Multi-stage three-dimensional target detection method, system and device fusing point cloud converter and spatial feature enhancement, and readable storage medium

Abstract

The invention discloses a multi-stage three-dimensional target detection method, system and device integrating a point cloud converter and spatial feature enhancement and a readable storage medium, and relates to the technical field of point cloud data processing, artificial intelligence and computer vision. Comprising the following steps: S1, acquiring an outdoor environment based on a distance sensing sensor to obtain original point cloud data; s2, performing denoising and downsampling processing on the obtained original point cloud data to obtain updated point cloud data; s3, constructing a multi-stage point cloud detection model based on Point TransformerV3 and PnP3D, and training the multi-stage point cloud detection model to obtain a trained point cloud detection model; and S4, detecting the updated point cloud data by using the trained point cloud detection model to obtain a point cloud detection result. According to the method, the expression ability of local space geometric features is enhanced, the precision and robustness of three-dimensional target detection in a complex scene are remarkably improved, and the method has important application value in the fields of automatic driving, robot navigation and the like.

Description

Technical Field

[0001] This invention relates to the fields of point cloud data processing, artificial intelligence, and computer vision, and specifically to a multi-stage three-dimensional target detection method, system, device, and readable storage medium that integrates point cloud transformers and spatial feature enhancement. Background Technology

[0002] Deep learning technology has demonstrated significant performance in 2D computer vision, including tasks such as object detection and instance segmentation. Compared to 2D scene understanding, 3D object detection plays a crucial role in many real-world applications such as autonomous driving and home robots. Although 2D detection algorithms developed in recent years can handle issues such as varying viewpoints and complex backgrounds in images well, point cloud-based 3D object detection still faces challenges such as irregular data formats and the enormous six-degree-of-freedom search space for 3D objects.

[0003] In applications such as autonomous driving, LiDAR (Light Detection and Ranging) is one of the most commonly used 3D sensors, capturing the 3D structure of a scene by generating 3D point clouds. The main challenge in point cloud-based 3D object detection lies in the inherent irregularity of point cloud data. Current mainstream 3D detection methods typically project the point cloud onto a bird's-eye view or front view, or convert it into a regular 3D voxel structure to draw upon established 2D detection frameworks. However, these methods inevitably suffer from information loss during the quantization process, and therefore are not optimal solutions.

[0004] To avoid the information loss associated with converting point clouds into voxels or other regularized structures for feature learning, researchers have proposed methods that can directly learn 3D representations from raw point cloud data, such as PointNet and its successors. These methods have been applied to point cloud classification, segmentation, and even 3D object detection tasks, the latter often using frustum point clouds cropped from 2D image detection results to estimate 3D bounding boxes. However, the performance of such methods largely depends on the quality of 2D detection and fails to fully utilize 3D information to generate robust bounding box proposals.

[0005] Unlike object detection based on 2D images, in scenarios such as autonomous driving, 3D objects are naturally defined by labeled 3D bounding boxes. This means that training data for 3D object detection directly provides semantic masks that can be used for 3D object segmentation, which is a key difference between 3D and 2D object detection training data. In 2D object detection tasks, bounding boxes typically only provide relatively weak supervision information for semantic segmentation.

[0006] Based on the above observations, we propose a multi-stage 3D target detection method, system, device, and readable storage medium that integrates point cloud transformer and spatial feature enhancement, referencing the PointRCNN framework. Summary of the Invention

[0007] In view of this, the present invention provides a multi-stage 3D target detection method, system, device and readable storage medium that integrates point cloud transformer and spatial feature enhancement. It can effectively solve the problems of information loss caused by converting point clouds into voxels or other regularized structures for feature learning in existing 3D target detection methods, as well as the problems of over-reliance on the quality of 2D detection and failure to fully utilize 3D information to generate robust bounding box proposals.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A multi-stage 3D target detection method integrating point cloud transformer and spatial feature enhancement includes the following steps:

[0010] S1. Acquire raw point cloud data of the outdoor environment based on distance sensing sensors;

[0011] S2. Perform denoising and downsampling processing on the acquired raw point cloud data to obtain updated point cloud data;

[0012] S3. Construct a multi-stage point cloud detection model based on PointTransformerV3 and PnP 3D, and train the multi-stage point cloud detection model to obtain the trained point cloud detection model;

[0013] S4. Use the trained point cloud detection model to detect the updated point cloud data and obtain the point cloud detection results;

[0014] The construction of the multi-stage point cloud detection model in S3 includes: the first stage of constructing the point cloud detection model and the second stage of constructing the point cloud detection model;

[0015] The first stage of constructing the point cloud detection model includes: extracting point cloud features through point cloud serialization and serialized attention mechanism, and generating three-dimensional candidate region features;

[0016] The second stage of constructing the point cloud detection model includes:

[0017] For the acquired 3D candidate region features, a local context fusion and global bilinear regularization mechanism is introduced. Through a hierarchical feature fusion strategy of point-neighborhood-patch, multi-scale spatial relationship expression data is generated. Based on the multi-scale spatial relationship expression data, the final 3D target detection result is output.

[0018] Preferably, the first stage of constructing the point cloud detection model includes:

[0019] The initialization module, encoding module, decoding module, and candidate box generation module are constructed sequentially.

[0020] The initialization module is used to receive raw point cloud data and sequentially perform serialization and embedding processing on the received raw point cloud data to obtain the initial feature vector of the point cloud for encoding.

[0021] Serialization processing includes: serializing the point cloud into a fixed-length representation;

[0022] Embedding processing includes: constructing initial feature vectors for the point cloud through linear transformations and activation functions;

[0023] The encoding module is used to generate high-dimensional feature representation data that combines global context awareness and local detail resolution capabilities. It includes: a raster pooling submodule, a sequential perturbation submodule, and a feature enhancement submodule connected in sequence.

[0024] Among them, the grid pooling submodule is used to divide fixed units in three-dimensional space and aggregate local point cloud features to generate neighborhood geometric features of the point cloud.

[0025] The sequence perturbation submodule is used to shuffle the initial arrangement order of the point cloud feature vectors;

[0026] The feature enhancement submodule is used to obtain long-distance dependencies between points in the point cloud; it includes multiple first-level feature enhancement submodules.

[0027] Each first-level feature enhancement submodule is used to obtain the dynamic spatial location information related to its own features for each point, including: an enhanced conditional location coding module, a layer normalization module, a self-attention module, a layer normalization module, and a multilayer perceptron module connected in sequence.

[0028] The encoding modules are stacked in S repetitions, and the first-level feature enhancement sub-modules are stacked in N repetitions, where S and N are positive integers.

[0029] The decoding module is used to generate three-dimensional candidate region features based on high-dimensional feature representation data. It includes several reverse-order structure units, which are stacked sequentially from top to bottom in stages.

[0030] Each reverse-order structural unit includes an upsampling module and multiple decoding sub-blocks in sequence;

[0031] The decoding sub-block has the same structure as the first-level feature enhancement sub-module in the encoding module;

[0032] The candidate box generation module is used to generate target localization and boundary prediction. Based on the target localization and boundary prediction, it generates 3D candidate region features. The candidate box generation module includes a 3D box generation submodule based on interval regression and a foreground point segmentation submodule.

[0033] The 3D frame generation submodule and the foreground point segmentation submodule are set up in parallel in the first stage structure;

[0034] The 3D bounding box generation submodule adopts a bucket regression structure; it performs feature transformation and prediction through multiple convolutional layers and fully connected layers, and outputs the center position and size information of the candidate bounding boxes.

[0035] The foreground point segmentation submodule performs binary classification prediction of foreground probability for each point in the point cloud through a multi-layer point feature transformation network.

[0036] Preferably, the foreground / background classification loss formula of the foreground segmentation submodule is L. cls =

[0037] -α t (1-p t ) γ log(p t );

[0038] Where L cls Let t be the foreground / background classification loss, and p be the ground truth label. t Let α be the probability that the model predicts the true class t. t γ is the category weight adjustment factor, usually set to 0.25, and γ is the focusing parameter, usually set to 2.

[0039] Preferably, the second stage of constructing the point cloud detection model includes:

[0040] Construct a point cloud region pooling module, a feature fusion module, a feature encoding module, and a final output module that are connected in sequence.

[0041] The point cloud region pooling module is used to generate a corresponding customized feature representation based on the features of each 3D candidate region, and to generate corresponding candidate region information by combining the corresponding customized feature representation. The point cloud region pooling module is connected in series with the candidate box generation module. The point cloud region pooling module includes an input port, a processing unit, an aggregation unit and an output port connected in sequence.

[0042] The input port is used to receive candidate region information and raw point cloud data;

[0043] The processing unit is used to define a subset of point cloud within each candidate region through geometric operations based on candidate region information and original point cloud data, and generate multiple point spatial coordinates and corresponding semantic features.

[0044] The aggregation unit is used to accommodate and organize multiple point spatial coordinates and corresponding semantic features to generate local point cloud information within each candidate region. The local point cloud information includes point spatial information and semantic information.

[0045] The output port is used to output the local point cloud information of each candidate region;

[0046] The feature fusion module includes a coordinate transformation submodule and a multilayer perceptron submodule;

[0047] The coordinate transformation submodule is used to receive local point cloud information within the candidate region, perform spatial standardization processing on the input local point cloud information, and output the transformed local point set.

[0048] The multilayer perceptron submodule is connected to the coordinate transformation submodule. The multilayer perceptron submodule is used to receive the transformed local point set and regional semantic feature information and generate a fused feature representation.

[0049] The feature encoding module is constructed based on a point cloud feature extraction network.

[0050] The point cloud feature extraction network adopts the PointNet++ architecture;

[0051] The PointNet++ structure includes: multiple feature abstraction sub-modules stacked in hierarchical order, each feature abstraction sub-module having a Set Abstraction layer and a spatial feature enhancement sub-module connected in sequence, and a single Set Abstraction layer connected after the multiple feature abstraction sub-modules stacked in hierarchical order.

[0052] The spatial feature enhancement submodule adopts a PnP 3D structure, which connects the local context fusion submodule and the global bilinear regularization submodule to generate multi-scale spatial relationship expression data based on the fused feature representation;

[0053] The final output module includes a 3D bounding box fine-tuning submodule based on interval division and a confidence prediction submodule;

[0054] The interval-based 3D bounding box fine-tuning submodule includes: several cascaded interval geometric parameter processing components and residual correction components;

[0055] The confidence prediction submodule includes: a set of feature mapping components and a scalar output component;

[0056] Among them, the three-dimensional bounding box fine-tuning submodule and the confidence prediction submodule based on interval division are both configured with independent input components and output components;

[0057] The input component is used to receive multi-scale spatial relationship representation data generated by the feature encoding module, and the output component is used to output the geometric parameter set and scalar value sequence.

[0058] Preferably, the obtained trained point cloud detection model includes:

[0059] Obtain the labeled public dataset KITTI and randomly sample 20% of the data.

[0060] The data preprocessing pipeline built from the input data samples generates a normalized input tensor;

[0061] The total loss scalar is generated by combining the standardized input tensor and the weighted loss function module.

[0062] The weights of the point cloud detection model are updated by combining the total loss scalar and using the AdamW optimizer and cosine annealing strategy;

[0063] After 80 training cycles, training is stopped and the corresponding point cloud detection model is saved to generate the trained point cloud detection model.

[0064] Preferably, the data preprocessing pipeline includes: a point cloud file loading unit, a 3D annotation loading unit, a point cloud range filtering component, a target range filtering component, a target sampling unit, a 3D random flipping component, a target noise addition unit, a global rotation, scaling and translation unit, a second point cloud range filtering component, a point cloud sampling component, a point cloud disorder component, and a final 3D detection input encapsulation unit;

[0065] Preferably, the weighted loss function module includes:

[0066] The loss function input port, multiple parallel loss calculation paths, central summing unit, and loss function output port are connected in sequence.

[0067] The input port of the loss function is used to converge the network's predicted values ​​and the true labels;

[0068] The loss calculation path includes: sequentially connected independent loss metric units and weight application components;

[0069] Independent loss metric units are used to generate the corresponding loss values;

[0070] The weighting component has a built-in preset scalar weight value, which is used to adjust the corresponding loss value generated by the independent loss metric unit to generate an independent loss value;

[0071] The central summing unit is used to accumulate all weighted independent loss values ​​to generate a total loss scalar;

[0072] The output port of the loss function is used to pass gradient information to the optimizer.

[0073] A multi-stage 3D target detection system integrating a point cloud transformer and spatial feature enhancement, applying the aforementioned multi-stage 3D target detection method integrating point cloud transformer and spatial feature enhancement, includes the following sequentially connected components:

[0074] The system includes an acquisition unit, a preprocessing unit, a model building unit, and a point cloud detection unit.

[0075] An acquisition unit is used to acquire point cloud data, which is obtained by sensor devices collecting data from the target scene.

[0076] The preprocessing unit is used to perform preprocessing on the point cloud data, such as noise filtering and downsampling, to obtain updated point cloud data;

[0077] The model building unit is used to build a multi-stage 3D target detection model that integrates point cloud transformer and spatial feature enhancement, and to train the point cloud detection model to obtain the trained point cloud detection model.

[0078] The point cloud detection unit is used to detect the updated point cloud data using the trained point cloud detection model to obtain point cloud detection results.

[0079] The model building unit includes: a model building sub-unit and a model training sub-unit;

[0080] The model building sub-units include: initialization module, encoding module, decoding module, candidate box generation module, point cloud region pooling module, feature fusion module, feature encoding module, and final output module;

[0081] The model training subunit includes: acquisition module, preprocessing module, function construction module, training module, and logic module.

[0082] A multi-stage 3D target detection device integrating a point cloud transformer and spatial feature enhancement, realizing the aforementioned multi-stage 3D target detection method integrating a point cloud transformer and spatial feature enhancement, includes: an I / O interface; a communication component; a multimedia component; a processor; and a memory storing a computer program.

[0083] A readable storage medium for multi-stage 3D target detection that integrates a point cloud transformer and spatial feature enhancement is provided. The medium stores computer program instructions, which, when executed by a processor of the device where the storage medium is located, cause the processor to execute the aforementioned multi-stage 3D target detection method that integrates a point cloud transformer and spatial feature enhancement.

[0084] As can be seen from the above technical solution, compared with the prior art, the multi-stage three-dimensional target detection method, system, device, and readable storage medium of the present invention, which integrates point cloud transformer and spatial feature enhancement, has the following beneficial effects:

[0085] To address the information loss caused by converting point clouds into voxels or other regularized structures for feature learning, and the problem of over-reliance on 2D detection quality and failure to fully utilize 3D information to generate robust bounding box proposals, this paper proposes a multi-stage 3D object detection algorithm that integrates point cloud transformers and spatial feature enhancement. By introducing the PointTransformerV3 backbone network and the PnP 3D feature enhancement module, efficient point cloud feature extraction is achieved using point cloud serialization and serialization attention mechanisms. Furthermore, local context fusion and global bilinear regularization mechanisms enhance the expressive power of local spatial geometric features. Experiments on the KITTI dataset demonstrate that this method significantly improves the average detection accuracy on car, pedestrian, and bicycle datasets, especially in the case of sparse points with small targets, further confirming the effectiveness of the algorithm.

[0086] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0087] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0088] Figure 1 This is a schematic diagram of the process of a multi-stage three-dimensional target detection method that integrates point cloud transformer and spatial feature enhancement as described in an embodiment of the present invention;

[0089] Figure 2 This is a schematic diagram of the detection model in a multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement, as described in an embodiment of the present invention.

[0090] Figure 3 This is a schematic diagram of the encoding module structure of the detection model in a multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement, as described in an embodiment of the present invention.

[0091] Figure 4 This is a schematic diagram of the decoding module structure of the detection model in a multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement, as described in an embodiment of the present invention.

[0092] Figure 5This is a schematic diagram of the feature encoding module structure of the detection model in a multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement, as described in an embodiment of the present invention.

[0093] Figure 6 This is a schematic diagram of the spatial feature enhancement submodule structure of the detection model in a multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement, as described in this embodiment of the invention.

[0094] Figure 7 This is a schematic diagram of the structure of a multi-stage three-dimensional target detection device that integrates a point cloud transformer and spatial feature enhancement, as described in an embodiment of the present invention.

[0095] Figure 8 This is a schematic diagram of a multi-stage 3D target detection device that integrates a point cloud transformer and spatial feature enhancement, as described in an embodiment of the present invention; wherein, 500 is a multi-stage 3D target detection device that integrates a point cloud transformer and spatial feature enhancement; 501 is an IO interface; 502 is a communication component; 503 is a processor; 504 is a memory; and 505 is a multimedia component. Detailed Implementation

[0096] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0097] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0098] This invention can be used in a wide variety of general-purpose or special-purpose computing environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor devices, distributed computing environments including any of the above devices, etc.

[0099] Reference Figure 1 As shown, this invention discloses a multi-stage 3D target detection method that integrates a point cloud transformer and spatial feature enhancement, comprising the following steps:

[0100] S1. Acquire raw point cloud data of the outdoor environment based on distance sensing sensors;

[0101] S2. Perform denoising and downsampling processing on the acquired raw point cloud data to obtain updated point cloud data;

[0102] S3. Construct a multi-stage point cloud detection model based on PointTransformerV3 and PnP 3D, and train the multi-stage point cloud detection model to obtain the trained point cloud detection model;

[0103] S4. Use the trained point cloud detection model to detect the updated point cloud data and obtain the point cloud detection results;

[0104] The construction of the multi-stage point cloud detection model in S3 includes: the first stage of constructing the point cloud detection model and the second stage of constructing the point cloud detection model;

[0105] The first stage of constructing the point cloud detection model includes: extracting point cloud features through point cloud serialization and serialized attention mechanism, and generating three-dimensional candidate region features;

[0106] The second stage of constructing the point cloud detection model includes:

[0107] For the acquired 3D candidate region features, a local context fusion and global bilinear regularization mechanism is introduced. Through a hierarchical feature fusion strategy of point-neighborhood-patch, multi-scale spatial relationship expression data is generated. Based on the multi-scale spatial relationship expression data, the final 3D target detection result is output.

[0108] Preferably, the first stage of constructing the point cloud detection model includes:

[0109] The initialization module, encoding module, decoding module, and candidate box generation module are constructed sequentially.

[0110] The initialization module is used to receive raw point cloud data and sequentially perform serialization and embedding processing on the received raw point cloud data to obtain the initial feature vector of the point cloud for encoding.

[0111] Serialization processing includes: serializing the point cloud into a fixed-length representation;

[0112] Embedding processing includes: constructing initial feature vectors for the point cloud through linear transformations and activation functions;

[0113] Furthermore, the initialization module receives the original point cloud data and sequentially processes it through serialization and embedding to obtain feature embedding information for encoding. The serialization process serializes the point cloud into a fixed-length representation, and the embedding process constructs the initial features of the point cloud through linear transformation and activation function.

[0114] The core design of the initialization module lies in solving the problems of disorder and unstructured raw point cloud data. Through serialization, raw point cloud data of different scales are unified into a standardized fixed-length representation, thus enabling compatibility with mainstream encoder architectures. Embedding further maps the original spatial coordinates of points to a higher-dimensional feature space. Through learned linear transformations and nonlinear activations, the feature representation capability of a single point is greatly enriched. This transformation from geometric information to feature information allows the model to capture more discernible local geometric patterns in the initial stage, providing high-quality input for the subsequent network to perform accurate and efficient object detection tasks.

[0115] like Figure 3 As shown, the encoding module is used to generate high-dimensional feature representation data that combines global context awareness and local detail resolution capabilities, including: a raster pooling submodule, a sequential perturbation submodule, and a feature enhancement submodule connected in sequence;

[0116] Among them, the grid pooling submodule is used to divide fixed units in three-dimensional space and aggregate local point cloud features to generate neighborhood geometric features of the point cloud.

[0117] The sequence perturbation submodule is used to shuffle the initial arrangement order of the point cloud feature vectors;

[0118] The feature enhancement submodule is used to obtain long-distance dependencies between points in the point cloud; it includes multiple first-level feature enhancement submodules.

[0119] Each first-level feature enhancement submodule is used to obtain the dynamic spatial location information related to its own features for each point, including: an enhanced conditional location coding module, a layer normalization module, a self-attention module, a layer normalization module, and a multilayer perceptron module connected in sequence.

[0120] The encoding modules are stacked in S repetitions, and the first-level feature enhancement sub-modules are stacked in N repetitions, where S and N are positive integers.

[0121] Furthermore, it's worth noting that the raster pooling submodule first structurally regularizes the discrete point cloud, effectively capturing geometric features within the microscopic neighborhood through local feature aggregation, and significantly reducing the computational complexity of subsequent processing. The sequence perturbation submodule, as a data augmentation strategy, disrupts the inherent order of the input sequence, forcing the network to learn permutation invariance to the disorder of the point cloud, significantly improving the model's generalization ability and robustness. The core feature enhancement submodule adopts a Transformer-like architecture. Its self-attention mechanism can establish long-distance dependencies between points, thus transcending the limitations of local receptive fields and understanding the scene from a global perspective. Enhanced conditional position encoding provides each point with dynamic spatial location information related to its own features, ensuring that the model does not lose crucial absolute and relative positional awareness when modeling global relationships, ultimately generating a high-dimensional feature representation that combines global context awareness with local detail resolution.

[0122] like Figure 4 As shown, the decoding module is used to generate three-dimensional candidate region features based on high-dimensional feature expression data. It includes several reverse-order structure units, which are stacked sequentially from top to bottom in stages.

[0123] Each reverse-order structural unit includes an upsampling module and multiple decoding sub-blocks in sequence;

[0124] The decoding sub-block has the same structure as the first-level feature enhancement sub-module in the encoding module;

[0125] Furthermore, the decoding module adopts a top-down hierarchical architecture symmetrical to the encoding module, fusing with low-level features from the encoding process. This design can progressively recover the spatial resolution and detail information of the point cloud, thereby achieving refined target localization and boundary prediction while preserving rich context awareness.

[0126] The candidate box generation module is used to generate target localization and boundary prediction. Based on the target localization and boundary prediction, it generates 3D candidate region features. The candidate box generation module includes a 3D box generation submodule based on interval regression and a foreground point segmentation submodule.

[0127] The 3D frame generation submodule and the foreground point segmentation submodule are set up in parallel in the first stage structure;

[0128] The 3D bounding box generation submodule adopts a bucket regression structure; it performs feature transformation and prediction through multiple convolutional layers and fully connected layers, and outputs the center position and size information of the candidate bounding boxes.

[0129] The foreground point segmentation submodule performs binary classification prediction of foreground probability for each point in the point cloud through a multi-layer point feature transformation network.

[0130] Furthermore, the candidate box generation module adopts a parallel dual-task design.

[0131] A submodule for generating 3D boxes based on interval regression is constructed to efficiently generate high-quality 3D target proposals. The bucketed regression structure transforms the difficult direct regression problem into a more stable classification and residual correction task.

[0132] A site segmentation submodule is constructed, and parallel foreground segmentation provides pixel-level precise supervision for candidate box generation, which significantly improves the localization accuracy and recall rate of the initial candidate boxes.

[0133] Preferably, the foreground / background classification loss formula of the foreground segmentation submodule is L. cls =

[0134] -α t (1-p t ) γ log(p t );

[0135] Where L cls Let t be the foreground / background classification loss, and p be the ground truth label. t Let α be the probability that the model predicts the true class t. t γ is the category weight adjustment factor, usually set to 0.25, and γ is the focusing parameter, usually set to 2.

[0136] Preferably, the second stage of constructing the point cloud detection model includes:

[0137] Construct a point cloud region pooling module, a feature fusion module, a feature encoding module, and a final output module that are connected in sequence.

[0138] The point cloud region pooling module is used to generate a corresponding customized feature representation based on the features of each 3D candidate region, and to generate corresponding candidate region information by combining the corresponding customized feature representation. The point cloud region pooling module is connected in series with the candidate box generation module. The point cloud region pooling module includes an input port, a processing unit, an aggregation unit and an output port connected in sequence.

[0139] The input port is used to receive candidate region information and raw point cloud data;

[0140] The processing unit is used to define a subset of point cloud within each candidate region through geometric operations based on candidate region information and original point cloud data, and generate multiple point spatial coordinates and corresponding semantic features.

[0141] The aggregation unit is used to accommodate and organize multiple point spatial coordinates and corresponding semantic features to generate local point cloud information within each candidate region. The local point cloud information includes point spatial information and semantic information.

[0142] The output port is used to output the local point cloud information of each candidate region;

[0143] Furthermore, the point cloud region pooling module extracts customized feature representations for each candidate region, accurately aggregates the point cloud subsets within each candidate box through geometric constraints, and aggregates them into regular feature vectors. This transforms unstructured regional point clouds into structured inputs that can be processed by subsequent networks, providing crucial and discriminative regional-level information for the final target classification and location refinement.

[0144] The feature fusion module includes a coordinate transformation submodule and a multilayer perceptron submodule;

[0145] The coordinate transformation submodule is used to receive local point cloud information within the candidate region, perform spatial standardization processing on the input local point cloud information, and output the transformed local point set.

[0146] The multilayer perceptron submodule is connected to the coordinate transformation submodule. The multilayer perceptron submodule is used to receive the transformed local point set and regional semantic feature information and generate a fused feature representation.

[0147] Furthermore, in order to enable the fused features to have stronger expressive and discriminative capabilities, the feature fusion module is structurally divided into two complementary sub-modules: a coordinate transformation sub-module and a multilayer perceptron sub-module.

[0148] The coordinate transformation submodule transforms the local point cloud to the standard coordinate system, eliminating feature differences caused by translation or rotation of objects.

[0149] The multilayer perceptron submodule nonlinearly fuses standardized geometric information with existing semantic features to generate an enhanced feature that is insensitive to object pose while possessing both geometric details and semantic information.

[0150] Feature encoding module such as Figure 5 As shown, a feature encoding module is constructed based on a point cloud feature extraction network;

[0151] Furthermore, the feature encoding module is significantly enhanced on top of PointNet++'s hierarchical feature extraction framework through a cascaded structure. It not only inherits PointNet++'s powerful ability to capture local geometric structures, but also significantly improves the model's understanding of multi-scale spatial relationships by fusing local context and global bilinear regularization, thereby generating feature representations that are more robust to scene geometric changes and have stronger discriminative power.

[0152] The spatial feature enhancement submodule is described in detail below. Figure 6 .

[0153] The point cloud feature extraction network adopts the PointNet++ architecture;

[0154] The PointNet++ structure includes: multiple feature abstraction sub-modules stacked in hierarchical order, each feature abstraction sub-module having a Set Abstraction layer and a spatial feature enhancement sub-module connected in sequence, and a single Set Abstraction layer connected after the multiple feature abstraction sub-modules stacked in hierarchical order.

[0155] The spatial feature enhancement submodule adopts a PnP 3D structure, which connects the local context fusion submodule and the global bilinear regularization submodule to generate multi-scale spatial relationship expression data based on the fused feature representation.

[0156] The final output module includes a 3D bounding box fine-tuning submodule based on interval division and a confidence prediction submodule;

[0157] The interval-based 3D bounding box fine-tuning submodule includes: several cascaded interval geometric parameter processing components and residual correction components;

[0158] The confidence prediction submodule includes: a set of feature mapping components and a scalar output component;

[0159] Among them, the three-dimensional bounding box fine-tuning submodule and the confidence prediction submodule based on interval division are both configured with independent input components and output components;

[0160] The input component is used to receive multi-scale spatial relationship representation data generated by the feature encoding module, and the output component is used to output the geometric parameter set and scalar value sequence.

[0161] Furthermore, a 3D bounding box fine-tuning submodule based on interval partitioning is constructed, which decomposes the complex 3D bounding box direct regression problem into two-stage tasks of interval classification and residual correction, which are easier to converge.

[0162] A confidence prediction submodule is constructed to provide a reliable quality assessment for each output box, ensuring that the final detection result has high positioning accuracy while effectively suppressing false detections.

[0163] Preferably, the obtained trained point cloud detection model includes:

[0164] Obtain the labeled public dataset KITTI and randomly sample 20% of the data.

[0165] It should be noted that this invention selects 20% of the KITTI dataset as training data. This proportion ensures sufficient data to support model learning while reducing training and debugging costs to some extent and improving experimental efficiency.

[0166] The data preprocessing pipeline built from the input data samples generates a normalized input tensor;

[0167] The total loss scalar is generated by combining the standardized input tensor and the weighted loss function module.

[0168] The weights of the point cloud detection model are updated by combining the total loss scalar and using the AdamW optimizer and cosine annealing strategy;

[0169] After 80 training cycles, training is stopped and the corresponding point cloud detection model is saved to generate the trained point cloud detection model.

[0170] Preferably, the data preprocessing pipeline includes: a point cloud file loading unit, a 3D annotation loading unit, a point cloud range filtering component, a target range filtering component, a target sampling unit, a 3D random flipping component, a target noise addition unit, a global rotation, scaling and translation unit, a second point cloud range filtering component, a point cloud sampling component, a point cloud disorder component, and a final 3D detection input encapsulation unit;

[0171] Furthermore, the data preprocessing pipeline systematically transforms the raw, sparse, and unstructured point cloud data into standardized input tensors suitable for training deep learning models. This pipeline enriches the diversity of training samples through a series of cascaded data augmentation operations, effectively suppressing model overfitting and significantly improving the model's generalization ability and robustness to complex and varied driving scenarios in the real world. Moreover, by introducing target sampling and multi-stage filtering and downsampling components, it addresses the problem of imbalanced target classes in the dataset, ensuring the efficiency and consistency of the input data.

[0172] Among them, deep learning models are used to accurately identify and locate targets in the scene, enabling effective detection of key objects and improving the system's automatic perception capability in complex environments.

[0173] Preferably, the weighted loss function module includes:

[0174] The loss function input port, multiple parallel loss calculation paths, central summing unit, and loss function output port are connected in sequence.

[0175] The input port of the loss function is used to converge the network's predicted values ​​and the true labels;

[0176] The loss calculation path includes: sequentially connected independent loss metric units and weight application components;

[0177] Independent loss metric units are used to generate the corresponding loss values;

[0178] The weighting component has a built-in preset scalar weight value, which is used to adjust the corresponding loss value generated by the independent loss metric unit to generate an independent loss value;

[0179] The central summing unit is used to accumulate all weighted independent loss values ​​to generate a total loss scalar;

[0180] The output port of the loss function is used to pass gradient information to the optimizer.

[0181] It should be noted that the weighted loss function is specifically L. total =λ1L cls1 +λ2L reg1 +λ3L seg +λ4L cls2 +λ5L reg2 L cls1 L represents the cross-entropy of the first-stage binary classification. reg1 L represents the regression loss of the first-stage 3D candidate boxes. seg L represents the segmentation loss. cls2 L represents the second-stage classification loss. reg2 λ represents the regression loss of the 3D candidate boxes in the second stage, and λ1 to λ5 represent the weighting coefficients for adjusting each loss term.

[0182] and Figure 1 Correspondingly, this embodiment also discloses a multi-stage 3D target detection system that integrates a point cloud transformer and spatial feature enhancement, such as... Figure 7 As shown, the multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement described above includes the following sequentially connected components:

[0183] The system includes an acquisition unit, a preprocessing unit, a model building unit, and a point cloud detection unit.

[0184] An acquisition unit is used to acquire point cloud data, which is obtained by sensor devices collecting data from the target scene.

[0185] The preprocessing unit is used to perform preprocessing on the point cloud data, such as noise filtering and downsampling, to obtain updated point cloud data;

[0186] The model building unit is used to build a multi-stage 3D target detection model that integrates point cloud transformer and spatial feature enhancement, and to train the point cloud detection model to obtain the trained point cloud detection model.

[0187] The point cloud detection unit is used to detect the updated point cloud data using the trained point cloud detection model to obtain point cloud detection results.

[0188] The model building unit includes: a model building sub-unit and a model training sub-unit;

[0189] The model building sub-units include: initialization module, encoding module, decoding module, candidate box generation module, point cloud region pooling module, feature fusion module, feature encoding module, and final output module;

[0190] The model training subunit includes: acquisition module, preprocessing module, function construction module, training module, and logic module.

[0191] and Figure 1 Correspondingly, this embodiment also discloses a multi-stage three-dimensional target detection device that integrates a point cloud transformer and spatial feature enhancement, such as... Figure 8 As shown, the multi-stage 3D target detection method that integrates point cloud transformer and spatial feature enhancement described above includes: an I / O interface; a communication component; a multimedia component; a processor; and a memory storing a computer program.

[0192] A framework structure for a multi-stage 3D target detection device 500 that integrates a point cloud transformer and spatial feature enhancement. This multi-stage 3D target detection device 500 includes, but is not limited to, one or more of the following: an I / O interface 501, a communication component 502, a processor 503, a memory 504, and a multimedia component 505. These components work collaboratively to support the implementation of a multi-stage 3D target detection method that integrates a point cloud transformer and spatial feature enhancement.

[0193] In the data processing flow, the communication component 502 first completes the data exchange task between the device 500 and external systems or sensors. The communication component can support a variety of communication methods, including but not limited to Wi-Fi, Bluetooth, Near Field Communication (NFC), and cellular mobile communication, and therefore can be configured with corresponding wireless communication modules to adapt to the above communication protocols.

[0194] The acquired external input data can be stored in memory 504. This memory is used to persistently store various data and application instructions required for device operation, including control logic for executing target detection algorithms, intermediate processing results, and related auxiliary data (such as images, audio, communication records, user data, etc.). The memory can be constructed using any one or a combination of storage media such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, read-only memory, flash memory, disk drive, or optical disk.

[0195] The multimedia component 505 plays a role in data acquisition and human-computer interaction. The display module may include a touchscreen for information display and operation response; the audio module may include a microphone and speaker for acquiring external audio signals and providing feedback on processing results. The acquired audio can be stored in the memory 504 or transmitted to other terminals via the communication component 502.

[0196] Subsequently, the processor 503 reads and schedules the program instructions stored in the memory 504 to execute a multi-stage point cloud target detection task, including point cloud preprocessing, feature extraction, candidate box generation, spatial feature enhancement, and generation of the final detection result. The processor 503 can be implemented using hardware resources such as application-specific integrated circuits, digital signal processors, digital signal processing equipment, programmable logic devices, field-programmable gate arrays, controllers, microcontrollers, or microprocessors, and has efficient neural network inference capabilities.

[0197] The IO interface 501 serves as a data channel between the processor and external interactive devices, and can connect to various input methods such as physical buttons, virtual buttons, keyboards, and mice, ensuring the operational flexibility and compatibility of the device.

[0198] In another exemplary embodiment, the present invention also provides a computer-readable storage medium pre-stored thereon with program code executable by processor 503 to complete all the steps of the aforementioned multi-stage 3D target detection method integrating point cloud transformer and spatial feature enhancement. This medium may include the aforementioned memory 504, or other independent storage devices such as USB flash drives, solid-state drives, SD cards, etc., suitable for distributing and deploying related detection systems.

[0199] A readable storage medium for multi-stage 3D target detection that integrates a point cloud transformer and spatial feature enhancement is provided. The medium stores computer program instructions, which, when executed by a processor of the device where the storage medium is located, cause the processor to execute the aforementioned multi-stage 3D target detection method that integrates a point cloud transformer and spatial feature enhancement.

[0200] Another aspect of the present invention provides a computer-readable storage medium for implementing a technical solution corresponding to the aforementioned multi-stage 3D target detection method that integrates a point cloud transformer and spatial feature enhancement. The computer-readable storage medium pre-stores a set of computer program instructions that can be called and executed by a processor. When executed by the processor, these instructions can complete the various processing flows described in the above method embodiments, including but not limited to point cloud data acquisition, preprocessing, feature extraction, spatial feature enhancement, and target recognition.

[0201] The physical form of the computer-readable storage medium can be various types of storage devices, including but not limited to: USB flash drives, portable hard disk drives, optical storage media, disks, solid-state storage devices, or read-only memory, random access memory, etc. inside a computer, used to store program code or data that can be read by a computing device in a non-transitory manner.

[0202] By storing the program instructions in a computer-readable storage medium and loading and running them by a processor in a computing device, the multi-stage three-dimensional target detection method proposed in this invention, which integrates point cloud transformers and spatial feature enhancement, can be executed to achieve efficient target detection in complex three-dimensional scenes.

[0203] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0204] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both.

[0205] To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality above. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0206] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-stage 3D target detection method integrating point cloud transformer and spatial feature enhancement, characterized in that, Includes the following steps: S1. Acquire raw point cloud data of the outdoor environment based on distance sensing sensors; S2. Perform denoising and downsampling processing on the acquired raw point cloud data to obtain updated point cloud data; S3. Construct a multi-stage point cloud detection model based on PointTransformerV3 and PnP 3D, and train the multi-stage point cloud detection model to obtain the trained point cloud detection model; S4. Use the trained point cloud detection model to detect the updated point cloud data and obtain the point cloud detection results; The construction of the multi-stage point cloud detection model in S3 includes: the first stage of constructing the point cloud detection model and the second stage of constructing the point cloud detection model; The first stage of constructing the point cloud detection model includes: extracting point cloud features through point cloud serialization and serialized attention mechanism, and generating three-dimensional candidate region features; The second stage of constructing the point cloud detection model includes: For the acquired 3D candidate region features, a local context fusion and global bilinear regularization mechanism is introduced. Through a hierarchical feature fusion strategy of point-neighborhood-patch, multi-scale spatial relationship expression data is generated. Based on the multi-scale spatial relationship expression data, the final 3D target detection result is output.

2. The multi-stage three-dimensional target detection method integrating point cloud transformer and spatial feature enhancement according to claim 1, characterized in that, The first stage of constructing the point cloud detection model includes: The initialization module, encoding module, decoding module, and candidate box generation module are constructed sequentially. The initialization module is used to receive raw point cloud data and sequentially perform serialization and embedding processing on the received raw point cloud data to obtain the initial feature vector of the point cloud for encoding. Serialization processing includes: serializing the point cloud into a fixed-length representation; Embedding processing includes: constructing initial feature vectors for the point cloud through linear transformations and activation functions; The encoding module is used to generate high-dimensional feature representation data that combines global context awareness and local detail resolution capabilities. It includes: a raster pooling submodule, a sequential perturbation submodule, and a feature enhancement submodule connected in sequence. Among them, the grid pooling submodule is used to divide fixed units in three-dimensional space and aggregate local point cloud features to generate neighborhood geometric features of the point cloud. The sequence perturbation submodule is used to shuffle the initial arrangement order of the point cloud feature vectors; The feature enhancement submodule is used to obtain long-distance dependencies between points in the point cloud; it includes multiple first-level feature enhancement submodules. Each first-level feature enhancement submodule is used to obtain the dynamic spatial location information related to its own features for each point, including: an enhanced conditional location coding module, a layer normalization module, a self-attention module, a layer normalization module, and a multilayer perceptron module connected in sequence. The encoding modules are stacked in S repetitions, and the first-level feature enhancement sub-modules are stacked in N repetitions, where S and N are positive integers. The decoding module is used to generate three-dimensional candidate region features based on high-dimensional feature representation data. It includes several reverse-order structure units, which are stacked sequentially from top to bottom in stages. Each reverse-order structural unit includes an upsampling module and multiple decoding sub-blocks in sequence; The decoding sub-block has the same structure as the first-level feature enhancement sub-module in the encoding module; The candidate box generation module is used to generate target localization and boundary prediction. Based on the target localization and boundary prediction, it generates 3D candidate region features. The candidate box generation module includes a 3D box generation submodule based on interval regression and a foreground point segmentation submodule. The 3D frame generation submodule and the foreground point segmentation submodule are set up in parallel in the first stage structure; The 3D bounding box generation submodule adopts a bucket regression structure; it performs feature transformation and prediction through multiple convolutional layers and fully connected layers, and outputs the center position and size information of the candidate bounding boxes. The foreground point segmentation submodule performs binary classification prediction of foreground probability for each point in the point cloud through a multi-layer point feature transformation network.

3. The multi-stage three-dimensional target detection method integrating point cloud transformer and spatial feature enhancement according to claim 2, characterized in that, The foreground point segmentation submodule includes: The foreground / background classification loss formula for the foreground segmentation submodule is L. cls =-α t (1-p t ) γ log(p t ); Where L cls The loss is for foreground / background classification, where t is the true label and p is the background label. t Let α be the probability that the model predicts the true class t. t γ is the category weight adjustment factor, usually set to 0.25, and γ is the focusing parameter, usually set to 2.

4. The multi-stage three-dimensional target detection method integrating point cloud transformer and spatial feature enhancement according to claim 2, characterized in that, The second stage of constructing the point cloud detection model includes: Construct a point cloud region pooling module, a feature fusion module, a feature encoding module, and a final output module that are connected in sequence. The point cloud region pooling module is used to generate a corresponding customized feature representation based on the features of each 3D candidate region, and to generate corresponding candidate region information by combining the corresponding customized feature representation. The point cloud region pooling module is connected in series with the candidate box generation module. The point cloud region pooling module includes an input port, a processing unit, an aggregation unit and an output port connected in sequence. The input port is used to receive candidate region information and raw point cloud data; The processing unit is used to define a subset of point cloud within each candidate region through geometric operations based on candidate region information and original point cloud data, and generate multiple point spatial coordinates and corresponding semantic features. The aggregation unit is used to accommodate and organize multiple point spatial coordinates and corresponding semantic features to generate local point cloud information within each candidate region. The local point cloud information includes point spatial information and semantic information. The output port is used to output the local point cloud information of each candidate region; The feature fusion module includes a coordinate transformation submodule and a multilayer perceptron submodule; The coordinate transformation submodule is used to receive local point cloud information within the candidate region, perform spatial standardization processing on the input local point cloud information, and output the transformed local point set. The multilayer perceptron submodule is connected to the coordinate transformation submodule. The multilayer perceptron submodule is used to receive the transformed local point set and regional semantic feature information and generate a fused feature representation. The feature encoding module is constructed based on a point cloud feature extraction network. The point cloud feature extraction network adopts the PointNet++ architecture; The PointNet++ structure includes: multiple feature abstraction sub-modules stacked in hierarchical order, each feature abstraction sub-module having a Set Abstraction layer and a spatial feature enhancement sub-module connected in sequence, and a single Set Abstraction layer connected after the multiple feature abstraction sub-modules stacked in hierarchical order. The spatial feature enhancement submodule adopts a PnP 3D structure, which connects the local context fusion submodule and the global bilinear regularization submodule to generate multi-scale spatial relationship expression data based on the fused feature representation; The final output module includes a 3D bounding box fine-tuning submodule based on interval division and a confidence prediction submodule; The interval-based 3D bounding box fine-tuning submodule includes: several cascaded interval geometric parameter processing components and residual correction components; The confidence prediction submodule includes: a set of feature mapping components and a scalar output component; Among them, the three-dimensional bounding box fine-tuning submodule and the confidence prediction submodule based on interval division are both configured with independent input components and output components; The input component is used to receive multi-scale spatial relationship representation data generated by the feature encoding module, and the output component is used to output the geometric parameter set and scalar value sequence.

5. The multi-stage three-dimensional target detection method integrating point cloud transformer and spatial feature enhancement according to claim 4, characterized in that, The trained point cloud detection model includes: Obtain the labeled public dataset KITTI and randomly select 20% of the data samples; The data preprocessing pipeline built from the input data samples generates a normalized input tensor; The total loss scalar is generated by combining the standardized input tensor and the weighted loss function module. The weights of the point cloud detection model are updated by combining the total loss scalar and using the AdamW optimizer and cosine annealing strategy; After 80 training cycles, training is stopped and the corresponding point cloud detection model is saved to generate the trained point cloud detection model.

6. The multi-stage three-dimensional target detection method integrating point cloud transformer and spatial feature enhancement according to claim 5, characterized in that, The data preprocessing pipeline includes: a point cloud file loading unit, a 3D annotation loading unit, a point cloud range filtering component, a target range filtering component, a target sampling unit, a 3D random flipping component, a target noise addition unit, a global rotation, scaling, and translation unit, a second point cloud range filtering component, a point cloud sampling component, a point cloud disordering component, and the final 3D detection input encapsulation unit.

7. The multi-stage three-dimensional target detection method integrating point cloud transformer and spatial feature enhancement according to claim 6, characterized in that, The weighted loss function module includes: The loss function input port, multiple parallel loss calculation paths, central summing unit, and loss function output port are connected in sequence. The input port of the loss function is used to merge the network's predicted values ​​with the true labels; The loss calculation path includes: sequentially connected independent loss metric units and weight application components; Independent loss metric units are used to generate the corresponding loss values; The weighting component has a built-in preset scalar weight value, which is used to adjust the corresponding loss value generated by the independent loss metric unit to generate an independent loss value; The central summing unit is used to accumulate all weighted independent loss values ​​to generate a total loss scalar; The output port of the loss function is used to pass gradient information to the optimizer.

8. A multi-stage three-dimensional target detection system integrating a point cloud transformer and spatial feature enhancement, characterized in that, A multi-stage 3D target detection method integrating a point cloud transformer and spatial feature enhancement as described in any one of claims 1-7, comprising the following sequentially connected components: The system includes an acquisition unit, a preprocessing unit, a model building unit, and a point cloud detection unit. An acquisition unit is used to acquire point cloud data, which is obtained by sensor devices collecting data from the target scene. The preprocessing unit is used to perform preprocessing on the point cloud data, such as noise filtering and downsampling, to obtain updated point cloud data; The model building unit is used to build a multi-stage 3D target detection model that integrates point cloud transformer and spatial feature enhancement, and to train the point cloud detection model to obtain the trained point cloud detection model. The point cloud detection unit is used to detect the updated point cloud data using the trained point cloud detection model to obtain point cloud detection results. The model building unit includes: a model building sub-unit and a model training sub-unit; The model building sub-units include: initialization module, encoding module, decoding module, candidate box generation module, point cloud region pooling module, feature fusion module, feature encoding module, and final output module; The model training subunit includes: acquisition module, preprocessing module, function construction module, training module, and logic module.

9. A multi-stage three-dimensional target detection device integrating a point cloud transformer and spatial feature enhancement, characterized in that, When the computer program is configured to be executed by the processor, it implements a multi-stage three-dimensional target detection method according to any one of claims 1-7, comprising: an I / O interface; a communication component; a multimedia component; a processor; and a memory storing the computer program thereon.

10. A readable storage medium for multi-stage three-dimensional target detection that integrates a point cloud transformer and spatial feature enhancement, characterized in that, The device stores computer program instructions, which, when executed by a processor of the device where the storage medium is located, cause the processor to perform a multi-stage three-dimensional target detection method that integrates a point cloud transformer and spatial feature enhancement as described in any one of claims 1-7.

Images