3D point cloud target detection method and system based on random mapping

Abstract

The invention relates to a 3D point cloud target detection method and system based on random mapping, and the method comprises the steps: compressing a point cloud implicit vector of a 3D point cloud target detection image into a low-dimensional vector through a lower projection matrix, carrying out the L2 norm normalization processing, and obtaining a normalized vector; projecting the normalized vector to k one-dimensional axes through k independent random linear mapping functions to generate k projection value sequences, screening out the projection value sequence with the maximum variance as an optimal sequence, and calculating a vector difference value between the normalized query vector and the normalized key vector based on the optimal sequence so as to realize approximate nearest neighbor search. And the trained Transform target detection model carries out target detection on the 3D point cloud target detection image according to the sparse attention operator to obtain a target detection result. Therefore, the detection efficiency and the detection accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a 3D point cloud target detection method based on random mapping. Background Technology

[0002] When performing target detection on 3D point cloud images, the unordered, sparse, large-scale, and unstructured nature of point cloud data necessitates that standard global self-attention mechanisms construct relationships between all point pairs during detection. This leads to a quadratic increase in both time and space complexity with the number of points. This computational bottleneck severely limits its practical application in high-density, large-scale point cloud scenarios, making it difficult to meet the requirements of real-time systems in terms of inference speed and memory usage, thus impacting detection efficiency. Furthermore, global attention mechanisms tend to introduce a large amount of irrelevant noise when processing discrete point clouds, weakening the perception of key local structures and easily leading to homogenization, thereby reducing the model's ability to distinguish geometric details and salient features, and affecting the accuracy of detection results.

[0003] To address these challenges, existing research has proposed local attention mechanisms. These methods are typically based on spatial partitioning strategies, such as k-nearest neighbor graphs or voxel grids, which restrict attention computation to the local neighborhood of each point, preserving fine-grained structural details while reducing computational overhead. However, due to their limited receptive field, these methods struggle to model non-local semantic relationships and global structural regularities, leading to performance degradation in tasks requiring holistic scene understanding. Their performance is limited in tasks requiring holistic scene understanding and fails to achieve performance comparable to global attention mechanisms, thus affecting the accuracy of detection results. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: to provide a 3D point cloud target detection method and system based on random mapping, which improves the accuracy of detection results while improving detection efficiency.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a 3D point cloud target detection method based on random mapping, comprising: Obtain the latent vector of the 3D point cloud target detection image, compress the latent vector into a low-dimensional vector through the down projection matrix, and perform L2 norm normalization on the low-dimensional vector to obtain a normalized vector, which includes a normalized query vector and a normalized key vector. The normalized vector is projected onto k one-dimensional axes using k independent random linear mapping functions to generate k sequences of projected values ​​containing k projected values. The variance of the k sequences of projected values ​​is calculated and the sequence with the largest variance is selected as the optimal sequence. Based on the optimal sequence, the vector difference between the normalized query vector and the normalized key vector is calculated. Based on the vector difference, m normalized key vectors similar to the normalized query vector are obtained to achieve approximate nearest neighbor search. The m normalized key vectors are mapped back to the high-dimensional space through the up projection matrix to obtain m value values. The m value values ​​are weighted and summed through attention scores to obtain the weighted sum result. The weighted sum result is used as the sparse attention operator. A trained Transformer object detection model is obtained, and the trained Transformer object detection model performs object detection on the 3D point cloud object detection image according to the sparse attention operator to obtain the object detection result.

[0006] The beneficial effects of this invention are as follows: By using a projection matrix, the latent vectors of 3D point cloud target detection images are compressed into low-dimensional vectors. This involves projecting high-dimensional point cloud data into a low-dimensional space and performing L2 norm normalization, eliminating norm differences in low-dimensional vectors and the influence of vector length on the dot product. This ensures that subsequent approximate nearest neighbor search can be performed without significant accuracy loss. The normalized vectors are projected onto k one-dimensional axes using k random linear projection functions, generating a sequence of projection values ​​containing k projection values. The sequence with the largest variance is used as the optimal sequence for approximate nearest neighbor search, avoiding similarity calculations between irrelevant information and overcoming the limitations of similarity calculations between different dimensions. While enhancing the qualitative aspects, it also improves semantic judgment capabilities. It retains the most significant features in the latent vectors of point clouds and achieves fine-grained modeling of sparse attention through dynamic approximate nearest neighbor search. This enables approximate nearest neighbor search of relevant key vectors centered on a single query vector in the multidimensional latent space, while keeping the complexity of attention score calculation within a linear range. This meets the requirements of real-time systems in terms of inference speed and memory usage, thereby improving detection efficiency. The calculated sparse attention operator is combined with the trained Transformer object detection model to perform object detection on 3D point cloud object detection images, improving the accuracy of object detection results.

[0007] Optionally, the random linear mapping function is: ; ; ; ; in, Indicates about Let j be the j-th random linear mapping function, and k represent the total number of random linear mapping functions. Indicates about linear matrix and All elements in the dataset follow a standard normal distribution. This represents the i-th normalized vector. This indicates that the dimension is d.

[0008] As described above, the elements in the linear matrix all follow a standard normal distribution. Therefore, the parameters of the stochastic linear mapping function also follow a standard state distribution. By projecting the normalized vector onto the stochastic linear mapping function with a standard normal distribution, the relative distance between point pairs is preserved, which enables the effective estimation of the Euclidean distance between high-dimensional points in a low-dimensional space.

[0009] Optionally, the normalized vector is projected onto k one-dimensional axes using k independent random linear mapping functions to generate a sequence of projected values ​​containing k projected values: The normalized query vector and the normalized key vector are projected onto k one-dimensional axes through a first linear layer of k independent random linear mapping functions. At the same time, a sequence of projected values ​​containing k projected values ​​is generated through a second linear layer of k independent random linear mapping functions. The first linear layer and the second linear layer are not the same.

[0010] As described above, by projecting the normalized query vector and the normalized key vector onto the same linear layer, the resulting query vector and key vector are kept consistent, thus eliminating feature space bias, enhancing feature distribution alignment, and reducing model parameter size and computational overhead.

[0011] Optionally, the step of calculating the variance of the k projected value sequences and selecting the projected value sequence with the largest variance as the optimal sequence includes: The normalized vectors are sorted in ascending order according to their corresponding variances and placed into a contiguous memory block. The normalized vectors are then extracted from the contiguous memory block in parallel through a sliding window of length m to perform memory reorganization on the normalized vectors. The resulting memory-reorganized normalized vectors are then selected, and the sequence of projected values ​​with the largest variance is chosen as the optimal sequence.

[0012] As described above, memory reorganization of normalized vectors optimizes memory usage, thereby integrating and optimizing near nearest neighbor search and attention score calculation into standard matrix operations, improving hardware utilization and training efficiency.

[0013] Secondly, the present invention provides a 3D point cloud target detection system based on random mapping, comprising: The compression module is used to obtain the latent vector of the point cloud in the 3D point cloud target detection image, compress the latent vector of the point cloud into a low-dimensional vector through the lower projection matrix, and perform L2 norm normalization on the low-dimensional vector to obtain a normalized vector. The normalized vector includes a normalized query vector and a normalized key vector. The projection module is used to project the normalized vector onto k one-dimensional axes through k independent random linear mapping functions, generate k sequences of projection values ​​containing k projection values, calculate the variance of the k sequences of projection values, and select the sequence of projection values ​​with the largest variance as the optimal sequence. The generation module is used to calculate the vector difference between the normalized query vector and the normalized key vector based on the optimal sequence, obtain m normalized key vectors similar to the normalized query vector based on the vector difference to achieve approximate nearest neighbor search, map the m normalized key vectors back to the high-dimensional space through the up projection matrix to obtain m value values, perform weighted summation on the m value values ​​through attention scores to obtain the weighted summation result, and use the weighted summation result as a sparse attention operator; The detection module is used to obtain a trained Transformer object detection model. The trained Transformer object detection model performs object detection on the 3D point cloud object detection image according to the sparse attention operator to obtain the object detection result.

[0014] The beneficial effects of this invention are as follows: By using a projection matrix, the latent vectors of 3D point cloud target detection images are compressed into low-dimensional vectors. This involves projecting high-dimensional point cloud data into a low-dimensional space and performing L2 norm normalization, eliminating norm differences in low-dimensional vectors and the influence of vector length on the dot product. This ensures that subsequent approximate nearest neighbor search can be performed without significant accuracy loss. The normalized vectors are projected onto k one-dimensional axes using k random linear projection functions, generating a sequence of projection values ​​containing k projection values. The sequence with the largest variance is used as the optimal sequence for approximate nearest neighbor search, avoiding similarity calculations between irrelevant information and overcoming the limitations of similarity calculations between different dimensions. While enhancing the qualitative aspects, it also improves semantic judgment capabilities. It retains the most significant features in the latent vectors of point clouds and achieves fine-grained modeling of sparse attention through dynamic approximate nearest neighbor search. This enables approximate nearest neighbor search of relevant key vectors centered on a single query vector in the multidimensional latent space, while keeping the complexity of attention score calculation within a linear range. This meets the requirements of real-time systems in terms of inference speed and memory usage, thereby improving detection efficiency. The calculated sparse attention operator is combined with the trained Transformer object detection model to perform object detection on 3D point cloud object detection images, improving the accuracy of object detection results.

[0015] Optionally, the random linear mapping function is: ; ; ; ; in, Indicates about Let j be the j-th random linear mapping function, and k represent the total number of random linear mapping functions. Indicates about linear matrix and All elements in the dataset follow a standard normal distribution. This represents the i-th normalized vector. This indicates that the dimension is d.

[0016] As described above, the elements in the linear matrix all follow a standard normal distribution. Therefore, the parameters of the stochastic linear mapping function also follow a standard state distribution. By projecting the normalized vector onto the stochastic linear mapping function with a standard normal distribution, the relative distance between point pairs is preserved, which enables the effective estimation of the Euclidean distance between high-dimensional points in a low-dimensional space.

[0017] Optionally, the projection module specifically comprises: The normalized query vector and the normalized key vector are projected onto k one-dimensional axes through a first linear layer of k independent random linear mapping functions. At the same time, a sequence of projected values ​​containing k projected values ​​is generated through a second linear layer of k independent random linear mapping functions. The first linear layer and the second linear layer are not the same.

[0018] As described above, by projecting the normalized query vector and the normalized key vector onto the same linear layer, the resulting query vector and key vector are kept consistent, thus eliminating feature space bias, enhancing feature distribution alignment, and reducing model parameter size and computational overhead.

[0019] Optionally, the projection module includes: The memory optimization module is used to sort the normalized vectors in ascending order according to their corresponding variances into a contiguous memory block, and extract the normalized vectors from the contiguous memory block in parallel through a sliding window of length m, so as to realize the memory reorganization of the normalized vectors, obtain the memory reorganized normalized vectors, and select the sequence of projection values ​​with the largest variance as the optimal sequence.

[0020] As described above, memory reorganization of normalized vectors optimizes memory usage, thereby integrating and optimizing near nearest neighbor search and attention score calculation into standard matrix operations, improving hardware utilization and training efficiency. Attached Figure Description

[0021] Figure 1 This is a flowchart of a 3D point cloud target detection method based on random mapping provided in this embodiment; Figure 2 This is a schematic diagram of the overall process of a 3D point cloud target detection method based on random mapping provided in this embodiment; Figure 3 This is a schematic diagram of the structure of a 3D point cloud target detection system based on random mapping provided in this embodiment.

[0022] [Explanation of Labels in the Attached Image] 1. A 3D point cloud target detection system based on random mapping; 2. Compression module; 3. Projection module; 31. Memory optimization module; 4. Generation module; 5. Detection module. Detailed Implementation

[0023] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0024] Example 1 Please refer to Figures 1 to 2 This invention provides a 3D point cloud target detection method based on random mapping, comprising the following steps: S1. Obtain the latent vector of the point cloud in the 3D point cloud target detection image, compress the latent vector of the point cloud into a low-dimensional vector through the down projection matrix, and perform L2 norm normalization on the low-dimensional vector to obtain a normalized vector. The normalized vector includes a normalized query vector and a normalized key vector. In this embodiment, as Figure 2As shown, the latent vectors of the point cloud are obtained. These latent vectors refer to all query vectors and all key vectors in the high-dimensional space. The lower projection matrix is ​​used to compress all query vectors and key vectors into low-dimensional vectors. That is, the high-dimensional point cloud data is compressed into low-dimensional point cloud data. The low-dimensional vectors are then normalized using the L2 norm to obtain normalized vectors. The lower projection matrix is: ; ; in, Let d represent the learnable parameter matrix, and let d represent the dimension of the point cloud latent vectors. l Indicates the dimension of a low-dimensional vector; A low-dimensional vector can be represented as: ; A normalized vector can be represented as: ; Where j represents the total number of hidden vectors in the point cloud.

[0025] S2. Project the normalized vector onto k one-dimensional axes using k independent random linear mapping functions to generate k projection value sequences containing k projection values. Calculate the variance of the k projection value sequences and select the projection value sequence with the largest variance as the optimal sequence. At this point, step S2, which involves projecting the normalized vector onto k one-dimensional axes using k independent random linear mapping functions to generate a sequence of projected values ​​containing k projected values, includes: S21. The normalized query vector and the normalized key vector are projected onto k one-dimensional axes through the first linear layer of k independent random linear mapping functions. At the same time, a sequence of projected values ​​containing k projected values ​​is generated through the second linear layer of k independent random linear mapping functions. The first linear layer and the second linear layer are not the same.

[0026] In this embodiment, as Figure 2 As shown, a normalized vector is projected onto k one-dimensional axes using k independent random linear mapping functions. Specifically, each of the k independent random linear mapping functions projects the normalized vector onto a point on the one-dimensional axis, i.e., a scalar value, thus obtaining the corresponding projection value. The projection values ​​of each normalized vector in k different directions are collected to form a sequence of projection values ​​containing k projection values. The variance of the k projection value sequences is calculated, and the sequence with the largest variance is selected as the optimal sequence. In other words, the sequence with the most dispersed data distribution that best reflects the difference between the query vector and the key vector is chosen.

[0027] In this embodiment, the random linear mapping function is: ; ; ; ; in, Indicates about Let j be the j-th random linear mapping function, and k represent the total number of random linear mapping functions. Indicates about linear matrix and All elements in the dataset follow a standard normal distribution. This represents the i-th normalized vector. This indicates that the dimension is d.

[0028] because All elements in the array follow a standard normal distribution, i.e., Therefore, the parameters of the random linear mapping function are actually independently sampled from the standard normal distribution. When projecting through the random linear mapping function, the same linear layer is used to project the query vector and key vector onto k one-dimensional axes, while different linear layers are used to generate a sequence of projected values ​​containing k projected values.

[0029] At this point, step S2, which involves calculating the variance of the k projected value sequences and selecting the projected value sequence with the largest variance as the optimal sequence, includes: S22. Sort the normalized vectors in ascending order according to their corresponding variances and extract the normalized vectors from the continuous memory blocks in parallel through a sliding window of length m, so as to realize the memory reorganization of the normalized vectors, obtain the memory reorganized normalized vectors, and select the projection value sequence with the largest variance as the optimal sequence.

[0030] In this embodiment, as Figure 2 As shown, the normalized vectors are sorted in ascending order according to the variance of their corresponding projection value sequences and placed into contiguous memory blocks. The normalized vectors are then extracted from the contiguous memory blocks in parallel through a sliding window of length m, thereby reorganizing the memory of the normalized vectors and obtaining the reorganized normalized vectors. This enables the subsequent approximate nearest neighbor search and attention score calculation to be integrated and optimized into standard matrix operations, and the projection value sequence with the largest variance is selected as the optimal sequence.

[0031] S3. Calculate the vector difference between the normalized query vector and the normalized key vector based on the optimal sequence. Obtain m normalized key vectors similar to the normalized query vector based on the vector difference to achieve approximate nearest neighbor search. Map the m normalized key vectors back to the high-dimensional space through the up projection matrix to obtain m value values. Perform a weighted summation on the m value values ​​through attention scores to obtain a weighted summation result. Use the weighted summation result as a sparse attention operator.

[0032] In this embodiment, since the normalized vectors, namely the normalized query vector and the normalized key vector, have already undergone L2 norm normalization, the Euclidean distance between a single normalized query vector and all normalized key vectors can directly reflect the magnitude of their dot product. This point can be demonstrated through the following derivation: Define two point cloud latent vectors of dimension d. , ; Its Euclidean distance formula is: ; Its dot product formula is: ; Based on the above Euclidean distance formula and dot product formula, we can know that: ; when and After being normalized by the L2 norm: ; That is, the square of ED is negatively correlated with DP. Specifically, the larger the DP, the higher the similarity between the two vectors, and the smaller the corresponding ED. The Euclidean distance between a single normalized query vector and all normalized key vectors can directly reflect the size of their dot product.

[0033] At this point, the problem of selecting key normalized query vector-key vector point pairs is transformed into an approximate nearest neighbor search problem based on Euclidean distance. Theoretically, approximate nearest neighbor search can be achieved by calculating the Euclidean distance between all normalized query vectors and all normalized key vectors and selecting m nearest neighbor normalized key vectors for each normalized query vector. However, this method is too brute-force, leading to a second increase in computational complexity and inefficiency when processing large-scale point cloud data. Therefore, in this embodiment, as... Figure 2As shown, the vector difference between the normalized query vector and the normalized key vector is calculated based on the optimal sequence. Based on the vector difference, m normalized key vectors similar to the normalized query vector are obtained to achieve approximate nearest neighbor search. The m normalized key vectors are then mapped back to the high-dimensional space through an up-projection matrix to obtain m value values. The m value values ​​are then weighted and summed using attention scores to obtain a weighted sum result, which is used as a sparse attention operator.

[0034] Furthermore, in this embodiment, when projecting the latent vectors of the point cloud from the high-dimensional space to the low-dimensional space, the relative distance relationship between point pairs can still be maintained, thereby achieving the subsequent approximate nearest neighbor search. This can be demonstrated as follows: Two point cloud latent vectors of dimension d are input into a random linear mapping function and the following calculations are performed: ; ; ; This represents the projection difference of the j-th random linear projection function; make and The expected value is: ; Right now: ; If there exist k random linear projection functions, then they can be generated independently. k times, therefore through the mean absolute difference Approaching This improves the approximate accuracy of high-dimensional Euclidean distance, namely: ; S4. Obtain the trained Transformer object detection model. The trained Transformer object detection model performs object detection on the 3D point cloud object detection image according to the sparse attention operator to obtain the object detection result.

[0035] In this embodiment, as Figure 2 As shown, the sparse attention operator calculated in step S3 is combined with the trained Transformer object detection model to perform object detection on the 3D point cloud object detection image. That is, the trained Transformer object detection model performs object detection on the 3D point cloud object detection image according to the sparse attention operator to obtain the object detection result.

[0036] The 3D point cloud target detection method in this embodiment can be applied to multiple scenarios, such as 3D point cloud shape classification and 3D point cloud semantic segmentation.

[0037] The following specific application scenario, 3D point cloud shape classification, further demonstrates that the 3D point cloud target detection method based on random mapping provided in this embodiment improves detection efficiency and accuracy.

[0038] The shape classification performance was evaluated on two classic 3D point cloud benchmark datasets, ModelNet40 and ScanObjectNN. ModelNet40 contains 40 classes, and the evaluation was conducted using its official 9843 training samples and 2468 test samples. The performance metric used was average accuracy. Overall accuracy (OA). ScanObjectNN includes 15 categories, from which 36,542 training samples and 10,240 test samples were selected to evaluate the same metrics. All samples maintained a size of 2048 points in the experiments. Full attention (FA) was selected as the benchmark, and based on the object of attention, it was divided into two paradigms: Point-Wise (point-to-point) and Neighborhood-Wise (neighborhood points). Performance was quantitatively evaluated through a unified comparative experiment.

[0039] Point-Wise: Point-to-point: It takes the coordinates and features of the original point cloud as direct input and performs fine-grained attention modeling at the level of basic elements (i.e., individual points) of the 3D point cloud. It captures point-by-point semantic relationships through global point interactions, thereby modeling the contextual relationships between all points.

[0040] Neighborhood-Wise: First, n center points are selected using a farthest-point sampling algorithm. Then, a local neighborhood is divided for each center point based on a K-nearest neighbor or sphere query strategy, constructing sub-region units containing structured spatial information. Finally, max pooling is performed on each neighborhood to obtain a unique local representation, ensuring permutation invariance within the neighborhood. Based on this, region-level attention computation is performed on each neighborhood to achieve aggregation of local geometric features and semantic refinement.

[0041] Table 1. Shape classification experiments of the present invention on ModelNet40 and ScanObjectNN

[0042] As shown in Table 1, the proposed 3D point cloud object detection method based on random mapping achieves state-of-the-art performance in shape classification tasks. Compared with the established full attention:FA baseline, this invention performs global key information filtering before attention semantic aggregation. This process filters out a large amount of irrelevant information, significantly improving the model's generalization ability. Compared to PW and NW, the latent features generated during the attention process contain broader contextual information, thus possessing superior shape understanding capabilities. In Table 1, PT represents Point Transformer: a point transformation neural network; PTv2 represents Point Transformer version 2: the second version of the point transformation neural network; PVT represents Point Visual Transformer: a point visual transformation neural network; Point-BERT represents: a point cloud BERT model; PCT represents: Point Cloud Transformer: a point cloud transformation neural network; LCPFormer represents: a local point cloud transformation neural network; MVTN represents: a multi-view transformation network; PW-FA (baseline): point-by-point feature aggregation (baseline); NW-FA (baseline): neighborhood point-by-point feature aggregation (baseline); Point2Vec represents: a model that converts point clouds into vectors; SPoTr represents: a structured point transformation neural network; Point-Tnt represents: a point cloud registration network; and oint-MAE represents: a point cloud mask autoencoder.

[0043] The following specific application scenario, 3D point cloud semantic segmentation, further demonstrates that the 3D point cloud target detection method based on random mapping provided in this embodiment improves detection efficiency and accuracy.

[0044] We conducted 3D semantic segmentation experiments on the ShapeNetPart and S3DIS datasets, and selected full attention:FA as the test baseline.

[0045] ShapeNetPart: This dataset is a classic benchmark dataset in the field of 3D computer vision for 3D shape part segmentation tasks, built on the large-scale 3D model library ShapeNet. The dataset contains 16,881 3D models across 16 object categories, such as cars, airplanes, and chairs. Each model is labeled with 2-5 semantic part labels, such as "fuselage," "wheel," and "chair leg." The data is stored in polygonal mesh format, and corresponding part segmentation annotation files are provided. As an authoritative testing platform for part segmentation tasks, its standard partitioning provides a unified benchmark for algorithm evaluation and is widely used to verify the performance of deep learning models such as point cloud segmentation and mesh segmentation. A dataset of 16,800 training samples and 2,900 test samples is constructed from this dataset, using the instance mean intersection-over-union ratio (IoU). Intersection over Union (IoU) of categories: Evaluate the semantic segmentation capability of this invention.

[0046] S3DIS: This dataset is a significant benchmark in 3D computer vision, focusing on semantic segmentation and instance segmentation tasks for indoor scenes. It is constructed from laser scan data of six large-scale indoor environments, such as offices and classrooms, providing dense point cloud annotations for 13 semantic categories, such as floors, walls, chairs, and desks, covering 2.5 billion points across 271 rooms. Region 5 is reserved for testing during training following a common protocol, and the overall point-by-point accuracy is: Category average accuracy: Intersection over Union (IoU) of categories: Evaluate the performance of S3DIS region 5.

[0047] Table 2. Semantic segmentation experiments of the present invention on ShapeNetPart and S3DIS

[0048] Table 2 shows the comparison results of the present invention with existing attention mechanisms on ShapeNetPart and S3DIS. The present invention outperforms existing methods on all evaluation metrics. Furthermore, on the S3DIS validation set, the present invention outperforms the baseline FA. The metrics improved significantly by 9.0%. In Table 2, PointNet represents the classic 3D point cloud processing model, PointNet++ represents an improved version of PointNet, SPoTr represents a structured point transformation neural network, PT represents Point Transformer, PVT represents Point Visual Transformer, PCT represents Point Cloud Transformer, FA (baseline) represents feature aggregation (baseline), CT2 represents a context transformation neural network, and PTv3 represents Point Transformer version 3.

[0049] Example 2 Please refer to Figure 3 The present invention provides a 3D point cloud target detection system 1 based on random mapping, comprising: a compression module 2, a projection module 3, a memory optimization module 31, a generation module 4, and a detection module 5.

[0050] Among them, compression module 2 is used to obtain the point cloud latent vector of the 3D point cloud target detection image, compress the point cloud latent vector into a low-dimensional vector through the lower projection matrix, and perform L2 norm normalization on the low-dimensional vector to obtain a normalized vector. The normalized vector includes a normalized query vector and a normalized key vector. Projection module 3 is used to project the normalized vector onto k one-dimensional number axes through k independent random linear mapping functions, generate a sequence of k projected values ​​containing k projected values, calculate the variance of the k projected value sequences, and select the projected value sequence with the largest variance as the optimal sequence. The generation module 4 is used to calculate the vector difference between the normalized query vector and the normalized key vector based on the optimal sequence, obtain m normalized key vectors similar to the normalized query vector based on the vector difference to achieve approximate nearest neighbor search, map the m normalized key vectors back to the high-dimensional space through the up projection matrix to obtain m value values, perform weighted summation on the m value values ​​through attention scores to obtain the weighted summation result, and use the weighted summation result as a sparse attention operator; The detection module 5 is used to obtain a trained Transformer object detection model. The trained Transformer object detection model performs object detection on the 3D point cloud object detection image according to the sparse attention operator to obtain the object detection result.

[0051] Specifically, the random linear mapping function is: ; ; ; ; in, Indicates about Let j be the j-th random linear mapping function, and k represent the total number of random linear mapping functions. Indicates about linear matrix and All elements in the dataset follow a standard normal distribution. This represents the i-th normalized vector. This indicates that the dimension is d.

[0052] Specifically, the projection module 3 is as follows: The normalized query vector and the normalized key vector are projected onto k one-dimensional axes through a first linear layer of k independent random linear mapping functions. At the same time, a sequence of projected values ​​containing k projected values ​​is generated through a second linear layer of k independent random linear mapping functions. The first linear layer and the second linear layer are not the same.

[0053] Specifically, the projection module 3 includes: The memory optimization module 31 is used to sort the normalized vectors in ascending order according to their corresponding variances into a contiguous memory block, and extract the normalized vectors from the contiguous memory block in parallel through a sliding window of length m, so as to realize the memory reorganization of the normalized vectors, obtain the memory reorganized normalized vectors, and select the projection value sequence with the largest variance as the optimal sequence.

[0054] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0055] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0057] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0058] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0059] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0060] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A 3D point cloud object detection method based on random mapping, characterized in that, The method comprises the following steps: obtain a point cloud hidden vector of a 3D point cloud target detection image, compress the point cloud hidden vector into a low-dimensional vector through a down-projection matrix, and perform L2 norm normalization processing on the low-dimensional vector to obtain a normalized vector, the normalized vector comprising a normalized query vector and a normalized key vector; project the normalized vector onto k one-dimensional number axes through k independent random linear mapping functions to generate k projection value sequences containing k projection values, calculate the variances of the k projection value sequences, and select the projection value sequence with the largest variance as an optimal sequence; calculate the vector difference between the normalized query vector and the normalized key vector based on the optimal sequence, obtain m normalized key vectors similar to the normalized query vector according to the vector difference to realize approximate nearest neighbor search, map the m normalized key vectors back to a high-dimensional space through an up-projection matrix to obtain m value values, and perform weighted summation on the m value values through attention scores to obtain a weighted summation result, which is taken as a sparse attention operator; obtain a trained Transformer target detection model, and perform target detection on the 3D point cloud target detection image according to the sparse attention operator to obtain a target detection result.

2. The 3D point cloud object detection method based on random mapping of claim 1, wherein, The random linear mapping function is: ； ； ； ； in, Indicates about Let j be the j-th random linear mapping function, and k represent the total number of random linear mapping functions. Indicates about linear matrix and All elements in the dataset follow a standard normal distribution. This represents the i-th normalized vector. This indicates that the dimension is d.

3. The 3D point cloud object detection method based on random mapping of claim 1, wherein, the normalized vector is projected onto k one-dimensional number axes through k independent random linear mapping functions to generate a projection value sequence containing k projection values; the normalized query vector and the normalized key vector are projected onto k one-dimensional number axes through a first linear layer of the k independent random linear mapping functions, and a projection value sequence containing k projection values is generated through a second linear layer of the k independent random linear mapping functions; the first linear layer and the second linear layer are different.

4. The 3D point cloud object detection method based on random mapping of claim 1, wherein, The calculation of the variances of the k projection value sequences and the selection of the projection value sequence with the largest variance as the optimal sequence comprise: the normalized vector is sorted in ascending order according to the corresponding variances into a continuous memory block, and the normalized vector is re-extracted from the continuous memory block through a sliding window with a length of m in parallel to realize memory reorganization of the normalized vector, obtain a memory-reorganized normalized vector, and select the projection value sequence with the largest variance as the optimal sequence.

5. A random mapping based 3D point cloud object detection system, characterized in that, The method comprises the following steps: a compression module is configured to obtain a point cloud hidden vector of a 3D point cloud target detection image, compress the point cloud hidden vector into a low-dimensional vector through a down-projection matrix, and perform L2 norm normalization processing on the low-dimensional vector to obtain a normalized vector, the normalized vector comprising a normalized query vector and a normalized key vector; a projection module is configured to project the normalized vector onto k one-dimensional number axes through k independent random linear mapping functions to generate k projection value sequences containing k projection values, calculate the variances of the k projection value sequences, and select the projection value sequence with the largest variance as an optimal sequence; The generating module is configured to calculate a vector difference between the normalized query vector and the normalized key vector based on the optimal sequence, obtain m normalized key vectors similar to the normalized query vector according to the vector difference, implement approximate nearest neighbor search, map the m normalized key vectors back to the high-dimensional space through an upper projection matrix to obtain m value values, perform weighted summation on the m value values through attention scores to obtain a weighted summation result, and take the weighted summation result as the sparse attention operator; The detecting module is configured to obtain a trained Transformer target detection model, and the trained Transformer target detection model performs target detection on the 3D point cloud target detection image according to the sparse attention operator to obtain a target detection result.

6. The random mapping based 3D point cloud object detection system of claim 5, wherein, The random linear mapping function is: ； ； ； ； in, Indicates about Let j be the j-th random linear mapping function, and k represent the total number of random linear mapping functions. Indicates about linear matrix and All elements in the dataset follow a standard normal distribution. This represents the i-th normalized vector. This indicates that the dimension is d.

7. The random mapping based 3D point cloud object detection system of claim 5, wherein, The projection module specifically is: The normalized query vector and the normalized key vector are projected to k one-dimensional number axes through a first linear layer of k independent random linear mapping functions, and a projection value sequence containing k projection values is generated through a second linear layer of the k independent random linear mapping functions. The first linear layer and the second linear layer are different.

8. The random mapping based 3D point cloud object detection system of claim 5, wherein, The projection module includes: The memory optimization module is configured to sort the normalized vectors in ascending order according to corresponding variances to a continuous memory block, and reextract the normalized vectors from the continuous memory block in parallel through a sliding window with a length of m to implement memory reorganization of the normalized vectors, obtain the normalized vectors after memory reorganization, and screen out a projection value sequence with the largest variance as the optimal sequence.

Images