Three-dimensional point cloud target detection method and device based on diffusion model

Abstract

The application relates to a three-dimensional point cloud target detection method based on a diffusion model. Point cloud data is acquired, voxelization is performed, and data enhancement is carried out. The voxelized data is sent to a voxel encoder to be converted into two-dimensional dense features in the perspective of a dense bird's-eye view. The labeled box in the perspective of the bird's-eye view is sent to a candidate box generator based on a diffusion model to return generated noise boxes and time step levels. Corresponding features are cropped on the two-dimensional features and sent to a detection decoder for prediction. A loss function is optimized, and iterative training is carried out until the model converges. The trained model is used for three-dimensional target detection. The application generates candidate boxes by using a diffusion model, can directly use random boxes from a Gaussian distribution as candidates during reasoning, avoids manually designing anchor box sizes in an experiential manner, makes the detection decoder more robust, can adapt to input with different noise levels, and obtains detection results. The application also provides a corresponding three-dimensional point cloud target detection device based on a diffusion model.

Description

Technical Field

[0001] This invention belongs to the fields of computer vision and image processing technology, and more specifically, relates to a method and apparatus for three-dimensional point cloud target detection based on a diffusion model. Background Technology

[0002] 3D point cloud object detection is a crucial research topic in the field of autonomous driving perception. It aims to detect targets of interest (such as vehicles and pedestrians) within point cloud data and output their corresponding categories and 3D locations. Point clouds refer to the point cloud data collected by LiDAR sensors mounted on autonomous vehicle platforms in urban street environments, containing information such as the 3D coordinates and reflection intensity of each object's reflection point. As a fundamental problem in autonomous driving, 3D point cloud object detection technology has broad and practical application prospects. Furthermore, 3D point cloud object detection technology is also the foundation for subsequent tasks such as target tracking and trajectory prediction.

[0003] Existing 3D point cloud object detection methods often rely on manually designed anchor boxes. However, in real-world road scenarios, due to the diverse types and sizes of vehicles, designers must carefully design the anchor box sizes; otherwise, the algorithm's detection performance may be severely affected. Furthermore, many existing detection methods predict numerous redundant boxes, potentially hindering algorithm deployment. Therefore, researching end-to-end 3D point cloud object detectors based on alternative anchor box setting methods has significant application value. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a 3D point cloud target detection method based on a diffusion model. This method can generate detection candidate boxes using a diffusion model, thereby maximizing the quality of the candidate boxes and recovering the target from normally randomly distributed candidate boxes during inference. This method is intended to alleviate the technical problem of performance loss caused by improper anchor box settings.

[0005] To achieve the above objectives, this embodiment provides a three-dimensional point cloud target detection method based on a diffusion model, the detection method comprising:

[0006] Step 10: Acquire point cloud data and perform voxelization and data augmentation;

[0007] Step 20: The voxelized data is fed into the voxel encoder to be converted into sparse 3D features, and then converted into 2D dense features from a bird's-eye view through the feature pyramid.

[0008] Step 30: Obtain labeled data, obtain noise time step level, and send the labeled boxes from the bird's-eye view into the candidate box generator based on the diffusion model to return the generated noise boxes and time step level;

[0009] Step 40: Based on the noise box and time step level, crop the corresponding features on the two-dimensional features and send them to the detection decoder for prediction;

[0010] Step 50: Perform label matching based on a one-to-one matching method, optimize the loss function, and perform iterative training until the model converges;

[0011] Step 60: Perform 3D object detection using the trained model.

[0012] In one embodiment of the present invention, in step 20, a voxel encoder based on the existing SECOND 3D backbone network is used to voxelize the point cloud data and encode it into bird's-eye view features, and multi-scale feature fusion is used to enhance the bird's-eye view features.

[0013] In one embodiment of the present invention, step 30 specifically includes:

[0014] During training, the bounding boxes are first repeated a total of N times. Each bounding box is a seven-dimensional vector (x, y, z, w, l, h, θ).

[0015] The components (x, y, w, l, θ) of the aforementioned seven-dimensional vector from a bird's-eye view are selected and, following a diffusion model, are gradually shifted towards Gaussian noise over time. The initial distribution is x0, and the preliminary candidate boxes after adding noise are x0. t The calculation formula is:

[0016]

[0017] in t = randint(1, T) max ), Based on the existing diffusion model, DDIM represents the change in the total variance of the corresponding noise, T max This represents the maximum range of values ​​for the time period.

[0018] In one embodiment of the present invention, the maximum time range value T max It is a dynamically changing integer that changes as the training process progresses, and the calculation formula is:

[0019]

[0020] Where ω and σ are hyperparameters, T is the maximum possible time value, and n is the number of training iterations.

[0021] In one embodiment of the present invention, in step 30, the implementation process of the candidate box generator based on the diffusion model is as follows:

[0022] Remove candidate boxes whose internal point cloud count is less than η from the damaged annotation boxes;

[0023] A set of Gaussian noise with a correlation coefficient is used to resample the number of candidate boxes to N. This correlation coefficient is applied to the (w, l) of the random vector. The two-dimensional Gaussian distribution with the correlation coefficient ρ is given by the following formula:

[0024]

[0025] in After obtaining W and L, they are scaled to a reasonable range of (0, w) and (0, l) to serve as the width and length of the candidate box sampled from the Gaussian distribution. The other components of the seven-dimensional vector are obtained directly from the normal distribution.

[0026] Iterate through the two processes described above until the number of point clouds contained in all pre-selected candidate boxes is greater than η.

[0027] During the reasoning process, random candidate boxes that satisfy the first three steps above are directly generated, and the integer t is directly returned as the maximum value T that can be taken in the aforementioned time.

[0028] In one embodiment of the present invention, in step 40, a plurality of target objects are obtained by the following formula:

[0029]

[0030] in, Indicates the features of a two-dimensional bird's-eye view. The candidate box is represented by t, which represents the noise level at the corresponding time step. The 3D position of each target object is represented by the center point (x, y, z), width w, length l, height h, and rotation angle θ of a 3D Box.

[0031] In one embodiment of the present invention, in step 50, tag matching is performed in a one-to-one matching manner:

[0032] The cost matrix between the prediction results and the labeled true values ​​is constructed according to the following formula.

[0033]

[0034] in, Let λ represent the classification loss, regression loss, and IoU loss from the bird's-eye view, respectively. cls , λ reg , λ IoU These are the weighting coefficients;

[0035] Representing the predicted results and ground truth bounding boxes, a one-to-one Hungarian matching method is used to match the target. for:

[0036]

[0037] In one embodiment of the present invention, in step 50, the total loss function of the training is calculated according to the following formula. Until the total loss function The predetermined value has been reached;

[0038]

[0039] Let λ represent the classification loss, regression loss, and 3D DIoU loss, respectively. cls , λ reg , λ DIoU These are the weighting coefficients.

[0040] According to another aspect of the present invention, a three-dimensional point cloud target detection device based on a diffusion model is also provided, including at least one processor and a memory, wherein the at least one processor and the memory are connected via a data bus, and the memory stores instructions that can be executed by the at least one processor, wherein the instructions, after being executed by the processor, are used to complete the three-dimensional point cloud target detection method based on the diffusion model.

[0041] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:

[0042] (1) The embodiments of the present invention can eliminate the need for manually designed anchor frame dimensions and precise proposed frame placement, thereby avoiding the use of manually set anchor frame dimensions based on experience.

[0043] (2) Furthermore, in response to the problem of label allocation during training, the one-to-one Hungarian matching method can be used to select the number of candidate boxes more flexibly and enable the model to achieve end-to-end in non-iterative inference, avoiding the complex post-processing methods used by common algorithms.

[0044] (3) Furthermore, due to the innovative use of the diffusion model, the detection decoder is able to receive randomly sampled candidate boxes, and the fully trained decoder introduces richer prior information and is therefore more robust. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating a training method for 3D target detection based on a diffusion model proposed in an embodiment of the present invention.

[0046] Figure 2 This is a schematic diagram of the overall framework for three-dimensional target detection based on a diffusion model proposed in an embodiment of the present invention;

[0047] Figure 3 This is a detection effect diagram of a 3D target detection based on a diffusion model provided by an embodiment of the present invention, wherein... Figure 3 (a) shows the detection results for a straight-ahead road scenario. Figure 3 (b) shows the detection results for the intersection scene. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0049] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates a three-dimensional target detection method based on a diffusion model according to the present invention. It should be particularly noted that the following description is exemplary and intended to further illustrate the present invention, rather than limiting the implementation methods.

[0050] Example 1:

[0051] This embodiment uses a deep neural network to extract features and complete training and inference to obtain a trained model. The trained model is then used for detection. This embodiment mainly explains how to train the model to obtain the desired result. Figure 1 and Figure 2 This embodiment provides a training method for a 3D target detection model based on a diffusion model, including:

[0052] Step 10: Acquire point cloud data and perform voxelization and data augmentation.

[0053] In this embodiment, the existing point cloud data and the 3D bounding box of the target to be detected are read. Data augmentation is performed on the point cloud data, including random flipping, random rotation, and resampling of the bounding box. After data augmentation, voxelization is performed with voxel sizes of (0.05m, 0.05m, 0.1m).

[0054] Step 20: The voxelized data is fed into the voxel encoder to be converted into sparse 3D features, and then converted into dense 2D features from a dense bird's-eye view through the feature pyramid.

[0055] A voxel encoder based on the existing SECOND 3D backbone network is used to voxelize point cloud data and encode it into bird’s-eye view features. Multi-scale feature fusion is used to enhance the bird’s-eye view features.

[0056] Specifically, by combining the existing SECOND 3D feature backbone network, 3D sparse convolution is used to extract sparse 3D features from voxel data, and then the features are stitched together in the height dimension. A three-layer feature pyramid is then used to aggregate the 2X, 4X, and 8X features, with the number of filters being 64, 128, and 256, respectively, finally resulting in dense bird's-eye view features.

[0057] Specifically, step 20 includes:

[0058] The acquired point cloud data is voxelized, and sparse 3D convolution is used for feature extraction.

[0059] The obtained feature maps are stitched together along the height direction to obtain a two-dimensional feature map;

[0060] The final bird's-eye view features are obtained by fusing features through a feature pyramid.

[0061] The features of a two-dimensional bird's-eye view can be obtained using the following formula:

[0062]

[0063] Where P represents a 3D point cloud and Encoder represents a voxel encoder.

[0064] Step 30: Obtain labeled data, obtain noise time step level, and send the labeled boxes in the bird's-eye view into the candidate box generator based on the diffusion model to return the generated noise boxes and time step level;

[0065] A candidate box generator based on a diffusion model is used to generate randomized candidate boxes, and the candidate boxes and their corresponding noise time step levels are returned. This step specifically includes:

[0066] During training, the bounding boxes are first repeated a total of N times. Each bounding box is a seven-dimensional vector (x, y, z, w, l, h, θ).

[0067] The components (x, y, w, l, θ) of the aforementioned seven-dimensional vector from a bird's-eye view are selected and, following a diffusion model, are gradually shifted towards Gaussian noise over time. The initial distribution is x0, and the preliminary candidate boxes after adding noise are x0. t The calculation formula is:

[0068]

[0069] in t = randint(1, T) max ), Based on the existing diffusion model, DDIM represents the change in the total variance of the corresponding noise.

[0070] The maximum value T in the above time range max It is a dynamically changing integer that changes as the training process progresses, and the calculation formula is:

[0071]

[0072] Where ω and σ are hyperparameters, T is the maximum possible time value, and n is the number of training iterations, from (1, T max A random integer t obtained directly is returned;

[0073] Remove candidate boxes whose internal point cloud count is less than η from the damaged annotation boxes;

[0074] A set of Gaussian noise with a correlation coefficient is used to resample the number of candidate boxes to N. This correlation coefficient acts on the (w, l) of the random vector. The two-dimensional Gaussian distribution with the correlation coefficient ρ is given by the following formula:

[0075]

[0076] in After obtaining W and L, they are scaled to a reasonable range of (0, w) and (0, l) to serve as the width and length of the candidate box sampled from the Gaussian distribution. The other components of the seven-dimensional vector are obtained directly from the normal distribution.

[0077] Iterate through the two processes described above until the number of point clouds contained in all pre-selected candidate boxes is greater than η.

[0078] During the reasoning process, random candidate boxes that satisfy the first three steps above are directly generated, and the integer t is directly returned as the maximum value T that can be taken in the aforementioned time.

[0079] For example: Using existing bounding box data, obtain bounding boxes (x, y, z, w, l, h, θ), and repeat these boxes until a total of 300 are obtained. Select the items from the bird's-eye view (x, y, w, l, θ) and add noise. z and h are set to half the scene height and the scene height by default. First, the maximum possible time step is preset to 1000, the total number of training iterations is 640, and n is the number of training iterations. T is calculated using the following formula. max :

[0080]

[0081] In the range (1, T) max Select a random integer t within the range, and use the formula:

[0082]

[0083] α tThis represents the change in the corresponding noise variance, resulting in the noise candidate box x0 with noise added at time step t. t .

[0084] The calculated candidate bounding boxes correspond to the number of points in the point cloud space. Candidate bounding boxes with fewer than 5 points in their internal point cloud are deleted, and the total number of boxes is replenished to 300. The replenished boxes satisfy the following conditions:

[0085] W = 0.8L + 0.6X

[0086] in W and L are scaled to the ranges (0, 5) and (0, 8) to serve as the width and length of candidate boxes sampled from a Gaussian distribution. The center point position and angle are randomly selected directly from a normal distribution. The number of point clouds contained within the resulting boxes is recalculated, and candidate boxes with fewer than 5 internal point clouds are deleted. Random boxes are then added again to bring the total number of boxes to 300. This process is repeated until all boxes contain at least 5 point clouds, with a maximum of 20 iterations.

[0087] Finally, the time step t and the obtained candidate boxes are returned.

[0088] Step 40: Based on the noise box and time step level, crop the corresponding features on the two-dimensional features and send them to the detection decoder for prediction;

[0089] The cropped feature representation with rotation is cropped from the bird's-eye view features obtained by the encoder;

[0090] Specifically, the time step t is processed through a multilayer perceptron to obtain the time embedding, and the bird's-eye view features corresponding to the candidate boxes are cropped. As a layer of the detection decoder, it includes... Self-attention operation, and With the addition of dynamic convolutional operations with temporal embedding, and the use of regression or classification prediction heads from multilayer perceptrons, the entire detection decoder consists of the same 6 layers, with the input of each layer coming from the output of the previous layer. Thus, the detection decoder regresses the final detection result step by step.

[0091] Step 50: Perform label matching according to the one-to-one matching method, optimize the loss function, and perform iterative training until the model converges;

[0092] The correlation cost between each output of the detector and each bounding box (M in total) is calculated to obtain a matching cost matrix of 300×M. The calculation formula is as follows:

[0093]

[0094] use The loss, specifically, if (x, y, z, w, l, h, θ) represents the true value, Let the predicted value of the detector decoder be:

[0095]

[0096]

[0097]

[0098]

[0099] The existing Focal loss was used. It directly represents the bird's-eye view perspective (IoU) between the predicted bounding box and the labeled bounding box.

[0100] Representing the predicted results and ground truth bounding boxes, a one-to-one Hungarian matching method is used to match the target. for:

[0101] The loss function is calculated directly between the matched predicted bounding box and the ground truth value.

[0102]

[0103] As mentioned above, This indicates the existing rotational DIoU loss.

[0104] After calculating the final loss value, gradient descent is performed to update the parameters and optimize the loss function. Iterative training is then conducted to obtain the final model.

[0105] Step 60: Perform 3D object detection using the trained model.

[0106] Furthermore, during the reasoning process, a diffusion approach involving multiple iterations can be used:

[0107] The prediction is gradually refined starting with candidate boxes sampled from a Gaussian distribution; in each sampling step, a random box or the estimated box from the previous step is noisily added and fed into the decoder to predict the result of the current stage. The candidate box for the next step can be calculated using the following formula:

[0108]

[0109]

[0110] Where, x t x t-s Candidate boxes representing two adjacent steps. It is the decoder's prediction offset, ∈ t It is Gaussian noise, α t α t-s This represents the change in noise variance between two adjacent steps. The number of sampling steps can be equal to or greater than 1. s is the starting time level (i.e., T) divided by the number of sampling steps.

[0111] When using a regression approach with multiple iterations, a non-maximum suppression (NMS) post-processing operation is required to filter out redundant predictions.

[0112] Example 2:

[0113] Based on the training method of the aforementioned Embodiment 1, this embodiment mainly explains how to use the obtained model for inference.

[0114] If only one iteration is performed, then during the reasoning process:

[0115] Step 10: Same as in Example 1.

[0116] Step 20: Same as in Example 1.

[0117] Step 30: For candidate boxes (cx, cy, cz, w, l, h, θ), sample 300 boxes directly from the normal distribution, where W = 0.8L + 0.6X. W and L are scaled to the ranges (0, 5) and (0, 8) to serve as candidate boxes w and l sampled from a Gaussian distribution. cz and h are fixed at half and full height of the point cloud scene, respectively. The number of points contained within each box is calculated, and candidate boxes with fewer than 5 points are removed. Random boxes are then added again to bring the total to 300, and this process is repeated until all boxes contain at least 5 points. The maximum number of iterations is set to 20. The time step t is set to 1000.

[0118] Step 40: Same as in Example 1.

[0119] Step 50: Does not exist during the reasoning process.

[0120] Step 60: Does not exist during the reasoning process.

[0121] If multiple iterations are required Then, in the reasoning process:

[0122] Based on only one iteration, save the output of step 40 as x. 1000 The candidate boxes and time that are re-entered into step 40 are calculated using the following formula:

[0123]

[0124]

[0125] Where, x t x t-s Candidate boxes representing two adjacent steps. It is the decoder's prediction offset, ∈ t It is Gaussian noise, α t α t-s This represents the change in noise variance between two adjacent steps.

[0126] x t-s Input the data into step 40, save the output again, and repeat the above steps to finally obtain x. 1000 x 1000-s , ..., x0.

[0127] The above results are combined and post-processed with non-maximum suppression (NMS) to obtain the final result.

[0128] Example 3:

[0129] Based on the training method of Embodiment 1 and the inference method of Embodiment 2, this embodiment presents the performance achieved for the car category on the mainstream 3D object detection dataset KITTI.

[0130] The KITTI dataset contains 3717 training samples and 3769 validation samples. Each sample includes point cloud data acquired by vehicle-mounted LiDAR collected on the streets of Germany, images collected by optical cameras, and high-quality bounding boxes. The dataset covers three types of targets: cars, bicycles, and pedestrians.

[0131] This embodiment trains on the training set and tests the performance of the car category on the validation set. The evaluation metric used is 3D mAP, where the threshold for the 3D bounding box intersection-over-union ratio is 0.7, and the calculation uses the results of 11 recall points. Figure 3 The image shown is a detection effect diagram of a 3D target detection based on a diffusion model provided by an embodiment of the present invention, wherein... Figure 3 (a) shows the detection results for a straight-ahead road scenario. Figure 3 (b) shows the detection results for the intersection scene.

[0132] The evaluation results of this embodiment under different inference iterations are shown in Table 1. Compared with the baseline model results based on manually designed anchor boxes, this embodiment can directly use random candidate boxes as input to the detection decoder and perform detection, and achieves a certain performance improvement.

[0133] Table 1 shows the results under partially labeled data settings.

[0134]

[0135] Example 4:

[0136] Furthermore, the present invention also provides a three-dimensional point cloud target detection device based on a diffusion model, including at least one processor and a memory, wherein the at least one processor and the memory are connected through a data bus, and the memory stores instructions that can be executed by the at least one processor. After being executed by the processor, the instructions are used to complete the three-dimensional point cloud target detection method based on the diffusion model.

[0137] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-dimensional point cloud target detection method based on a diffusion model, characterized in that, The detection method includes: Step 10: Acquire point cloud data and perform voxelization and data augmentation; Step 20: The voxelized data is fed into the voxel encoder to be converted into sparse 3D features, and then converted into 2D dense features from a bird's-eye view through the feature pyramid. Step 30: Obtain labeled data, obtain noise time step level, and send the labeled boxes from the bird's-eye view to the candidate box generator based on the diffusion model to return the generated noise boxes and time step level; Step 30 specifically includes: During the training process, the bounding boxes are first repeated N times, and each bounding box is a seven-dimensional vector (x,y,z,w,l,h,θ). The components (x, y, w, l, θ) of the aforementioned seven-dimensional vector from a bird's-eye view are selected and, following a diffusion model, are gradually shifted towards Gaussian noise over time. The initial distribution is x0, and the preliminary candidate boxes after adding noise are x0. t The calculation formula is: in t = randint(1, T) max ), Based on the existing diffusion model, DDIM represents the change in the total variance of the corresponding noise, T max This represents the maximum range of values ​​for the time period. In step 30, the implementation process of the candidate box generator based on the diffusion model is as follows: Remove candidate boxes whose internal point cloud count is less than η from the damaged annotation boxes; A set of Gaussian noise with a correlation coefficient is used to resample the number of candidate boxes to N. The correlation coefficient is applied to the (w,l) of the random vector. The two-dimensional Gaussian distribution with the correlation coefficient ρ is given by the following formula: in After obtaining W and L, they are scaled to a reasonable range of (0,w) and (0,l) to serve as the width and length of the candidate box sampled from the Gaussian distribution. The other components of the seven-dimensional vector are obtained directly from the normal distribution. Iterate through the two processes described above until the number of point clouds contained in all pre-selected candidate boxes is greater than η. During the reasoning process, random candidate boxes that satisfy the first three steps above are directly generated, and the integer t is directly returned as the maximum value T that can be taken in the aforementioned time. Step 40: Based on the noise box and time step level, crop the corresponding features on the two-dimensional features and send them to the detection decoder for prediction; Step 50: Perform label matching based on a one-to-one matching method, optimize the loss function, and perform iterative training until the model converges; Step 60: Perform 3D object detection using the trained model.

2. The three-dimensional point cloud target detection method based on a diffusion model according to claim 1, characterized in that, In step 20, a voxel encoder based on the existing SECOND 3D backbone network is used to voxelize the point cloud data and encode it into bird's-eye view features, and multi-scale feature fusion is used to enhance the bird's-eye view features.

3. The three-dimensional point cloud target detection method based on a diffusion model according to claim 1, characterized in that, The maximum time range value T max It is a dynamically changing integer that changes as the training process progresses, and the calculation formula is: Where ω and σ are hyperparameters, T is the maximum possible time value, and n is the number of training iterations.

4. The three-dimensional point cloud target detection method based on the diffusion model according to claim 1, characterized in that, In step 40, multiple target objects are obtained using the following formula: in, Indicates the features of a two-dimensional bird's-eye view. The candidate box is represented by t, which represents the noise level at the corresponding time step. The 3D position of each target object is represented by the center point (x,y,z) of a 3D Box, the width w, the length l, the height h, and the rotation angle θ.

5. The three-dimensional point cloud target detection method based on a diffusion model according to claim 1, characterized in that, In step 50, tag matching is performed using a one-to-one matching method: The cost matrix between the prediction results and the labeled true values ​​is constructed according to the following formula. in, Let λ represent the classification loss, regression loss, and IoU loss from the bird's-eye view, respectively. cls ,λ reg ,λ IoU These are the weighting coefficients; y j Representing the predicted results and ground truth bounding boxes, a one-to-one Hungarian matching method is used to match the target. for:

6. The three-dimensional point cloud target detection method based on the diffusion model according to claim 5, characterized in that, In step 50, the total training loss function is calculated according to the following formula. Until the total loss function The predetermined value has been reached; Let λ represent the classification loss, regression loss, and 3DDIoU loss, respectively. cls ,λ reg ,λ DIoU These are the weighting coefficients.

7. A three-dimensional point cloud target detection device based on a diffusion model, characterized in that: The method includes at least one processor and a memory, which are connected via a data bus. The memory stores instructions that can be executed by the at least one processor. After being executed by the processor, the instructions are used to complete the three-dimensional point cloud target detection method based on the diffusion model as described in any one of claims 1-6.

Images