High adaptability point cloud 3D target detection method based on active sampling

Abstract

The application provides a high adaptability point cloud 3D target detection method based on active sampling, which has the following characteristics and comprises the following steps: step S1, preprocessing input point cloud to obtain preprocessed point cloud; step S2, inputting the preprocessed point cloud into a point cloud 3D target detection model to obtain multiple initial targets containing target 3D bounding boxes and target categories; and step S3, performing target confidence filtering and non-maximum suppression on all initial targets to obtain final targets, wherein the point cloud 3D target detection model comprises: a point cloud feature extraction module, which is used for extracting multiple point cloud features of different depth levels from the preprocessed point cloud; a candidate target prediction module, which is used for obtaining candidate targets according to the point cloud features of the deepest level; and a multi-scale cascaded optimization module, which is used for performing multi-layer cascaded optimization on the candidate targets to obtain multiple optimized targets corresponding to different levels. In summary, the method can realize accurate 3D target detection on point cloud.

Description

Technical Field

[0001] This invention relates to a 3D target detection method for point clouds, specifically a highly adaptive 3D target detection method for point clouds based on active sampling. Background Technology

[0002] Locating and identifying specific targets in the three-dimensional world is a major challenge in computer vision, autonomous driving, and robotics. Point clouds play a crucial role in 3D detection tasks due to their superior spatial information representation and scene feature perception capabilities.

[0003] Unlike 2D detection tasks, point clouds with elevation information are typically generated by LiDAR or 4D imaging millimeter-wave radar, directly corresponding to discrete and disordered points in a 3D scene. LiDAR produces a large number of fixed-resolution points with coordinates and reflection intensities, while 4D radar produces sparser points with more features, such as Doppler velocity, radar cross section (RCS), and signal-to-noise ratio (SNR). Therefore, 3D detection in point clouds with complex geometry, low resolution, and diverse styles remains a highly challenging task, especially for small objects.

[0004] Most neural network models struggle to directly handle unstructured point clouds. Early voxel-based work attempted to organize point clouds into regular representations, such as multi-view projected images, subsequently employing image-oriented feature extraction and object detection techniques. Many later works introduced more voxelization methods, including dense, structured 3D meshes and pillars. While these methods have recently achieved significant accuracy and efficiency on LiDAR point clouds, they inevitably lose original information during voxelization, potentially making them unsuitable for 4D Radar point clouds with rich original information. Meanwhile, given the emergence of various models designed to learn from raw point clouds, several methods have focused on directly generating 3D bounding boxes from discrete points. Hierarchical downsampling is a crucial module for effective learning from raw point clouds; in this process, sampling a small subset of points relevant to the target from a large-scale scene point cloud is essential for accurate detection.

[0005] Existing methods typically employ uniform sampling techniques and heuristic strategies focused on feature or semantic information to manually select points that may be relevant to the object. However, describing the most suitable detection sampling strategy using rule-based methods is challenging, especially when dealing with different types of point clouds. Summary of the Invention

[0006] This invention is made to solve the above-mentioned problems, and aims to provide a highly adaptive point cloud 3D target detection method based on active sampling.

[0007] This invention provides a highly adaptive point cloud 3D target detection method based on active sampling, used to obtain the corresponding final target from the input point cloud. It includes the following steps: Step S1, preprocessing the input point cloud to obtain a preprocessed point cloud; Step S2, inputting the preprocessed point cloud into a point cloud 3D target detection model to obtain multiple initial targets containing target 3D bounding boxes and target categories; Step S3, performing target confidence filtering and non-maximum suppression on all initial targets to obtain the final target. The point cloud 3D target detection model includes: a point cloud feature extraction module, including a hierarchical point cloud encoder, used to extract multiple point cloud features of different depths from the preprocessed point cloud; a candidate target prediction module, including a multilayer perceptron, used to obtain candidate targets based on the deepest point cloud features; and a multi-scale cascaded optimization module, used to combine multiple shallow point cloud features to perform multi-scale cascaded optimization on the candidate targets, obtaining multiple optimized targets corresponding to different levels, where each optimized target and candidate target is an initial target.

[0008] The highly adaptable point cloud 3D target detection method based on active sampling provided by the present invention may also have the following features: Step S1 includes the following sub-steps: Step S1-1, perform point cloud quality detection on the input point cloud, screen out qualified point clouds, and obtain qualified point clouds; Step S1-2, perform point cloud region clipping on the qualified point clouds to obtain clipped point clouds; Step S1-3, perform point cloud sampling on the clipped point clouds to obtain preprocessed point clouds.

[0009] The highly adaptive point cloud 3D target detection method based on active sampling provided by this invention may also have the following features: the hierarchical point cloud encoder is PointNet++, and PointNet++ includes multiple point cloud feature extraction units connected in sequence. Each point cloud feature extraction unit includes a sampler and a group abstractor. The sampler is used to sample the input to obtain a subset, and the group abstractor is used to query the neighborhood of each point in the subset. The permutation-invariant neighborhood features are extracted as point-by-point features within the subset through a multilayer perceptron and a symmetric function.

[0010] The highly adaptable point cloud 3D target detection method based on active sampling provided by the present invention may also have the following features: the sampling mode of the sampler of the first point cloud feature extraction unit is uniform sampling, the sampling mode of the sampler of the second to last point cloud feature extraction units is active sampling, and the input of the second to last point cloud feature extraction units is point cloud features.

[0011] The highly adaptive point cloud 3D target detection method based on active sampling provided by this invention may also have the following features: the multilayer perceptron of the candidate target prediction module includes: a shared feedforward neural network, used to convert the deepest point cloud features into target semantic features; a 3D target bounding box regression head, used to obtain the target 3D bounding box based on the target semantic features; and a target classifier, used to obtain the target category based on the target semantic features.

[0012] The highly adaptive point cloud 3D target detection method based on active sampling provided by this invention may also have the following features: the multi-scale cascaded optimization module includes multiple cascaded optimization units connected in sequence, each cascaded optimization unit includes a group abstractor and a multilayer perceptron, the input of each cascaded optimization unit is the output of the previous cascaded optimization unit and the shallow point cloud features corresponding to that cascaded optimization unit, and the output of the previous cascaded optimization unit corresponding to the first cascaded optimization unit is the candidate target.

[0013] The highly adaptive point cloud 3D object detection method based on active sampling provided by this invention may also have the following features: the training process of the point cloud 3D object detection model includes the following steps: Step T1, constructing an initial point cloud 3D object detection model; Step T2, preprocessing and data augmenting the existing point cloud dataset to obtain a training dataset; Step T3, selecting training point clouds from the training dataset to input into the initial point cloud 3D object detection model to obtain the corresponding training output; Step T4, calculating the loss error based on the training output; Step T5, updating the parameters of the initial point cloud 3D object detection model through backpropagation based on the loss error; Step T6, repeating steps T3 to T5 until the training completion condition is met, then using the trained initial point cloud 3D object detection model as the point cloud 3D object detection model.

[0014] The highly adaptive point cloud 3D target detection method based on active sampling provided by the present invention may also have the following features: wherein, in step T4, the loss function used to calculate the loss error includes: bidirectional KL divergence loss of the auxiliary supervised sampling results, SmoothL1 loss of supervised 3D target box regression, and binary cross-entropy (BCE) loss of supervised target classification.

[0015] The highly adaptive point cloud 3D target detection method based on active sampling provided by this invention may also have the following features: Step S3 includes the following sub-steps: Step S3-1, perform validity detection on all initial targets, filter out initial targets that do not contain any points, and obtain multiple valid targets; Step S3-2, perform target confidence filtering on all valid targets according to a specified confidence threshold, and obtain multiple high-confidence targets; Step S3-3, perform non-maximum suppression on all high-confidence targets according to a specified intersection-union ratio threshold, and obtain the final target.

[0016] The role and effect of invention

[0017] According to the highly adaptive point cloud 3D target detection method based on active sampling of the present invention, the input point cloud is preprocessed through point cloud quality detection, point cloud region clipping, and point cloud sampling, thereby deleting poor-quality point clouds and reducing unnecessary computation; the active sampling sampler improves the encoder's adaptability to various point clouds; and the multi-scale cascaded optimization module effectively utilizes multi-scale features to obtain fine-grained bounding boxes. Therefore, the highly adaptive point cloud 3D target detection method based on active sampling of the present invention can achieve accurate 3D target detection of point clouds. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the highly adaptive point cloud 3D target detection method based on active sampling in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the structure and working principle of the point cloud 3D target detection model in an embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the training process of the point cloud 3D target detection model in an embodiment of the present invention. Detailed Implementation

[0021] To make the technical means, creative features, objectives and effects of this invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the highly adaptable point cloud 3D target detection method based on active sampling of this invention.

[0022] This embodiment provides a highly adaptive point cloud 3D target detection method based on active sampling, used to obtain the corresponding final target from the input point cloud. In this embodiment, the input point cloud is a radar point cloud collected in an autonomous driving scenario, and the corresponding final target is a detection target involving common traffic elements such as pedestrians, cars, and cyclists.

[0023] Figure 1 This is a flowchart illustrating the highly adaptive point cloud 3D target detection method based on active sampling in an embodiment of the present invention.

[0024] like Figure 1 As shown, the highly adaptive point cloud 3D target detection method based on active sampling in this embodiment includes the following steps:

[0025] Step S1: Preprocess the input point cloud to obtain a preprocessed point cloud.

[0026] Step S1 includes the following sub-steps:

[0027] Step S1-1: Perform point cloud quality inspection on the input point cloud, and filter out point clouds that meet the quality requirements to obtain a qualified point cloud. In this embodiment, the NumPy tool is used to delete illegal values ​​in the data, including non-real values ​​such as inf and nan, as well as delete data frames with significant missing points, thereby obtaining a qualified point cloud.

[0028] Steps S1-2 involve cropping the point cloud region from the qualified point cloud to obtain a cropped point cloud. In this embodiment, the point cloud region cropping operation for child laborers involves cropping out the region containing the target to be detected.

[0029] Steps S1-3 involve sampling the cropped point cloud to obtain a preprocessed point cloud. In this embodiment, a fixed number of points are sampled from the point cloud using random sampling to ensure that the point cloud size input to subsequent models is uniform, thus meeting the requirements for batch computation.

[0030] Step S2: Input the preprocessed point cloud into the point cloud 3D target detection model to obtain multiple initial targets containing target 3D bounding boxes and target categories.

[0031] Figure 2 This is a schematic diagram illustrating the structure and working principle of the point cloud 3D target detection model in an embodiment of the present invention.

[0032] like Figure 2 As shown, the point cloud 3D target detection model 100 includes a point cloud feature extraction module 10, a candidate target prediction module 20, and a multi-scale cascade optimization module 30.

[0033] The point cloud feature extraction module 10 includes a hierarchical point cloud encoder for extracting point cloud features at multiple different depth levels from the preprocessed point cloud. The hierarchical point cloud encoders include PointNeXt, KPConv, and PointNet++.

[0034] The hierarchical point cloud encoder in this embodiment is PointNet++. PointNet++ includes multiple point cloud feature extraction units 101 connected in sequence. Each point cloud feature extraction unit includes a sampler 1011 and a group abstractor 1012. In this embodiment, the point cloud 3D object detection model 100 includes three point cloud feature extraction units 101 connected in sequence. In this embodiment, the input of the first point cloud feature extraction unit 101 is the preprocessed point cloud, i.e., the point cloud itself. The corresponding output, that is, the input of the second point cloud feature extraction unit 101, is the sampling points. The output of the second point cloud feature extraction unit 101, which is the input of the third point cloud feature extraction unit 101, is the sampling point. The output of the third point cloud feature extraction unit 101, which is the input of the candidate target prediction module 20, is the sampling point.

[0035] Sampler 1011 is used to sample the input to obtain a subset.

[0036] Among them, the sampling method of the sampler 1011 of the first point cloud feature extraction unit 101 is uniform sampling, that is, sampling the farthest point FPS. The sampling method of the sampler 1011 of the second to last point cloud feature extraction units 101 is active sampling. The input of the second to last point cloud feature extraction units 101 is point cloud features.

[0037] In this embodiment, the active sampling operation is divided into two parts: value prediction and weighted sampling. Value prediction: The sampler 1011 first predicts the value of each point based on point-by-point features. This prediction is performed by a multilayer perceptron with two linear layers and one ReLU activation function. Weighted sampling: The values ​​are normalized to obtain a sampling probability distribution. This distribution is then used as weights, and a specified number of points are sampled from it using weighted random sampling without replacement. In addition to the above process, a pass-through strategy is used during model training to ensure the differentiability of the sampling process, allowing the multilayer perceptron used for value prediction to be optimized during end-to-end training.

[0038] The group abstractor 1012 is used to query the neighborhood of each point in the subset, and extracts permutation-invariant neighborhood features as point-by-point features within the subset through a multilayer perceptron and symmetric functions.

[0039] The candidate target prediction module 20 includes a multilayer perceptron 201, which is used to obtain candidate targets, i.e., predicted target boxes, based on the deepest point cloud features.

[0040] The multilayer perceptron 201 of the candidate target prediction module 20 includes: a shared feedforward neural network, a 3D target box regression head, and a target classifier.

[0041] A shared feedforward neural network is used to transform the deepest point cloud features into target semantic features. In this embodiment, the shared feedforward neural network includes multiple linear layers, a batch normalization layer, and a ReLU activation function layer.

[0042] The 3D target bounding box regression head is used to obtain the target's 3D bounding box based on the target's semantic features. In this embodiment, the target's 3D bounding box includes the center's three-dimensional coordinates, length, width, and height dimensions, and the target's heading angle.

[0043] The target classifier is used to determine the target category based on the target's semantic features. In this embodiment, the target categories include various categories from the background and the specified foreground.

[0044] The multi-scale cascaded optimization module 30 is used to combine multiple shallow point cloud features to perform multi-level cascaded optimization on candidate targets, resulting in multiple optimization targets corresponding to different levels.

[0045] The multi-scale cascaded optimization module 30 includes multiple sequentially connected cascaded optimization units 301. Each cascaded optimization unit 301 includes a group abstractor 3011 and a multilayer perceptron 3012. In this embodiment, both the group abstractor 3012 and the group abstractor 3011 are constructed based on the group abstractor in the PointNet++ single-layer local feature extractor SetAbstractionMSG. In this embodiment, the group abstractor 3011 is used to aggregate shallower point cloud semantic features, further extracting contextual information around the target and providing a more accurate target prediction. The corresponding sampling subset is the target center coordinates, i.e., the center three-dimensional coordinates.

[0046] The input to each cascaded optimization unit 301 is the output of the previous cascaded optimization unit 301 and the shallow point cloud features corresponding to that cascaded optimization unit 301. The output of the previous cascaded optimization unit corresponding to the first cascaded optimization unit 301 is the candidate target. In this embodiment, the point cloud 3D target detection model 100 includes two cascaded optimization units 301. The input to the first cascaded optimization unit 301 is the predicted target box. and sampling points The corresponding output is the predicted target box. The input to the second cascaded optimization unit 301 is the predicted target box. and sampling points The corresponding output is the predicted target box.

[0047] In this context, each optimization objective and candidate objective is an initial objective, meaning the initial objective includes the predicted target bounding box. Predicted target boxes and predicted target boxes

[0048] Figure 3 This is a schematic diagram of the training process of the point cloud 3D target detection model in an embodiment of the present invention.

[0049] like Figure 3 As shown, the training process of the point cloud 3D object detection model includes the following steps:

[0050] Step T1: Construct the initial point cloud 3D target detection model.

[0051] Step T2 involves preprocessing and augmenting the existing point cloud dataset to obtain the training dataset.

[0052] In this embodiment, the existing point cloud datasets are two publicly available autonomous driving scenario datasets: the KITTI dataset and the VoD dataset. The KITTI dataset is a dataset for research in the field of autonomous driving, jointly sponsored by the Karlsruhe Institute of Technology in Germany and Toyota Technological University at Chicago. It collects up to 6 hours of real traffic environment data. This dataset provides 14,999 frames of LiDAR point cloud data with different scenes for 3D object detection tasks. Each frame contains approximately 100,000 points, and the detected targets include common traffic elements such as pedestrians, cars, and cyclists. The VoD dataset provides 8,693 frames of 4D imaging millimeter-wave radar point clouds. In addition to coordinates, each point also includes attributes such as absolute Doppler velocity, ground Doppler velocity after vehicle motion compensation, and radar cross section. Due to the low resolution of 4D radar, a multi-frame aggregation method is used, where 5 adjacent frame point clouds are aggregated to form 1 frame point cloud, which contains an average of approximately 2,000 points. The detected targets also include common traffic elements such as pedestrians, cars, and cyclists.

[0053] In this embodiment, the preprocessing process for the downstream point cloud data is the same as the preprocessing process for the input point cloud, i.e., step S1. During point cloud region cropping, for the KITTI and VoD datasets, the target only exists within the observation area of ​​the forward-looking camera. Therefore, based on the radar and camera's intrinsic and extrinsic parameters and the projection matrix, all points in the point cloud are projected onto the forward-looking image plane, and points outside the plane are removed. Then, based on the maximum area where the target appears, points exceeding the area are removed. During point cloud sampling, the fixed number of sampling points is 16384 for the KITTI dataset and 4096 for the VoD dataset.

[0054] Data augmentation in this embodiment includes: 1) random flipping: random flipping around the x-axis of the point cloud coordinate system; 2) random rotation: random rotation around the z-axis of the point cloud coordinate system; 3) random shuffling: random shuffling of the order of points in the point cloud; 4) random target sampling: randomly selecting targets from other frames and copying and pasting them into the current frame; 5) random target noise: randomly rotating and randomly translating all targets around the z-axis of the point cloud coordinate system.

[0055] In this embodiment, the preprocessed point cloud corresponding to the KITTI dataset is a matrix of size B*16384*4, where B is the number of point cloud frames in each batch. In this embodiment, B=16, 16384 is the number of points contained in a frame of point cloud, i.e., the fixed number of sampling points, and 4 is the attribute contained in each point, including three-dimensional coordinates (x,y,z) and reflection intensity i.

[0056] Step T3 involves selecting training point clouds from the training dataset to input into the initial point cloud 3D object detection model, thereby obtaining the corresponding training output. In this embodiment, 16 frames of point clouds are used as a batch, and a total of 120 training rounds are conducted.

[0057] In this embodiment, the initial point cloud 3D target detection model includes four point cloud feature extraction modules 10. The outputs of the four modules are B*4096*64, B*1024*128, B*512*256, and B*256*512, respectively, with the number of points gradually decreasing and the feature dimension gradually increasing. The candidate target prediction module 20 of the initial point cloud 3D target detection model predicts the surrounding targets at each sampling point position based on B*256*512, resulting in a target prediction output of B*256*(7+cls), where 7 represents the 3D target box dimension, including the center three-dimensional coordinates (x,y,z), length, width, and height dimensions (l,w,h), and heading angle α, while cls represents the number of target categories to be detected, indicating the confidence level of the target belonging to each category. The multi-scale cascaded optimization module 30 of the initial point cloud 3D target detection model takes the center three-dimensional coordinates of all the predicted targets, performs a group abstractor operation, and outputs the predicted target center context features with a feature dimension of B*256*1024, thereby predicting more accurate targets. The multi-scale cascaded optimization module 30 of the initial point cloud 3D target detection model is stacked in three layers. The point cloud feature input of each layer's SetAbstractionMSG structure includes the output of the last three layers of the corresponding point cloud feature extraction module 10, namely B*256*512, B*512*256, and B*1024*128.

[0058] Step T4: Calculate the loss error based on the training output.

[0059] In step T4, the loss function used to calculate the loss error includes: bidirectional KL divergence loss for supervised sampling results, SmoothL1 loss for supervised 3D bounding box regression, and binary cross-entropy (BCE) loss for supervised target classification. In this embodiment, weight decay is also used in the loss function to prevent overfitting; the weight decay parameter is set to 1e-4.

[0060] Step T5: Update the parameters of the initial point cloud 3D object detection model through backpropagation based on the loss error.

[0061] Step T6: Repeat steps T3 to T5 until the training completion condition is met. Then, use the trained initial point cloud 3D object detection model as point cloud 3D object detection model 100. In this embodiment, the training completion condition is that the initial point cloud 3D object detection model converges.

[0062] In this embodiment, test data is constructed using existing point cloud datasets to test the trained point cloud 3D object detection model 100, and the model performance is measured by the average precision (AP).

[0063] The AP of point cloud 3D object detection model 100 is as follows, based on the test data constructed using the KITTI dataset:

[0064] 1) The AP rates for easy, average, and difficult car-related goals were 89.05%, 78.88%, and 77.45%, respectively; 2) The AP rates for easy, average, and difficult pedestrian-related goals were 65.08%, 59.74%, and 55.80%, respectively; 3) The AP rates for easy, average, and difficult cyclist-related goals were 86.01%, 71.34%, and 67.31%, respectively.

[0065] The AP of point cloud 3D object detection model 100 is as follows, based on the test data constructed using the VoD dataset:

[0066] 1) The APs for cars, pedestrians, and cyclists in the entire labeled area were 39.39%, 46.92%, and 76.16%, respectively; 2) The APs for cars, pedestrians, and cyclists in the driving corridor area were 73.93%, 57.92%, and 92.88%, respectively.

[0067] This demonstrates that the 3D point cloud target detection model 100 achieves remarkable performance on both lidar and 4D millimeter-wave radar point clouds, especially on small targets.

[0068] Step S3 involves applying target confidence filtering and non-maximum suppression to all initial targets to obtain the final target. In this embodiment, step S3 is considered post-processing, and the final target is the output target, which is the predicted target box.

[0069] Step S3 includes the following sub-steps:

[0070] Step S3-1: Perform validity checks on all initial targets, filtering out initial targets that do not contain any points, resulting in multiple valid targets. In this embodiment, the predicted bounding boxes of some initial targets may not contain any points because they are not responsible for detecting specific foregrounds or are affected by noise; therefore, they are considered invalid predictions and are directly deleted.

[0071] Step S3-2: Based on the specified confidence threshold, target confidence filtering is performed on all valid targets to obtain multiple high-confidence targets. In this embodiment, since the model does not explicitly determine whether the predicted target is foreground, but only outputs the foreground category with the highest confidence and the confidence value, a pre-set confidence threshold is used to filter out targets with low confidence.

[0072] Step S3-3: Based on the specified Intersection over Union (IoU) threshold, perform non-maximum suppression on all high-confidence targets to obtain the final target. In this embodiment, since multiple closely spaced sampling points may give similar predictions, i.e., multiple overlapping target boxes may be generated for the same target, a non-maximum suppression algorithm is used to retain target boxes with higher confidence based on a given IoU threshold and delete target boxes whose intersection with the target is greater than the threshold.

[0073] The role and effect of the embodiments

[0074] According to the highly adaptable point cloud 3D object detection method based on active sampling involved in this embodiment, the input point cloud is preprocessed through point cloud quality detection, point cloud region clipping, and point cloud sampling, thereby deleting poor-quality point clouds and reducing unnecessary computation. The active sampling sampler improves the encoder's adaptability to various point clouds. A multi-scale cascaded optimization module effectively utilizes multi-scale features to obtain fine-grained bounding boxes. In summary, this method can achieve accurate 3D object detection on point clouds.

[0075] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A highly adaptive point cloud 3D target detection method based on active sampling, used to obtain the corresponding final target based on the input point cloud, characterized in that, Includes the following steps: Step S1: Preprocess the input point cloud to obtain a preprocessed point cloud; Step S2: Input the preprocessed point cloud into the point cloud 3D target detection model to obtain multiple initial targets containing target 3D bounding boxes and target categories; Step S3: Perform target confidence filtering and non-maximum suppression on all the initial targets to obtain the final target. The point cloud 3D target detection model includes: The point cloud feature extraction module includes a hierarchical point cloud encoder for extracting point cloud features at multiple different depth levels from the preprocessed point cloud. The hierarchical point cloud encoder is PointNet++, which includes multiple point cloud feature extraction units connected in sequence. Each point cloud feature extraction unit includes a sampler and a group abstractor. The sampler is used to sample the input to obtain a subset. The sampling method of the sampler in the second to last point cloud feature extraction unit is active sampling. The group abstractor is used to query the neighborhood of each point in the subset and extract permutation-invariant neighborhood features as point-by-point features within the subset through a multilayer perceptron and a symmetric function. The candidate target prediction module includes a multilayer perceptron for obtaining candidate targets based on the deepest point cloud features; The multi-scale cascaded optimization module is used to combine multiple shallow point cloud features to perform multi-level cascaded optimization on the candidate targets, resulting in multiple optimized targets corresponding to different levels. Each of the optimization objectives and the candidate objectives are the initial objectives.

2. The highly adaptive point cloud 3D target detection method based on active sampling as described in claim 1, Its features are: Step S1 includes the following sub-steps: Step S1-1: Perform point cloud quality detection on the input point cloud, filter out point clouds with qualified quality, and obtain qualified point clouds; Step S1-2: Crop the point cloud region of the qualified point cloud to obtain the cropped point cloud; Steps S1-3: Perform point cloud sampling on the cropped point cloud to obtain the preprocessed point cloud.

3. The highly adaptive point cloud 3D target detection method based on active sampling according to claim 1, characterized in that: in, The sampling method of the sampler in the first point cloud feature extraction unit is uniform sampling. The input to each of the second to last point cloud feature extraction units is the point cloud feature.

4. The highly adaptive point cloud 3D target detection method based on active sampling according to claim 1, Its features are: The multilayer perceptron of the candidate target prediction module includes: A shared feedforward neural network is used to transform the deepest point cloud features into target semantic features; A 3D target bounding box regression head is used to obtain the 3D bounding box of the target based on the target semantic features; A target classifier is used to obtain the target category based on the target semantic features.

5. The highly adaptive point cloud 3D target detection method based on active sampling according to claim 1, characterized in that: in, The multi-scale cascaded optimization module includes multiple cascaded optimization units connected in sequence. The cascaded optimization unit includes a group abstractor and a multilayer perceptron. The input to each cascaded optimization unit is the output of the previous cascaded optimization unit and the shallow point cloud features corresponding to that cascaded optimization unit. The output of the previous cascaded optimization unit corresponding to the first cascaded optimization unit is the candidate target.

6. The highly adaptive point cloud 3D target detection method based on active sampling according to claim 1, Its features are: The training process of the point cloud 3D object detection model includes the following steps: Step T1: Construct an initial point cloud 3D object detection model; Step T2 involves preprocessing and augmenting the existing point cloud dataset to obtain the training dataset. Step T3: Select training point clouds from the training dataset and input them into the initial point cloud 3D object detection model to obtain the corresponding training output; Step T4: Calculate the loss error based on the training output; Step T5: Update the parameters of the initial point cloud 3D target detection model through backpropagation based on the loss error; Step T6: Repeat steps T3 to T5 until the training completion condition is met, then use the trained initial point cloud 3D object detection model as the point cloud 3D object detection model.

7. The highly adaptive point cloud 3D target detection method based on active sampling according to claim 6, characterized in that: in, In step T4, the loss function used to calculate the loss error includes: bidirectional KL divergence loss for the auxiliary supervised sampling results, SmoothL1 loss for supervised 3D target bounding box regression, and binary cross-entropy (BCE) loss for supervised target classification.

8. The highly adaptive point cloud 3D target detection method based on active sampling according to claim 1, Its features are: Step S3 includes the following sub-steps: Step S3-1: Perform validity checks on all the initial targets, filter out the initial targets that do not contain any points, and obtain multiple valid targets; Step S3-2: Based on the specified confidence threshold, perform target confidence filtering on all the valid targets to obtain multiple high-confidence targets; Step S3-3: Based on the specified intersection-union ratio threshold, perform non-maximum suppression on all the high-confidence targets to obtain the final target.

Images