Millimeter wave radar sparse point cloud diffusion model optimization method

Abstract

The invention belongs to the technical field of radar signal processing, and relates to a millimeter wave radar sparse point cloud diffusion model optimization method, which comprises the following steps that: 1, on the basis of a diffusion model, a baseline model adopts the weighted sum of mean square error and perception loss as a loss function; 2, a two-stage strategy is adopted, and on the premise of not sacrificing the strong denoising capability learned by the pre-training model, the ability of the model to generate the structure is liberated by adjusting the balance point of a loss function; 3, the radar sensing framework constructs a direction-dependent uncertainty model based on radar measurement physical characteristics through a self-adaptive error elliptic filter, and each radar reflection point is modeled into an elliptic region with directional error distribution; according to the method, the point cloud density, the integrity and the overall precision are remarkably improved. And particularly, the optimization of the order of magnitude is realized on CD and F-score indexes.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing technology, specifically relating to an optimization method for a sparse point cloud diffusion model of millimeter-wave radar. Background Technology

[0002] In the complex real world, the autonomous navigation capabilities of intelligent systems such as micro-air vehicles (MAVs) and self-driving vehicles heavily rely on accurate and robust perception of their surroundings. Optical sensors, such as LiDAR and visual cameras, have become the mainstream perception solutions due to their high resolution and precision. However, these optical sensors suffer a sharp decline in performance or even complete failure when faced with visually degraded environments (VDEs) such as rain, snow, fog, and dust storms. The scattering and absorption of light beams lead to sparse point clouds and blurred images, posing a significant challenge to the safe operation of the system. To overcome this limitation, millimeter-wave radar technology, with its unique physical characteristics, is attracting widespread attention from academia and industry. Millimeter-wave radar's millimeter waves (wavelengths of 1mm-10mm) can effectively penetrate tiny particles such as rain, snow, and fog, enabling stable detection in all weather conditions. Furthermore, millimeter-wave radar can directly measure target velocity information (Doppler effect), making it an ideal supplement or alternative to optical sensors in adverse weather conditions.

[0003] However, despite the all-weather advantage of millimeter-wave radar, its application still faces two major bottlenecks. First, there's the sparsity of the point cloud. Limited by antenna aperture and bandwidth, the angular resolution of commercial single-chip radar is far lower than that of lidar. This results in extremely sparse point clouds, making it difficult to form detailed environmental contours suitable for positioning and mapping. Second, there's high noise and multipath interference. Millimeter waves are highly susceptible to multipath reflection when propagating in environments (especially indoors or canyons). This causes radar to detect a large number of "false positives" (outliers) at locations outside the true range. Traditional radar signal processing methods, such as constant false alarm rate (CFAR) detectors, while able to extract targets from the raw signal, produce sparse and noisy point clouds, far from meeting the requirements of high-precision navigation.

[0004] Therefore, a method is needed to improve the overall performance of millimeter-wave radar and solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides the following technical solution: an optimization method for a sparse point cloud diffusion model of millimeter-wave radar, comprising: Step 1: Based on the Diffusion Models, the baseline model uses a weighted sum of mean squared error and perceptual loss as the loss function. Step 2: Adopt a two-stage fine-tuning strategy to liberate the model's ability to generate structures by adjusting the balance point of the loss function without sacrificing the powerful denoising ability already learned by the pre-trained model. Step 3: The radar sensing framework constructs a direction-dependent uncertainty model based on the physical characteristics of radar measurements using an adaptive error elliptic filter, modeling each radar reflection point as an elliptical region with a directional error distribution.

[0006] Preferably, in step 1, the loss function is: (1) (2) In equations (1) and (2), the weights are configured as follows: , To calculate the predicted value, It is a true value.

[0007] More preferably, in step 2, the two-stage finetuning strategy specifically includes: Phase 1, Pre-training: First, the model is pre-trained using the baseline model configuration, and checkpoints are saved; the model has already been... Under a conservative strategy, it fully learned the macro structure of the data and its basic denoising capabilities.

[0008] Phase Two, Finetuning: Loading the checkpoints from Phase One and performing coordinated adjustments: First, resetting the optimizer without loading the pre-trained opt*.pt states; this allows the model to "lighten its load" under the gradient of the new loss function, avoiding being pulled back to the local minimum of a "conservative strategy" by the momentum of the old optimizer. Second, significantly reducing and fixing the learning rate to 1e-5, a crucial "stabilizer" that ensures the model can smoothly transition when facing the gradient of the new loss function, rather than collapsing due to gradient shock.

[0009] More preferably, in stage two, the weights are adjusted to and By increasing the LPIPS weight from 0.2 to 0.4 and making it the main driving force in the Finetuning stage, the model is forced to prioritize addressing the "structural incompleteness" problem. The MSE weight is set to 0.6, acting as a "balance" or "anchor" to prevent the model from generating overly "aggressive" "ghost point clouds" to meet LPIPS requirements, thus ensuring that the positional accuracy of generated points does not decrease excessively.

[0010] Preferably, in step 3, the final decision criterion for the filter is: For each point in the point cloud First, utilize its distance Calculate the adaptive threshold and minimum number of support points Subsequently, by constructing KD Trees are used to efficiently retrieve neighboring points and compute their support sets. The size, that is ; A point is considered a structural point and is retained if and only if the size of the support set of a point satisfies equation (8): (8) Otherwise, the point is considered an outlier and is removed.

[0011] More preferably, the threshold for: (6) Where T0 is the basic threshold, This is a distance-dependent relaxation factor.

[0012] Minimum support points for: (7) in, It is the minimum number of support points required near the origin. It is the decay rate coefficient. It is the minimum support limit set to prevent excessive decay.

[0013] The beneficial effects of this invention are: 1. This invention achieves significant improvements in point cloud density, integrity, and overall accuracy. In particular, it achieves orders-of-magnitude optimization in CD and F-score metrics, demonstrating the superiority of the method in processing sparse radar point clouds.

[0014] 2. This invention uses visualization analysis to intuitively demonstrate the impact of different strategies on point cloud generation results, providing valuable reference for subsequent research. Attached Figure Description

[0015] Figure 1 This is a diagram showing the overall structure of the diffusion model for the millimeter-wave radar sparse point cloud diffusion model optimization method of the present invention. Figure 2 This is a schematic diagram illustrating the working principle of the MSE loss function of the present invention. Figure 3 The diagram shows the MSE+LPIPS loss function of this invention. Figure 4 This is a mathematical modeling diagram of the error ellipsoid and error ellipse of the present invention; Figure 5 This is an ellipse diagram showing the error as a function of distance in this invention; Figure 6 This is a scene visualization diagram of Outdoors-run9-717 of the present invention. Detailed Implementation

[0016] The related technologies of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figures 1-6 As shown, this implementation is based on diffusion models, with its core idea derived from non-equilibrium thermodynamics, and includes two key processes: a forward noise-adding process and a reverse noise-reducing process. In the forward process, a real data sample... exist Gaussian noise is gradually added over time steps until it becomes pure isotropic Gaussian noise. The reverse process involves training a deep neural network (typically a U-Net architecture) that learns how to perform tasks at a given time step. and noisy images Under these conditions, predict the recovery Key information. This is obtained from random noise during the inference phase. Initially, by iteratively applying this trained denoising network, the model can gradually "sculpt" new, clear samples that conform to the original data distribution. The overall structure of the diffusion model proposed in this paper is as follows: Figure 1 As shown.

[0018] The baseline model in this implementation uses a weighted sum of mean squared error (MSE) and perceptual loss (LPIPS) as the loss function, i.e.: (1) Its weight configuration is as follows .

[0019] This configuration is the root cause of the model generating a large number of "false negatives" and high HD metrics in sparse outdoor scenes. Radar bird's-eye view (BEV) images are essentially highly sparse binary images, with over 99% of the pixels being background (value 0) and very few pixels being foreground (value 1). The MSE loss function (such as...) Figure 2 (As shown) The error is penalized equally for all pixels. This loss function calculates the predicted value. with truth value The sum of squared differences between pixels is defined as follows: (2) exist Under the strong influence of this, the model adopts a "conservative strategy" to minimize the total loss, that is, it tends to predict the background (0), thereby avoiding the risk of making mistakes on 99% of the pixels. This mechanism severely inhibits the model's ability to generate sparse foreground points, resulting in low recall and high HD (GT -> Pred) metrics.

[0020] This implementation proposes a two-stage fine-tuning strategy. Its core idea is to "liberate" the model's ability to generate structures by adjusting the balance point of the loss function without sacrificing the powerful denoising capabilities already learned by the pre-trained model. Stage one (pre-training) first uses a baseline configuration... Pre-train the model (e.g., 330,000 steps) and save checkpoints. The model has already... The conservative strategy fully learns the macroscopic structure of the data and the basic denoising capabilities. Phase Two (Finetuning) loads the checkpoints from Phase One and performs a series of coordinated adjustments. First, the optimizer is reset without loading the pre-trained opt*.pt states, allowing the model to "lighten its load" under the gradient of the new loss function, avoiding being pulled back to the local minimum of the "conservative strategy" by the momentum of the old optimizer. Second, the learning rate is significantly reduced and fixed at 1e-5, a crucial "stabilizer" that ensures the model can smoothly transition when facing the gradient of the new loss function, rather than collapsing due to gradient shock.

[0021] The most crucial adjustment lies in rebalancing the loss weights: adjusting the weights to... and LPIPS (e.g.) Figure 3The image shown is a perceptual loss function that doesn't concern itself with absolute pixel-by-pixel position, but rather compares the "structural" and "textural" similarity of the image in the deep feature space. By increasing the LPIPS weight from 0.2 to 0.4, making it the main driver in the fine-tuning stage, the model is forced to prioritize addressing "structural incompleteness" (i.e., "missing points"). LPIPS "encourages" the model to generate points at perceptually reasonable locations to complete the contours of walls or obstacles. Meanwhile, the MSE weight is kept at 0.6, acting as a "balance" or "anchor." It prevents the model from generating overly "aggressive" "false positives" to satisfy LPIPS, ensuring that the precision of the generated points doesn't decrease excessively.

[0022] This series of adjustments (resetting the optimizer, low LR, 0.6 / 0.4 weights) worked together to bring the model to the "sweet spot" at finetuning 1000 steps, successfully achieving a new balance between recall (driven by LPIPS) and precision (constrained by MSE) that was far superior to the baseline.

[0023] In this embodiment's radar perception framework, a backend adaptive two-dimensional filter optimizes the point cloud data acquired from millimeter-wave radar sensors. Millimeter-wave radar point clouds are typically subject to noise interference and are sparse, making them prone to outliers. Noise interference is particularly severe in complex and dynamic outdoor environments, significantly impacting the performance of subsequent SLAM algorithms. To improve the robustness and structural reconstruction accuracy of millimeter-wave radar point clouds in complex outdoor scenarios, this embodiment proposes an Adaptive Elliptical Error Filter (AEEF). This filter constructs a direction-dependent uncertainty model based on the physical characteristics of radar measurements, modeling each radar reflection point as an elliptical region with a directional error distribution. Compared to traditional filtering methods based on Euclidean distance or isotropic Gaussian kernels, this embodiment, considering radar ranging and angle measurement errors and their variation with distance, achieves a physically consistent estimation of the local density of the point cloud.

[0024] Based on the accuracy characteristics of radar ranging and angle measurement, the measurement error is modeled as an approximation of a constant. The specific process is as follows: Figure 4 As shown, that is:

[0025] Since this implementation method mainly operates on a two-dimensional radar plane, elevation angle error is ignored. However, different radar models exhibit significant performance variations due to factors such as power, category, and operating frequency band. In particular, differences in angular resolution lead to substantial differences in azimuth error. For instance, a high-performance long-range millimeter-wave radar might have an azimuth error of only 0.5°, while a short-range, low-power radar with an angular resolution on the order of 10° would have an azimuth error of 5° or even higher. Therefore, for different radar models, it is necessary to consider radial error... as well as Set the parameters for separate debugging.

[0026] Projecting three-dimensional uncertainty onto After planarization, an anisotropic Gaussian kernel can be obtained (such as...). Figure 5 As shown), its spindle length is: (3) Where a represents the radial distance measurement error and b represents the displacement caused by the angular error in the tangential direction (azimuth direction).

[0027] For any radar point Defined in the local coordinate system centered at that point: (4) in, It is a unit vector in the radial direction. The tangential direction is perpendicular to it. For any neighboring point q, whether the square of its Mahalanobis distance in the elliptical coordinate system is within a pre-defined T is used to determine whether a structure is formed: (5) This implementation method uses the Mahalanobis distance threshold. T Minimum number of neighbors required for a valid support set The filter dynamically changes with distance r to adapt to the non-uniformity of point cloud density. The key parameters of this filter are defined as follows: 1. Adaptive threshold T( r The measurement model for distant points exhibits greater uncertainty, resulting in a sparser point cloud. To avoid accidental deletion, the statistical constraints on distant points are appropriately relaxed. The threshold T is designed as a function that increases linearly with distance. (6) Here, T0 is a basic threshold, typically taken from two degrees of freedom. Critical values ​​distributed at a 95% confidence level (e.g., 5.991). This is a key hyperparameter in the adaptive error elliptic filter, known as the range-dependent relaxation factor. The physical characteristics of millimeter-wave radar dictate that the greater the distance, the larger the measurement error and the sparser the point cloud. Without this... (Right now =0): The algorithm uses strict near-field criteria to measure distant points. Due to the larger natural error of far-field points, they are easily outside the strict elliptical range and are incorrectly eliminated. This causes distant walls and obstacles to disappear completely (low recall).

[0028] join in : The larger the value, the faster the allowable error range expands, thus preserving points at greater distances. It controls the rate at which the filter's decision threshold increases with distance, addressing the issue of inconsistent distribution characteristics of radar point clouds at different distances. 2. Adaptive minimum support points The physical characteristics of radar scanning dictate that point cloud density decreases with increasing distance. Therefore, it is unreasonable to require far-field points to have the same number of neighbors as near-field points. This implementation minimizes the number of support points. Designed as a function that decays exponentially with distance: (7) in, It is the minimum number of support points required near the origin. It is the decay rate coefficient. It is the minimum support limit set to prevent excessive decay (usually 1).

[0029] In summary, the final decision criterion for the filter is as follows: for each point in the point cloud... First, utilize its distance Calculate the adaptive threshold and minimum number of support points Subsequently, by constructing KD Trees are used to efficiently retrieve neighboring points and compute their support sets. The size, that is A point is considered a structural point and is retained if and only if the size of its support set satisfies the following condition: (8) Otherwise, the point is considered an outlier and is removed. In this way, the filter can accurately remove sparse outliers that cause HD metrics to deteriorate, while preserving the true target structure at a distance to the maximum extent, thus achieving effective optimization of the back-endpoint cloud map.

[0030] Example The data in this embodiment comes from the publicly available Colorado millimeter-wave radar (ColoRadar) dataset, which consists of 52 sequences recorded in mines, built environments, and urban creeks, totaling over 145 minutes of 3D FMCW radar, 3D lidar, and IMU data. The sensor equipment is shown in Table 1.

[0031]

[0032] To ensure a fair comparison with baseline work, the data preprocessing workflow fully follows the publicly released methods of the baseline work. This workflow primarily involves processing raw radar data into a range-azimuth heatmap (RAH) and converting synchronized LiDAR point clouds into sparse bird's-eye view (BEV) images through voxelization projection, ultimately generating RAH-BEV image pairs for model training. The dataset partitioning in this embodiment strictly follows the settings in the configuration file, with the training set consisting of 25 sequences and the test set consisting of 25 sequences. These sequences are derived from seven different environments in the ColoRadar dataset (longboard, aspen, outdoors, arpg_lab, ec_hallways, edgar_classroom, edgar_army). The training and test sets use different run sequences from the same environment to ensure evaluation independence.

[0033] To comprehensively evaluate the quality of the generated point cloud, this embodiment uses the following three core metrics to measure the predicted point cloud from different dimensions. With truth point cloud Consistency between them, that is: Hausdorff Distance (HD): This metric measures the maximum error in matching two point sets and is extremely sensitive to both "false positives" and "false negatives" in the point cloud. It is defined as follows: (9) The level of the HD index (especially) The component (i.e., the maximum distance from the true value to the prediction) is the most critical indicator for measuring whether the "missing point" bottleneck in this embodiment has been resolved.

[0034] Chamfer Distance (CD): This metric calculates the sum of the bidirectional average minimum distances between two point sets, used to measure the overall average accuracy of the generated point cloud and its fit to the ground truth point cloud. It is defined as follows: (10) F-score: This metric combines precision and recall to measure the balanced performance of a model on sparse point cloud classification tasks. It is set within a predetermined distance threshold. Inside: Precision: The percentage of predicted points that are considered "correct" (i.e., ...). The proportion of points with true values ​​within a certain distance. It is defined as: (11) Recall: The percentage of truth values ​​that the model "successfully predicted" (i.e., 100% of the total number of truth values). The proportion of points within a given distance (where there are predicted points). It is defined as: (12) The improvement in the F-score is a crucial indicator for verifying that the binary loss function in this embodiment achieves a better balance between improving recall (addressing missing points) and maintaining precision (avoiding the introduction of excessive noise). It is defined as: (13) All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 5090 GPU. The model was based on the PyTorch framework and accelerated using CUDA 12.8. Regarding training parameters, to address the data loading bottleneck and ensure stability under Windows conditions, this embodiment ultimately adopted a configuration of num_workers=0 (loaded by the main process) and global_batch_size=48. This embodiment employs a two-stage training strategy: In the first stage (pre-training), the model was trained for 330,000 steps using the baseline loss function (MSE+LPIPS) and an initial learning rate of 1e-4.

[0035] In the second stage (Finetuning), 330,000 checkpoints are loaded, and a reset optimizer is used (without loading opt*.pt states), switching to the binary balanced loss function proposed in this embodiment. To avoid overfitting and performance degradation under the new loss function, the initial learning rate in the Finetuning stage is significantly reduced and fixed at 1e-5 (without using a learning rate annealing strategy). The final loss function weight configuration is as follows: The balanced configuration, while ensuring that the overall correctness is not compromised, stimulates the model to generate points to compensate for missing points by increasing the proportion of LPIPS.

[0036] Furthermore, this embodiment notes that different sensors (such as radar) have different physical error characteristics (such as ranging error). and angle measurement error The differences are significant. For example, the error parameters used in the literature for long-range (400m) radar (such as...) This approach is not applicable to the short-range (16m), low-resolution radar used in this project. Therefore, any component that relies on a physical model must treat its error parameters as task-specific hyperparameters and recalibrate them experimentally to find optimal values, rather than directly applying irrelevant literature parameters. Ultimately, based on the characteristics of the radar used, an angular error of 8° and a radial error coefficient of 0.1 were selected.

[0037] Based on the above settings, this embodiment, to verify the effectiveness of Finetune, quantitatively compares it with several mainstream radar point cloud denoising methods on two outdoor scenes (outdoors and Longboard) on the Coloradar dataset, covering RPDNet, OS-CFAR, CFAR, and baseline methods such as EDM and CD. In particular, the baseline methods EDM and CD are analyzed. Table 2 shows the data for each method under the three detection standards of CD, HD, and F-Score.

[0038]

[0039] As shown in Table 2, this embodiment achieves orders of magnitude improvements in CD and F-score compared to traditional methods such as RPDNet, OS-CFAR, and CFAR, demonstrating the overwhelming advantage of the diffusion-based generative approach in point cloud density and integrity. While traditional methods achieve good CD and HD metrics through extremely aggressive point deletion, their F-score performance, considering both recall and precision, is only one-twentieth of that of the diffusion model.

[0040] CD and F-score (SOTA): In both scenarios, the finetuning strategy of this embodiment achieved state-of-the-art (SOTA) performance. In the Outdoors scenario, CD decreased by 21.6% (2.371 -> 1.858), and F-score increased by 13.5% (0.266 -> 0.302). In the Longboard scenario, CD decreased by 18.0% (5.469 -> 4.486). This strongly demonstrates that the "rebalancing" strategy of this invention (MSE=0.6, LPIPS=0.4) is far superior to the baseline conservative strategy (MSE=0.8, LPIPS=0.2) in improving the overall accuracy and structural integrity of the point cloud.

[0041] HD (Trade-off Analysis): In the Outdoors scenario, the HD metric (7.537m) is slightly worse than the baseline (7.158m). However, it must be pointed out that the low HD of the baseline model is an "illusion." The baseline model (MSE=0.8) generates almost no points through a "conservative strategy." Although it has no "outliers" (FP), its point cloud is filled with large areas of "missing points" (FN), resulting in extremely poor CD and F-scores. This invention completes these "missing points" through LPIPS-driven completion. Although a small number of "ghost points" are introduced in the worst case (HD), causing a slight increase in the metric, this is traded for a significant improvement in overall performance.

[0042] This embodiment conducted ablation experiments as shown in Table 2. Starting with EDM(330k) (the baseline model reproduced in this embodiment, CD 2.016, F-score 0.285), this model suffers from severe "missing points" (FN) due to the high MSE weights. First, the Finetune(330k+1k) strategy was validated, i.e., using... The rebalancing weights were fine-tuned over 1,000 steps. The results showed a significant success: in the Outdoors scene, the CD was significantly reduced by 7.5% to 1.864, and the F-score improved by 6.0% to 0.302, proving that the loss rebalancing strategy effectively fills in the "missing points." Secondly, the necessity of the "sweet spot" was verified. When the finetuning was extended to 4,000 steps in this embodiment (Finetune(330k+4k)), the performance deteriorated across the board (CD 2.100, F-score 0.282). This clearly demonstrates that the model begins to overfit to LPIPS after 1,000 steps, i.e., generating "ghost point clouds" (FP) to "appear structurally complete." Therefore, 1,000 steps is a necessary "early stopping" strategy. Finally, the necessity of the AEEF filter was verified. Finetune(330k+1k)+AEEF is the final complete solution in this embodiment. AEEF, as a back-end filter, successfully "cleans up" the FP noise introduced during the Finetuning stage to address FN. In the Outdoors scene, CD further decreased to 1.858, and HD decreased to 7.537; in the Longboard scene, F-score even improved to 0.143. This experiment strongly demonstrates the completeness and necessity of the collaborative framework of the front-end (Finetuning 1k) and back-end (AEEF) in this embodiment. The ablation experiment results of different modules are shown in Table 3.

[0043]

[0044] To demonstrate the impact of different strategies on SLAM, frame 717 of Outdoors-run9 was selected as a typical case for visualization analysis, such as... Figure 6 As shown, Figure 6 (a)GT; Figure 6 (b) EDM (330k); Figure 6 (c) Finetune(330k+1k) Figure 6 (d) Finetune(330k+4k). This figure shows the ground truth (GT) of Outdoors-run9-717, the graph generated by the EDM(330k) pre-trained model, the graph generated by the Finetune(330k+1k) model, and the graph generated by the Finetune(330k+4k) model. Purple is used as the background, bright yellow as the foreground, and cyan as the foreground prediction blur zone. Figure 6 It can be observed that the image generated by the 330k pre-trained model has missing points at the top right corner, causing breaks in the continuous wall surface and right-angle corner structure. This is an unavoidable problem when pre-training is dominated by the MSE loss function. However, after training for 1000 steps with a strengthened LPIPS loss function, the problem of broken corner structures was resolved, and a more balanced result was achieved with the MSE loss function. In comparison... Figure 6 (d) The LPIPS loss function in the Finetune(330k+4k) model generated the graph, leading to overfitting. This caused continuous structures to be generated where there were no wall structures, ultimately resulting in performance degradation. It is demonstrated that the Finetune(330k+1k) model achieved the optimal performance and balance.

[0045] In summary, this invention addresses the "missing point" bottleneck of baseline diffusion models in outdoor sparse radar point cloud generation tasks by proposing a front-end (Finetune) and back-end (AEEF) collaborative optimization framework. At the front end, this invention diagnoses that the high MSE weights of the baseline are the root cause of the "missing point" and proposes a "loss rebalancing fine-tuning" strategy: based on a 330k-step pre-trained model, a reset optimizer and a low learning rate of 1e-5 are used, and the loss weights are adjusted to... In the backend, this invention introduces an AEEF filter to accurately remove "ghost points" (FPs) introduced during frontend fine-tuning to improve recall. This filter is based on mathematical modeling of the error ellipsoid and error ellipse, and uses a set reasonable threshold to finely filter the point cloud generated by the frontend. Specifically, the AEEF filter first constructs an error ellipsoid model based on the physical error characteristics of the radar, such as ranging and angle measurement errors. This model accurately describes the possible location range of each point in the radar point cloud, providing a theoretical basis for subsequent filtering. Next, the filter further analyzes each point using the error ellipse model, calculating the distance between the point and the center of the error ellipse, as well as the distribution of the point within the error ellipse, to determine whether the point is a "ghost point." Points identified as "ghost points" are removed from the point cloud by the AEEF filter, resulting in cleaner and more accurate point cloud data. Through the synergistic effect of the frontend "loss rebalancing fine-tuning" strategy and the backend AEEF filter, this invention successfully solves the "missing point" bottleneck problem of the baseline diffusion model in outdoor sparse radar point cloud generation tasks, significantly improving the overall accuracy and structural integrity of the point cloud.

[0046] It should be emphasized that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. An optimization method for a sparse point cloud diffusion model of millimeter-wave radar, characterized in that, include: Step 1: Based on the Diffusion Models, the baseline model uses a weighted sum of mean squared error and perceptual loss as the loss function. Step 2: Adopt a two-stage fine-tuning strategy to liberate the model's ability to generate structures by adjusting the balance point of the loss function without sacrificing the powerful denoising ability already learned by the pre-trained model. Step 3: The radar sensing framework constructs a direction-dependent uncertainty model based on the physical characteristics of radar measurements using an adaptive error elliptic filter, modeling each radar reflection point as an elliptical region with a directional error distribution.

2. The optimization method for a sparse point cloud diffusion model of millimeter-wave radar according to claim 1, characterized in that, In step 1, the loss function is: （1） （2） In equations (1) and (2), the weights are configured as follows: , To calculate the predicted value, It is a true value.

3. The method for optimizing a sparse point cloud diffusion model for millimeter-wave radar according to claim 2, characterized in that, In step 2, the two-stage finetuning strategy specifically includes: Phase 1, Pre-training: First, the model is pre-trained using the baseline model configuration and checkpoints are saved; Phase 2, Finetuning: Load the checkpoints from Phase 1 and perform coordinated adjustments: First, reset the optimizer and do not load the pre-trained opt*.pt states; second, significantly reduce the learning rate and fix it at 1e-5.

4. The optimization method for a sparse point cloud diffusion model of millimeter-wave radar according to claim 3, characterized in that, In the second stage, the weights are adjusted to and .

5. The optimization method for a sparse point cloud diffusion model of millimeter-wave radar according to claim 1, characterized in that, In step 3, the final decision criterion for the filter is: For each point in the point cloud First, utilize its distance Calculate the adaptive threshold and minimum number of support points ; Subsequently, by constructing KD Trees are used to efficiently retrieve neighboring points and compute their support sets. The size, that is ; A point is considered a structural point and is retained if and only if the size of the support set of a point satisfies equation (8): （8） Otherwise, the point is considered an outlier and is removed.

6. The method for optimizing a sparse point cloud diffusion model for millimeter-wave radar according to claim 5, characterized in that, The threshold for: （6） Where T0 is the basic threshold, For distance dependence relaxation factor; The minimum number of support points for: （7） in, It is the minimum number of support points required near the origin. It is the decay rate coefficient. It is the minimum support limit set to prevent excessive decay.

Images