4D millimeter wave radar point cloud enhancement and target detection method based on Bayesian fusion

Abstract

The invention provides a Bayesian fusion-based 4D millimeter wave radar point cloud enhancement and target detection method, and the method comprises the steps: 1, projecting a 4D millimeter wave radar point cloud to an image plane, carrying out the matching of an image instance segmentation result, and obtaining an image semantic category and image segmentation confidence; 2, constructing a class conditional probability density model based on the physical characteristics of the radar points; based on a Bayesian probability framework, carrying out weighted fusion on the image segmentation confidence and the class probability density to obtain fusion confidence; 3, taking the fusion confidence and semantic category codes as augmented features, and performing feature splicing on the augmented features and the 4D millimeter-wave radar point cloud to generate a semantic enhanced point cloud; and sampling the semantic enhanced point cloud by adopting a sampling method fusing confidence guidance, and inputting the sampled semantic enhanced point cloud into a detection network for training and reasoning. According to the method, the technical problems of unidirectional dependence, shallow geometric matching and insufficient utilization of radar physical characteristics in the existing image and radar point cloud fusion technology are solved.

Description

Technical Field

[0001] This invention relates to the technical field of environmental perception for autonomous driving, and in particular to a method for enhancing 4D millimeter-wave radar point clouds and detecting targets based on Bayesian fusion. Background Technology

[0002] Currently, environmental perception technologies for autonomous driving and intelligent transportation systems are developing towards multi-sensor fusion to address the challenges of complex scenarios and adverse weather conditions. Cameras provide high-resolution texture and color information, making them key sensors for achieving fine target classification and segmentation; however, their performance is susceptible to changes in lighting and weather. LiDAR can generate dense 3D geometric point clouds, but it is costly and its performance degrades in fog, rain, and dust. Traditional millimeter-wave radar offers advantages such as accurate velocity measurement and all-weather operation, but its sparse point clouds and lack of elevation information limit its perception dimensions. In recent years, 4D millimeter-wave radar has emerged, providing dense point clouds containing range, azimuth, elevation, and Doppler velocity, significantly improving target contour resolution and spatial information recognition, and providing rich and reliable information support for building high-performance fusion perception systems.

[0003] In image-radar point cloud fusion schemes, a typical existing technique is a geometric matching method based on image segmentation and radar point cloud projection. The basic process is as follows: First, a two-dimensional semantic mask of the target is extracted using an image segmentation network; second, a 4D millimeter-wave radar point cloud is unidirectionally projected onto the image plane using an extrinsic parameter matrix; finally, a determination is made solely based on spatial geometric positional relationships, i.e., directly determining whether the projected point is contained within the image mask region. If the containment relationship is satisfied, the camera-perceived category label is forcibly assigned to the corresponding radar point.

[0004] However, the core drawback of the aforementioned existing technologies lies in their unidirectional and shallow fusion architecture. On the one hand, this technology heavily relies on camera perception results and lacks robustness in handling image segmentation failures. Due to the "camera-centric, radar-assisted" fusion mode, any false detections or missed detections in image segmentation will propagate to the radar modality through the projection process, resulting in the radar point cloud being assigned incorrect semantic labels. On the other hand, this technology has low information utilization of 4D millimeter-wave radar. The fusion process only stays at the spatial geometric matching level and fails to effectively extract the deep semantic value of features such as radial velocity and radar cross-section, thus failing to effectively unleash the modal advantages of 4D radar in multidimensional physical perception. Summary of the Invention

[0005] The purpose of this invention is to provide a 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion, which solves the technical problems of unidirectional dependence, shallow geometric matching and insufficient utilization of radar physical features in existing image and radar point cloud fusion technologies.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion, comprising the following steps: Step 1: Project the 4D millimeter-wave radar point cloud onto the image plane and match it with the image instance segmentation results to obtain the image semantic category and image segmentation confidence of each radar point; Step 2: Based on the physical features of radar points in the 4D millimeter-wave radar point cloud, construct class-conditional probability density models for different semantic categories; map the physical features to the logarithmic space and perform standardization to obtain standardized log-likelihood values; in the logarithmic domain, use the image segmentation confidence as a prior term and perform weighted fusion with the standardized log-likelihood values ​​to obtain the fusion confidence of each radar point belonging to each semantic category; Step 3: The fusion confidence score and the corresponding semantic category encoding are used as augmenting features and concatenated with the 4D millimeter-wave radar point cloud to generate a semantically enhanced point cloud; the semantically enhanced point cloud is sampled using a key point sampling method guided by fusion confidence score, and the sampled point cloud is input into a 3D target detection network for training and inference to achieve target detection.

[0007] Furthermore, step one specifically includes: The radar point coordinates are transformed to the camera coordinate system using the extrinsic parameter matrix from the radar coordinate system to the camera coordinate system. By using the camera intrinsic parameter matrix, the radar point coordinates in the camera coordinate system are projected onto the image pixel coordinate system to obtain the projected pixel coordinates of the radar point on the image. Based on the binary mask of each semantic category in the image obtained by the instance segmentation network, it is determined whether the projected pixel coordinates fall within the binary mask region of each semantic category. If so, the image semantic category corresponding to the radar point is determined, and the segmentation confidence corresponding to the category is used as the image segmentation confidence.

[0008] Furthermore, in step two, the physical features include absolute radial velocity and radar cross-section; the construction of class-conditional probability density models for different semantic categories specifically includes: For each semantic category, based on the statistical distribution of the absolute radial velocity and radar cross-section of radar points belonging to that category in the training data, a Gaussian distribution model is used to parameterize and fit the class probability density of the absolute radial velocity and radar cross-section respectively.

[0009] Furthermore, in step two, within the logarithmic domain, the image segmentation confidence is used as a priori and weighted with the standardized log-likelihood value to obtain the fused confidence of each radar point belonging to each semantic category, specifically including: For radar points, they belong to semantic categories. Fusion confidence Its logarithmic form satisfy:

[0010] in, Segment the image with confidence; The velocity feature score is calculated and standardized based on the class conditional probability of absolute radial velocity; The radar cross-section feature score is calculated based on the prior probability of the radar cross-section and then standardized. and The fusion weights are obtained through training and satisfy the following conditions: + =1.

[0011] Furthermore, the velocity feature score The calculation formula is:

[0012] in, , The absolute radial velocity calculated based on the Gaussian distribution model. Class-conditional probability density; Indicates that the training set belongs to category The mean of the samples in the logarithmic probability space. This represents the corresponding standard deviation.

[0013] Furthermore, the radar cross-section feature score The calculation formula is:

[0014] in, , Radar cross-section calculated based on Gaussian distribution model The prior probability density; Indicates that the training set belongs to category The mean of the samples in the logarithmic probability space. This represents the corresponding standard deviation.

[0015] Furthermore, in step three, the keypoint sampling method guided by fusion confidence is an improved farthest point sampling method, and its sampling weighted distance metric formula is as follows:

[0016] in, This represents the minimum Euclidean distance from the current point to be examined to the existing set of key points. The preset semantic guidance weight coefficient, Normalized fusion confidence score (values ​​ranging from 0) 1), For the comprehensive weighted distance used for sampling, in each iteration, select The largest point is selected as the new key point.

[0017] Furthermore, the three-dimensional target detection network uses PV-RCNN as the detection framework, and the improved farthest point sampling method is used to improve the key point sampling strategy in the network.

[0018] Compared with the prior art, the present invention has at least the following beneficial effects: This invention constructs a unified Bayesian probabilistic fusion framework to deeply fuse the semantic confidence of image segmentation with the physical features (absolute radial velocity, radar cross-section) of 4D millimeter-wave radar. This core approach breaks the unidirectional dependency pattern of "camera-centric, radar-secondary" in existing technologies. By establishing statistical distribution models (such as Gaussian distribution) of radar physical features for different target categories and using them as a likelihood term for weighted fusion with the prior confidence of the image, this method achieves probabilistic cross-validation and error correction of image segmentation results. This means that when image segmentation suffers from misjudgment or reduced confidence due to factors such as illumination or occlusion, the physical evidence of the radar modality can provide effective correction through the likelihood term, thereby significantly improving the overall robustness of the multi-sensor fusion system in complex scenarios.

[0019] This invention fully utilizes the unique multidimensional physical information of 4D millimeter-wave radar. Traditional methods only use the spatial location of point clouds for geometric matching, while this invention delves into the statistical value of radial velocity and radar cross-section for category discrimination. By parametrically fitting, logarithmically transforming, and standardizing these features, standardized feature scores are generated that can be directly compared and fused with image confidence scores under a unified dimension. This extraction and utilization of the deep value of radar physical features fully releases the modal advantages of 4D millimeter-wave radar, achieving a leap from shallow data alignment to deep feature and decision fusion, effectively overcoming the bottleneck of low utilization rate of non-geometric information from radar in existing technologies.

[0020] This invention innovatively applies the "fusion confidence score" generated by fusion to the key point cloud sampling stage of the downstream 3D object detection network. The proposed "fusion confidence score-guided farthest point sampling method" introduces the fused multimodal confidence score as a weight into the sampling distance metric, enabling the sampling process to prioritize the retention of high-confidence foreground object points while maintaining spatial coverage, thus solving the problem of easily losing distant, sparse, small objects. Experimental data shows that on the VoD dataset, after applying this method to the PV-RCNN framework, its average detection accuracy (mAP) in the entire annotation region and region of interest is improved from 44.637% and 68.657% to 54.027% and 76.213%, respectively, effectively verifying the significant effect of this method on improving the accuracy of object detection, especially difficult object detection. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This embodiment provides a flowchart of a 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion. Figure 2 This embodiment provides a farthest point sampling method based on fusion confidence. Figure 3 The method provided in this embodiment is based on the test evaluation results of the VoD dataset; Figure 4 The method provided in this embodiment is based on the visualization results of the VoD dataset. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] This embodiment provides a 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion. Please refer to... Figure 1 As shown, this method comprises three core steps: First, the 4D millimeter-wave radar point cloud is projected onto the image plane through coordinate transformation (using extrinsic and intrinsic parameter matrices), and then matched with the image semantic mask and confidence obtained through the instance segmentation network, thereby assigning a preliminary image semantic category and confidence to each radar point.

[0025] Secondly, based on the physical characteristics of the radar points themselves (absolute radial velocity, radar cross-section), statistical probability models (class conditional probabilities) are constructed for different categories (such as cars and pedestrians). Under the Bayesian framework, this model is used as the "likelihood" and weighted and fused with the image confidence score from the previous step (as the "prior") to calculate the final "fusion confidence score" of each radar point belonging to each category.

[0026] Finally, these confidence scores and category codes are used as augmented features and concatenated with the 4D millimeter-wave radar point cloud to form a "semantically enhanced point cloud". Before performing 3D target detection on this point cloud, an improved sampling method is used, namely, fusing confidence-guided farthest point sampling, to select key points, which are then input into 3D detection networks such as PV-RCNN to complete training and inference.

[0027] In summary, this method constructs a complete deep fusion process based on probabilistic reasoning, unifying image semantics and radar physical features at the decision-making level, rather than simply geometric matching or feature stitching. It addresses the problems of existing technologies such as over-reliance on single sensors, poor robustness, and insufficient utilization of deep radar information. By achieving cross-validation between sensors through probabilistic fusion, it improves the overall accuracy and reliability of target detection in complex scenarios.

[0028] I. This invention is specifically implemented through the following methods: Appendix Figure 1 The overall flowchart of this invention is shown, which mainly includes the following three steps, as detailed below: Step 1: Achieve fusion and matching of radar point cloud and image segmentation data through coordinate mapping and mask filtering. Consider a frame of 4D millimeter-wave radar point cloud, which is a collection of multiple radar points. We use... Let represent any one of the radar points, where This represents the coordinate vector of the radar point in the radar coordinate system. Indicates radar cross-section, Represents relative radial velocity, Indicates absolute radial velocity. This represents the fusion timestamp of the point cloud. (This is used to merge radar points.) Matching with the image mask requires transforming it from the radar coordinate system to the camera coordinate system. To satisfy matrix multiplication operations, the extrinsic parameter matrix obtained from the calibration of the radar to the camera is used. And the coordinate vector of the radar point in the radar coordinate system. Represented as a homogeneous coordinate vector, i.e. (superscript) (This indicates transpose). Finally, the coordinate vector of the radar point in the camera coordinate system. , can be represented as

[0029] Obtain the coordinate vector of the radar point in the camera coordinate system. Then, through the camera intrinsic parameter matrix Projecting onto the pixel coordinate system. This process involves two steps: first, calculating the homogeneous pixel coordinate vector of the normalized camera coordinates. It can be represented as

[0030] Final radar point pixel coordinates on the image It can be represented as

[0031] Finally, the input image is processed by a pre-trained deep learning instance segmentation network to obtain pixel-level prediction results. The instance segmentation network is configured to identify multiple preset target categories (cars, pedestrians, and cyclists), and output a binary mask corresponding to each target category. and split confidence .

[0032] The projected pixel coordinates calculated based on the aforementioned steps Determine whether the projected pixel coordinates fall within the binary mask region of each semantic category: if satisfied... Then the radar point is determined. Category And the corresponding image segmentation confidence score This serves as a reliable feature for image segmentation of the radar point.

[0033] Step 2: Construct a category prior model and perform Bayesian probability fusion. For each 4D millimeter-wave radar point projected into the segmented mask region, this invention constructs a conditional probability density model based on its physical observations, including absolute radial velocity. With radar cross-section By statistically analyzing the physical quantities of different categories of radar points, a statistical distribution model of each category in terms of physical characteristics is established. In this invention, it is observed that... and The distribution exhibits a unimodal characteristic, therefore a Gaussian distribution model is used for parameter fitting.

[0034] Based on the above analysis, the categories are defined. absolute radial velocity The conditional probability density function on is , can be represented as

[0035] in, and These are the categories that belong to in the training data. The statistical mean and standard deviation of the absolute radial velocities of all radar points were obtained through offline statistics. It is the standard mathematical symbol for an exponential function.

[0036] Similarly, define categories In radar cross-section The prior probability density function on is , can be represented as

[0037] in, , These are the categories that belong to in the training data. The statistical mean and standard deviation of the radar cross-section of all radar points.

[0038] To address the challenge of directly weighting image segmentation confidence scores with the probability densities of radar physical features, this invention employs a logarithmic normalization method to map the probability densities of physical features to a unified metric space. Specifically, the logarithm of the conditional probabilities of the absolute radial velocity and radar cross-section of the radar point is first taken to obtain the log-likelihood value of the absolute radial velocity. Log-likelihood of radar cross-section , can be represented as

[0039] Subsequently, the standardized velocity feature score is transformed using pre-statistical distribution parameters to eliminate dimensional differences. , can be represented as

[0040] in, Indicates that the training set belongs to category The mean of the samples in the logarithmic probability space. This represents the corresponding standard deviation.

[0041] Similarly, the standardized radar cross-section feature score , can be represented as

[0042] in, Indicates that the training set belongs to category The mean of the samples in the logarithmic probability space. This represents the corresponding standard deviation.

[0043] The image segmentation confidence score obtained in step one is weighted and fused with the standardized physical feature scores to obtain the logarithmic form of the final fused confidence score. , can be represented as

[0044] in The category output by the image segmentation network in step one. Confidence level, in logarithmic form As a prerequisite for fusion. and The fusion weights for absolute radial velocity and radar cross-section, respectively, are obtained through training and satisfy the normalization condition in this invention.

[0045] Finally, the logarithmic fusion results The exponential operation is performed to obtain the final fusion confidence score. , can be represented as

[0046] The fusion confidence score indicates the category to which the radar point projected onto the instance belongs after integrating the radar physical characteristics. The estimated value.

[0047] Step 3: Generate fused and enhanced point clouds and perform 3D network training and target detection. The fusion confidence score, which characterizes the multi-sensor fusion result, calculated in step two, is used as the basis for this calculation. and semantic category encoding As an augmentation feature, it is stitched with the 4D millimeter-wave radar point cloud to construct an enhanced fused point cloud. Specifically, for any point in the enhanced point cloud... , can be represented as

[0048] The enhanced point cloud contains multi-dimensional information including geometry, motion, and deep semantics. It is fed as input data into a 3D object detection network for feature extraction and object regression.

[0049] In the specific implementation of training and inference of the 3D detection network, this invention selects PV-RCNN as the detection framework. To further improve detection performance, especially the feature extraction capability for sparse small targets, this invention improves the keypoint sampling strategy in the 3D detection network and proposes a farthest point sampling method guided by fused confidence. This method considers not only the spatial distance of points during sampling but also introduces a fused confidence level. As weights, these ensure that the sampled keypoints maintain a uniform spatial distribution while also tending to be distributed in regions with clear semantic information and high credibility. (See attached image) Figure 2 The specific process is as follows: 1. Algorithm Input The algorithm's input includes: the fused and enhanced set of point cloud coordinates. Each point contains three-dimensional spatial coordinates and other physical attributes; the fusion confidence score, calculated from the preceding steps and corresponding to each point, is... ; and the preset number of sampling key points .

[0050] 2. Initialization First, initialize an empty set of key points. A set for recording each point in the point cloud. minimum distance array (Initially infinity), and an array to mark whether a point has been visited. Then iterate through the input point cloud, select the point with the highest fusion confidence score as the first keypoint, and add it to the set. and update the distance array. .

[0051] 3. Weighted iterative sampling In each subsequent round of sampling, the minimum Euclidean distance from the remaining points to the current set of keypoints is calculated. This invention constructs the following weighted distance metric formula, expressed as follows:

[0052] in, Normalized fusion confidence score (values ​​ranging from 0) 1), These are semantically guided weighting coefficients used to adjust the degree of influence of confidence on sampling. This serves as a comprehensive indicator for final site selection, replacing the single geometric distance evaluation indicator.

[0053] According to the above formula, it has a high confidence level, that is... Radar points with larger values ​​will have a larger weighted distance calculated. Therefore, it is selected first under the farthest point sampling mechanism.

[0054] 4. Output Repeat the above calculations and select The largest radar point is added to the set. Continue until the number of sampling points reaches the preset value. The final set of keypoints, which balances spatial coverage and target semantic information, will serve as input for feature interpolation and region of interest extraction in subsequent PV-RCNN processes.

[0055] II. Testing and Evaluation of the Method Proposed in the Invention The method proposed in this invention will now be tested: All test data is based on the VoD (View-of-Delft) dataset, which contains 8693 frames. The dataset is divided into training, validation, and test sets at a ratio of 59%, 15%, and 26%, respectively. Annotated targets include pedestrians, cyclists, and cars. During training, training and validation were performed on three categories: cars, pedestrians, and cyclists. On the VoD dataset, mean accuracy (mAP) was chosen as the evaluation metric to test the performance of various existing object detection algorithms across the entire annotated region and the region of interest. The entire annotated region refers to a 90° fan-shaped area 50m in front of the camera, while the region of interest refers to a specific area in the driving space near the front of the vehicle within the camera coordinate system. .

[0056] The tests were conducted using the PyTorch framework, trained and evaluated on hardware equipped with four 4090 GPUs and an Intel i7-12700K CPU. The network was trained for 80 epochs using the Adam optimizer, with a mini-batch size of 4. The initial learning rate was 0.001, dynamically adjusted using a OneCycle strategy. The keypoint sampling number was set to 512.

[0057] Appendix Figure 3 The evaluation results of this invention on the VoD dataset are presented and compared with other existing object detection algorithms. The results show that this invention demonstrates a significant advantage over the baseline model PV-RCNN, improving the mAP value from 44.637% and 68.657% to 54.027% and 76.213%, respectively. (Appendix) Figure 4 The visualization results of the present invention on the VoD dataset are shown. As can be seen from the figure, the present invention can accurately identify the target in the scene in complex road scenarios.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enhancing 4D millimeter-wave radar point clouds and detecting targets based on Bayesian fusion, characterized in that, Includes the following steps: Step 1: Project the 4D millimeter-wave radar point cloud onto the image plane and match it with the image instance segmentation results to obtain the image semantic category and image segmentation confidence of each radar point; Step 2: Based on the physical features of radar points in the 4D millimeter-wave radar point cloud, construct class-conditional probability density models for different semantic categories; map the physical features to the logarithmic space and perform standardization to obtain standardized log-likelihood values; in the logarithmic domain, use the image segmentation confidence as a prior term and perform weighted fusion with the standardized log-likelihood values ​​to obtain the fusion confidence of each radar point belonging to each semantic category; Step 3: The fusion confidence score and the corresponding semantic category encoding are used as augmenting features and concatenated with the 4D millimeter-wave radar point cloud to generate a semantically enhanced point cloud; the semantically enhanced point cloud is sampled using a key point sampling method guided by fusion confidence score, and the sampled point cloud is input into a 3D target detection network for training and inference to achieve target detection.

2. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to claim 1, characterized in that, Step one specifically includes: The radar point coordinates are transformed to the camera coordinate system using the extrinsic parameter matrix from the radar coordinate system to the camera coordinate system. By using the camera intrinsic parameter matrix, the radar point coordinates in the camera coordinate system are projected onto the image pixel coordinate system to obtain the projected pixel coordinates of the radar point on the image. Based on the binary mask of each semantic category in the image obtained by the instance segmentation network, it is determined whether the projected pixel coordinates fall within the binary mask region of each semantic category. If so, the image semantic category corresponding to the radar point is determined, and the segmentation confidence corresponding to the category is used as the image segmentation confidence.

3. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to claim 1, characterized in that, In step two, the physical features include absolute radial velocity and radar cross-section; the construction of class-conditional probability density models for different semantic categories specifically includes: For each semantic category, based on the statistical distribution of the absolute radial velocity and radar cross-section of radar points belonging to that category in the training data, a Gaussian distribution model is used to parameterize and fit the class probability density of the absolute radial velocity and radar cross-section respectively.

4. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to claim 3, characterized in that, In step two, within the logarithmic domain, the image segmentation confidence is used as a priori and weighted with the standardized log-likelihood value to obtain the fused confidence of each radar point belonging to each semantic category. Specifically, this includes: For radar points, they belong to semantic categories. Fusion confidence Its logarithmic form satisfy: ； in, Segment the image with confidence; The velocity feature score is calculated and standardized based on the class conditional probability of absolute radial velocity; The radar cross-section feature score is calculated based on the prior probability of the radar cross-section and then standardized. and The fusion weights are obtained through training and satisfy the following conditions: + =1.

5. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to claim 4, characterized in that, The velocity feature score The calculation formula is: ； in, , The absolute radial velocity calculated based on the Gaussian distribution model. Class-conditional probability density; Indicates that the training set belongs to category The mean of the samples in the logarithmic probability space. This represents the corresponding standard deviation.

6. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to claim 4, characterized in that, Radar cross-sectional feature score The calculation formula is: ； in, , Radar cross-section calculated based on Gaussian distribution model The prior probability density; Indicates that the training set belongs to category The mean of the samples in the logarithmic probability space. This represents the corresponding standard deviation.

7. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to any one of claims 1-6, characterized in that, In step three, the keypoint sampling method guided by fusion confidence is an improved farthest point sampling method, and its sampling weighted distance metric formula is as follows: ； in, This represents the minimum Euclidean distance from the current point to be examined to the existing set of key points. The preset semantic guidance weight coefficient, Normalized fusion confidence score (values ​​ranging from 0) 1), For the comprehensive weighted distance used for sampling, in each iteration, select The largest point is selected as the new key point.

8. The 4D millimeter-wave radar point cloud enhancement and target detection method based on Bayesian fusion according to claim 7, characterized in that, The three-dimensional target detection network uses PV-RCNN as the detection framework and improves the key point sampling strategy in the network using the improved farthest point sampling method.

Images