature weight distribution of each element is used to generate a pseudo-image feature map F' after preprocessing the pseudo-supervoxel point cloud. pi Complete the pseudo-hypervoxel point cloud preprocessing task.

[0101] The technical effects of this invention are verified below using simulation experiments and experimental data:

[0102] Example 7:

[0103] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Examples 1-6:

[0104] I. Simulation Experiment Conditions:

[0105] Dataset: The dataset used is the KITTI dataset, one of the most authoritative computer vision evaluation datasets in the world, which contains real images and point cloud data collected from urban, rural and highway scenes.

[0106] II. Simulation Content

[0107] Based on the PointPillars algorithm framework, this invention is directly placed after Pillar Feature Net, which is a pseudo-image generation network. The pseudo-image feature map generated by Pillar Feature Net is preprocessed with pseudo-supervoxels and then input into the backbone to complete subsequent operations. The experimental results are compared with those of PointPillars on the KITTI dataset for target detection. The results are shown in Table 1. Table 1 shows the comparison results of pseudo-supervoxel preprocessing after inserting the PointPillars pseudo-image generation network into this invention.

[0108] Table 1 Comparison results of the present invention after inserting the PointPillars pseudo-image generation network.

[0109]

[0110] As shown in Table 1, after inputting pseudo-images into this invention, the network's detection accuracy for all three target types was improved to varying degrees, with a more significant improvement for small targets. Compared to the classic target detection algorithm PointPillars, the average accuracy for car detection was improved by 1.81%, with the most significant improvements observed in (bev, easy) and (3d, easy) difficulty levels, at 2.71% and 2.67%, respectively. The average accuracy for pedestrian detection was improved by 2.06%, with the largest improvement reaching 3.72%. The average accuracy for cyclists was improved by 2.99%, with an average improvement of 4.62% for easy difficulty, 2.44% for medium difficulty, and 1.92% for hard difficulty. In each metric, this invention improved the detection accuracy for each target type by at least 0.98% and up to 4.85%. This embodiment demonstrates from experimental data that using this invention for pseudo-supervoxel preprocessing improves the detection accuracy for all target types to varying degrees.

[0111] Example 8:

[0112] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Examples 1-6:

[0113] I. Simulation Experiment Conditions

[0114] Dataset: Same as in Example 7, using the KITT dataset for simulation experiments.

[0115] II. Simulation Content:

[0116] In PointPillars, the pseudo-images generated by the columnar feature network are essentially feature maps. Furthermore, in point cloud object detection algorithms, feature maps of different resolutions are generated at each stage, and these feature maps can characterize the spatial relationships of point clouds within a certain spatial range. Therefore, this invention, used for plug-and-play pseudo-supervoxel point cloud preprocessing, can be applied not only to pseudo-images but also to any feature maps generated in the intermediate stages of the algorithm network.

[0117] This invention is inserted into the backbone of PointPillars, where feature maps of different resolutions obtained from the backbone network are input into this invention, and then the original upsampling and deconvolution operations of the backbone are continued. See also Figure 4 , Figure 4 This is a partial structural diagram of the simulation experiment of this invention. Convolution operations are used to reduce the resolution of the feature maps while increasing their dimensionality. Then, the pseudo-image feature maps obtained by convolution are processed using the present invention with pseudo-supervoxels. Finally, deconvolution operations are used to upsample the three feature maps to the same size. Experimental results are as follows: Figure 5 As shown, Figure 5 The average detection accuracy changes after inserting the present invention into feature maps of different resolutions.

[0118] Figure 5 This demonstrates the improvement in average detection accuracy for three types of targets after preprocessing feature maps of different resolutions, from smallest to largest, using this invention. The horizontal axis corresponds to... Figure 4 Feature maps of different resolutions are shown: (4C, H / 8, W / 8) represents the feature map after the third convolution; (2C, H / 4, W / 4) represents the feature map after the second convolution; (C, H / 2, W / 2) represents the feature map after the first convolution; and (C, H / 2, W / 2) represents the unconvolved feature map, corresponding to the simulation experiment in Example 7. The vertical axis represents the improved detection accuracy, in percentage (%). The experimental results and data trends show that the higher the resolution of the feature map, the more significant the improvement in average detection accuracy after preprocessing with this invention. This is obviously because a higher resolution feature map contains more target information. The average detection accuracy improvement is greatest for cyclists, ranging from 2.52% to 3.33%. This example demonstrates from experimental data that using this invention to perform pseudo-supervoxel preprocessing on feature maps of different resolutions improves the detection accuracy of various targets to varying degrees.

[0119] Example 9:

[0120] A pseudo-supervoxel point cloud preprocessing method for point cloud target detection is the same as in Examples 1-6:

[0121] I. Simulation Experiment Conditions

[0122] Similar to Example 7, the KITT dataset was used for simulation experiments.

[0123] II. Simulation Content

[0124] Based on the voxelization method, experiments were conducted on VoxelNet, SECOND, and PointPillars. This invention was applied after the feature learning networks of VoxelNet and SECOND, enabling the sparse 4D tensors obtained from the feature learning networks to utilize the semantic information of the point cloud space. The results are shown in Tables 2 and 3, which represent the experimental results on the BEV benchmark and the 3D benchmark, respectively.

[0125] Tables 2 and 3 present the experimental results on the BEV and 3D benchmarks, respectively. According to the results in both tables, this invention has a positive impact on the voxelization method. On the BEV benchmark, the average improvement to the original network is above 2%. The highest improvement is seen in PointPillars for easy cyclist detection, increasing from 82.36% to 86.87%. On the 3D benchmark, the average improvement to the original network is also above 2%. The highest improvement is seen in SECOND for easy pedestrian detection, increasing from 55.10% to 59.49%. This embodiment demonstrates from experimental data that incorporating this invention into the feature learning network of the classic voxelization method, followed by pseudo-supervoxel preprocessing, improves the detection accuracy of various methods for various targets to varying degrees.

[0126] Table 2 shows the experimental results on the BEV testing benchmark.

[0127]

[0128] Table 3 shows the experimental results on the 3D detection benchmark.

[0129]

[0130] In summary, this invention provides a pseudo-supervoxel point cloud preprocessing method for point cloud target detection. It addresses the technical problems of most point cloud data preprocessing methods failing to handle noise, causing noise to interfere with target detection; the easy loss of local details and features when simplifying or compressing point cloud data; and the time-consuming nature of directly segmenting point cloud voxels into supervoxels. The method achieves the following: obtaining a pseudo-image feature map, acquiring a category segmentation mask, constructing a local mean pseudo-image that preserves spatial details, and generating a pseudo-image feature map after pseudo-supervoxel point cloud preprocessing. The innovation of this invention lies in: using a superpixel algorithm to achieve a clustering effect similar to that of a supervoxel algorithm on voxels on pseudo-image feature maps, extracting semantic information from supervoxels, and achieving better speed and computational complexity than the supervoxel algorithm. Furthermore, by utilizing the correspondence between the category segmentation mask generated by the superpixel segmentation algorithm and the pseudo-image feature map, local meanization is performed to generate a local mean pseudo-image. A cross-channel attention vector and a supervoxel factor matrix are constructed to adjust the feature weight distribution, enhancing the spatial information semantic expression capability of the pseudo-image feature map. The advantages are: preserving more spatial detail features of the original point cloud, improving target detection accuracy, requiring less time to obtain supervoxel semantic information than supervoxel segmentation, weakening random noise and reducing noise impact, and being directly applicable to feature maps of different resolutions in any point cloud target detection method without fundamentally modifying the algorithm architecture. It can be used in applications such as autonomous driving and robot navigation.

Claims

1. A pseudo-hyper voxel point cloud preprocessing method for point cloud target detection, characterized in that, The correspondence between the class segmentation mask generated by the superpixel segmentation algorithm and the pseudo-image feature map is used to perform local mean valueization to generate a local mean value pseudo-image, and the pseudo-image and the local mean value pseudo-image are learned and fused to adjust the weight distribution of the features in the pseudo-image feature map, including the following steps: (1) Obtain a pseudo-image feature map: obtain the original point cloud to be pre-processed, perform voxelization processing, divide the point cloud space into regular columnar voxels in the x-y direction, encode each columnar voxel using maximum pooling and a simplified PointNet network to obtain voxel features, arrange the voxel features according to their positions in the three-dimensional space to obtain a pseudo-image feature map F pi ; (2) Obtain the class segmentation mask: For the pseudo image feature map F pi Use the pseudo-hyper voxel segmentation module to perform pseudo-hyper voxel segmentation using a simple linear iterative clustering (SLIC) algorithm, only for the pseudo image feature map F pi The first three channels are pseudo-hyper voxel segmented, and the number of hyper pixel labels is specified according to the desired number of hyper pixel regions to generate a class segmentation mask T mask Each element in the class segmentation mask T mask represents the hyper pixel label to which the corresponding pixel point belongs. (3) constructing a local mean module to obtain a local mean pseudo-image retaining spatial detail features: for the pseudo-image feature map F pi and the category segmentation mask T mask , constructing a local mean module; for the pseudo-image feature map F pi , averaging the feature values of the pixel points belonging to the same category and taking them as the feature values of the pixel points in the corresponding position of the local mean pseudo-image F lm , using the corresponding relationship between the category segmentation mask T mask and the pseudo-image feature map F pi , obtaining the local mean pseudo-image F lm retaining spatial detail features in the constructing of the local mean module. (4) A design factor extraction aggregation module is used to generate a pseudo image feature map after preprocessing of the pseudo hyper voxel point cloud: by designing a factor extraction aggregation module, the pseudo image feature map F pi and the local mean pseudo image F lm are learned and fused to adjust the weight distribution of the features in the pseudo image feature map F pi , fuse the hyper voxel space semantic information, generate a pseudo image feature map F' pi after preprocessing of the pseudo hyper voxel point cloud, and complete the pseudo hyper voxel point cloud preprocessing task for the point cloud target detection of the original point cloud to be preprocessed.

2. The pseudo-hyper voxel point cloud preprocessing method for point cloud target detection according to claim 1, characterized in that, The local mean valueization module in step (3) is constructed to obtain a local mean value pseudo-image that retains spatial detail features, including: 3.1) Scan class segmentation mask: For class segmentation mask T mask starting from class 1, scan the elements in the class segmentation mask T mask row by row; 3.2) Calculate the mean value of the regions in the same category: scan the category segmentation mask T mask During the process, record the position of the pixel point with category 1, extract the feature value of the corresponding position in the pseudo-image feature map F pi and calculate the mean value of the extracted feature value. 3.3) constructing a local mean pseudo-image F lm The pixel points in the category 1 are assigned values: according to the positions of the pixel points in the category 1 and the calculated mean value, a local mean pseudo-image F is constructed lm The obtained mean value is taken as the local mean pseudo-image F lm The feature value at the corresponding position 3.4) Construct a local mean pseudo-image F that preserves spatial detail features. lm Repeat steps (3.1)-(3.3) until the category segmentation mask T is traversed. mask For all categories, calculate the mean feature corresponding to each category, and then assign the mean feature to each category in the pseudo-image feature map F. pi The arrangement of positions on the image ultimately generates a local mean pseudo-image F that preserves spatial detail features. lm .

3. The pseudo-hyper voxel point cloud preprocessing method for point cloud object detection according to claim 2, characterized in that, The recording category division mask T described in step (3.2) mask the positions in which the category is 1, and extracts the pseudo image F belonging to the category corresponding to 1 according to the recorded positions pi the values of the pixel points in the positions in which the category is 1 and sums and calculates the average value, including: 3.2.1) Generate a Boolean tensor mask indicating whether the corresponding element is of the specified class 1: For each element in the class segmentation mask T mask , if the value of the element is equal to 1, the result is True, otherwise False, for the elements in the class segmentation mask T mask , after the scan is completed, a Boolean tensor mask of the same shape as the class segmentation mask T mask is generated, where True indicates that the pixel at that position belongs to class 1, and False indicates that it does not belong. 3.2.3) Root Boolean tensor mask extracts pseudo image F pi All pixels with category 1: According to the Boolean tensor mask, if the pixel value at a certain position in the Boolean tensor mask is 1, the feature value of the pixel point corresponding to the position in the pseudo image F pi is extracted; if the pixel value at a certain position in the Boolean tensor mask is 0, the feature value of the pixel point corresponding to the position in the pseudo image F pi is not operated. 3.2.3) calculating the average value of the extracted feature values on the pixel points.

4. The pseudo-hyper voxel point cloud preprocessing method for point cloud target detection according to claim 1 or 2 or 3, characterized in that, The factor extraction and aggregation module in step (4) is designed to generate a pseudo-super voxel point cloud preprocessed pseudo-image feature map, including: 4.1) generating a collectible pseudo-image F pi and expressive statistical information of the local mean pseudo-image F lm a cross-channel expressive attention vector I from the pseudo-image F pi and the local mean pseudo-image F lm , generating a cross-channel expressive attention vector I: I = F gp (F pi ||F lm ) where || denotes concatenating the two pseudo images in the channel dimension, F gp denotes the global average pooling operation; 4.2) generating a pseudo image F characterizing said pseudo image F pi The super voxel factor matrix W of the weight distribution of each voxel after super voxel segmentation: The super voxel factor matrix W is learned from the cross-channel representation attention vector I: W = σ (F mlp (I)) where F mlp denotes the MLP network, and σ denotes the relu activation function. 4.3) adjusting the pseudo-image feature map F according to the super voxel factor matrix W pi , generating the pseudo-image feature map F of the pre-processed pseudo-super voxel point cloud pi : multiplying the pseudo-image feature map F pi and the super voxel factor matrix W, fusing the super voxel space semantic information, and generating the pseudo-image feature map F of the pre-processed pseudo-super voxel point cloud pi : F' pi = W * F pi The pseudo super voxel point cloud preprocessed pseudo image feature map F pi The pseudo image feature map F pi The shape is the same.

5. The pseudo-hyper voxel point cloud preprocessing method for point cloud target detection according to claim 1 or 2 or 3 or 4, characterized in that, The pseudo-super voxel point cloud preprocessing for plug-and-play is directly applied to different resolution feature maps in any point cloud target detection method without fundamentally modifying the algorithm architecture, including: a) In the backbone network of PointPillars, different resolution feature maps are processed using the pseudo-super voxel point cloud preprocessing method, and then the original upsampling operation in the PointPillars backbone network is continued; b) In the feature learning network of the VoxelNet and SECOND algorithms, the 4D tensor obtained by the feature learning network is processed using the pseudo-super voxel point cloud preprocessing method, and then the predetermined operation of the VoxelNet and SECOND algorithms is continued.

Images