A radar point cloud denoising and enhancement method based on adaptive filtering

Abstract

This invention discloses a radar point cloud denoising and enhancement method based on adaptive filtering, comprising the following steps: Step 1: acquiring an initial point cloud dataset; Step 2: preprocessing to generate a structured point cloud dataset; Step 3: performing point-by-point denoising and reflection intensity enhancement processing using an improved adaptive Kalman filter algorithm; Step 4: performing density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset using the density multi-scale sampling module of an improved PointNeXt network to generate a multi-scale fused feature tensor; Step 5: generating a joint geometry-reflection feature tensor through a residual feature extraction structure; Step 6: performing channel weighting and residual fusion, applying inter-frame geometric constraints and confidence consistency constraints to generate an enhanced feature tensor; Step 7: performing multi-head reconstruction to generate an enhanced point cloud dataset. This invention improves the denoising accuracy, reflection intensity restoration accuracy, and inter-frame consistency stability of radar point clouds.

Description

Technical Field

[0001] This invention relates to the fields of intelligent signal processing and 3D point cloud modeling technology, and in particular to a method for denoising and enhancing radar point clouds based on adaptive filtering. Background Technology

[0002] With the development of intelligent driving systems, mobile robots, and 3D environmental perception technologies, radar point cloud technology, as a low-cost and robust spatial perception technology, is widely used in various dynamic perception and recognition scenarios. However, raw radar point cloud signals often face the following key problems in complex environments:

[0003] Due to weather interference, reflection occlusion, and multipath effects, raw point cloud data contains a large amount of unstructured noise and void regions, resulting in uneven point cloud density distribution and blurred boundaries, which seriously affects the spatial geometric modeling and structural feature extraction of targets. Existing filtering methods mostly use fixed thresholds or regular templates, which are difficult to adapt to local changes in point clouds at different densities and scales, and lack the ability to adaptively preserve key structural regions. In the point cloud enhancement process, the information fusion processing between the reflection intensity channel and the geometric channel is insufficient. Current mainstream methods often ignore the dynamic adjustment mechanism of confidence and other weight information, which cannot fully improve the structural integrity and semantic expression ability of point cloud reconstruction. In continuous frame point cloud scenarios, the lack of effective temporal consistency modeling can easily cause spatial drift and inconsistency in confidence of the target structure between multiple frames, which restricts the stability and accuracy of subsequent perception and decision-making modules.

[0004] Therefore, how to provide a radar point cloud denoising and enhancement method based on adaptive filtering is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a radar point cloud denoising and enhancement method based on adaptive filtering. This invention utilizes an improved adaptive Kalman filter algorithm and an improved PointNeXt network, and describes in detail the acquisition of raw radar echo signals, confidence modeling of filtered point clouds, multi-scale feature fusion, reflection feature enhancement, and multi-task point cloud reconstruction. This invention has the advantages of strong noise resistance, high structural integrity preservation, stable reflection characteristic enhancement, and high point cloud reconstruction accuracy, and can significantly improve the geometric perception quality and target recognition reliability of radar point clouds in complex scenes.

[0006] A radar point cloud denoising and enhancement method based on adaptive filtering according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Collect the raw echo signals from the radar sensor to generate an initial point cloud dataset;

[0008] Step 2: Perform spatial coordinate alignment, reflection intensity normalization, and outlier removal on the initial point cloud dataset to generate a structured point cloud dataset;

[0009] Step 3: Perform point-by-point denoising and reflection intensity enhancement on the structured point cloud dataset using the improved adaptive Kalman filter algorithm to generate a filtered point cloud dataset;

[0010] Step 4: Construct an improved PointNeXt network. The improved PointNeXt network includes a density multi-scale sampling module, a residual feature extraction structure, a confidence adjustment module, a temporal consistency module, and a multi-head reconstruction module. The density multi-scale sampling module performs density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset to generate a multi-scale fused feature tensor.

[0011] Step 5: Extract joint features from the multi-scale fusion feature tensor using a residual feature extraction structure to generate a joint geometry-reflection feature tensor. The residual feature extraction structure incorporates reflection intensity information to construct the reflection feature tensor.

[0012] Step 6: Perform channel weighting and residual fusion on the joint geometry-reflection feature tensor through the confidence adjustment module, and perform inter-frame geometric constraints and confidence consistency constraints through the temporal consistency module to generate the enhanced feature tensor;

[0013] Step 7: Use the multi-head reconstruction module to predict the coordinate residuals, reflection enhancement values, and confidence scores of the enhanced feature tensor to generate an enhanced point cloud dataset.

[0014] Optionally, step one specifically includes:

[0015] Set the transmission parameters of the radar sensor and determine the transmission time. The transmission parameters include the transmission waveform type, pulse repetition frequency, bandwidth, and sampling rate.

[0016] The original echo signal received by the radar sensor is converted from analog to digital to obtain an echo discrete sequence. The echo discrete sequence consists of several sampling points, and each sampling point corresponds to a discrete time sampling value.

[0017] Waveform demodulation and down-conversion processing are performed on the discrete echo sequence to obtain a complex baseband echo sequence. The complex amplitude of each sampling point in the complex baseband echo sequence represents the echo amplitude and phase information.

[0018] Multiply the difference between the arrival time and the transmission time of each sampling point by the speed of electromagnetic wave propagation in air, and divide by two to calculate the polar coordinate distance of each sampling point.

[0019] Set a minimum distance threshold and a maximum distance threshold, perform distance threshold filtering on the polar coordinate distance of all sampling points, and retain only sampling points whose polar coordinate distance is greater than or equal to the minimum distance threshold and less than or equal to the maximum distance threshold to obtain a valid echo sampling point sequence;

[0020] Based on the azimuth and elevation angles of the radar array, the polar coordinate distance of each effective echo sampling point is converted into spatial rectangular coordinates to obtain the three-dimensional spatial coordinates of each effective echo sampling point.

[0021] Based on the complex baseband echo sequence, the square of the complex amplitude of each valid echo sampling point is taken as the reflection intensity;

[0022] The initial point cloud data for each valid echo sampling point is composed of three-dimensional spatial coordinates and reflection intensity. All initial point cloud data are aggregated to form an initial point cloud dataset.

[0023] Optionally, step two specifically includes:

[0024] A rigid transformation matrix is ​​set to convert the 3D spatial coordinates of all points in the initial point cloud dataset into homogeneous coordinates to obtain aligned spatial coordinates.

[0025] The reflection intensity values ​​of all points are normalized using the min-max normalization method to obtain normalized reflection intensity values.

[0026] Set a spatial neighborhood radius and a threshold for the number of neighborhood points, remove outliers from the initial point cloud dataset, and retain only valid echo sampling points within the spatial neighborhood radius whose number of neighboring points is greater than or equal to the threshold for the number of neighboring points, to obtain a sequence of valid neighborhood points;

[0027] The aligned spatial coordinates and normalized reflection intensity values ​​of all valid neighboring points are combined to form a structured point cloud dataset.

[0028] Optionally, step three specifically includes:

[0029] Setting the initial state of the filter specifically includes: initializing the predicted state vector, state covariance matrix, system noise covariance matrix, state transition matrix, and observation matrix;

[0030] Calculate the local density value of each neighborhood valid point in the structured point cloud dataset. The local density value is obtained by counting the number of neighbor points contained in a spatial neighborhood of a set radius and dividing by the volume of the spatial neighborhood.

[0031] Based on local density values, the system noise covariance matrix is ​​dynamically updated using an exponential decay method.

[0032] Define a system noise term, which follows a multivariate Gaussian distribution with a mean of zero and a covariance equal to the updated system noise covariance matrix;

[0033] Update the predicted state vector and the predicted covariance matrix. The predicted state vector is the product of the state transition matrix and the predicted state vector at the previous time step, plus the system noise term. The predicted covariance matrix is ​​the product of the state transition matrix and the state covariance matrix at the previous time step, plus the system noise covariance matrix.

[0034] Calculate the product of the predicted state vector and the observation matrix to obtain the predicted observation vector, and use the vector difference between the current observation vector and the predicted observation vector as the observation residual vector.

[0035] The square norm of the observed residual vector is mapped to the confidence score of the current neighborhood valid points through a negative exponential function;

[0036] The confidence gating function is set using the Sigmoid gating function, which maps the confidence scores of the current neighborhood valid points to the confidence gating coefficients.

[0037] Multiply the standard Kalman gain by the confidence gating coefficient to generate the confidence-gated Kalman gain;

[0038] The confidence-gated Kalman gain is multiplied by the observation residual vector and added to the predicted state vector to obtain the filtered state vector of the current neighborhood effective point.

[0039] Based on the filtered state vector, three-dimensional spatial coordinates and reflection intensity are extracted, and the three-dimensional spatial coordinates, reflection intensity and confidence score are combined to form a filtered point cloud data item. All filtered point cloud data items are aggregated to generate a filtered point cloud dataset.

[0040] Optionally, the step of performing density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset through the density multi-scale sampling module to generate a multi-scale fused feature tensor specifically includes:

[0041] Define a multi-scale neighborhood set, which includes neighborhood radii of M scales;

[0042] Calculate the local density value of each filtered point cloud data item in the multi-scale neighborhood set;

[0043] For each filtered point cloud data item, the reciprocal of the local density value at each scale is taken to obtain the reciprocal density value at each scale, and the reciprocal density values ​​at all scales are normalized to obtain the neighborhood sampling weight at each scale.

[0044] Each filtered point cloud data item is used as the center point, and weighted K-nearest neighbor sampling is performed based on the neighborhood radius at each scale to obtain the neighborhood point set of M scales for each filtered point cloud data item.

[0045] For each scale's neighborhood point set, the spatial relative position offset and reflection intensity difference between the neighborhood points and the center point are calculated, where the position offset is the difference in three-dimensional spatial coordinates and the intensity difference is the difference in reflection intensity. The position offset and intensity difference are then concatenated along the feature dimension to generate a local fusion feature matrix for each scale.

[0046] The local structural features of the M-scale local fusion feature matrices are extracted by passing them through one-dimensional convolutional units to obtain M-scale local structural feature matrices.

[0047] The local structural feature matrix is ​​weighted and fused based on the neighborhood sampling weight to generate a multi-scale fused feature matrix for each filtered point cloud data item. The multi-scale fused feature matrices of all filtered point cloud data items are then aggregated to form a multi-scale fused feature tensor.

[0048] Optionally, step five specifically includes:

[0049] The multi-scale fused feature tensor is input into the residual feature extraction structure of the improved PointNeXt network, which includes a backbone feature extraction branch and a reflection feature enhancement branch.

[0050] The main feature extraction branch includes L layers of PointNeXtBlock. The input features of the first layer of PointNeXtBlock are multi-scale fused feature tensors, and the output feature tensor of the current layer of PointNeXtBlock is the input feature tensor of the next layer of PointNeXtBlock. Specifically:

[0051] In the current layer PointNeXtBlock, layer normalization is performed on the input feature tensor to obtain a normalized feature tensor; the normalized feature tensor is passed through two fully connected layers and generated into a nonlinear mapping feature tensor through the GELU activation function and DropPath technique; the input feature tensor and the nonlinear mapping feature tensor are added by residual addition to obtain the output feature tensor of the current layer PointNeXtBlock.

[0052] The output feature tensor of the Lth layer PointNeXtBlock is used as the backbone geometric feature tensor;

[0053] In the reflection feature enhancement branch, the reflection intensity data is extracted from the filtered point cloud dataset to obtain the reflection intensity channel vector;

[0054] The reflection intensity channel vector is modeled for local intensity changes and global feature convergence in the feature dimension through two layers of one-dimensional convolution, and the intermediate feature tensor is output.

[0055] Global average pooling and max pooling are performed on the intermediate feature tensor in the point cloud dimension to obtain average pooling vector and max pooling vector respectively. The average pooling vector and max pooling vector are then concatenated in the channel dimension to obtain the channel description vector. The channel description vector is then input into a fully connected network and a sigmoid activation function to obtain the channel attention weight vector.

[0056] The channel attention weight vector and the intermediate feature tensor are weighted by attention to obtain the reflection feature tensor;

[0057] The core geometric feature tensor and the reflection enhancement feature tensor are concatenated to generate a joint geometric-reflection feature tensor.

[0058] Optionally, step six specifically includes:

[0059] Layer normalization is performed on the joint geometry-reflection feature tensor to obtain the normalized joint feature tensor;

[0060] Extract confidence score data from the filtered point cloud dataset to form a confidence feature vector;

[0061] Softmax normalization is performed on the confidence feature vectors, and the channel dimension is expanded to be consistent with the normalized joint feature tensor to obtain the confidence weighted matrix;

[0062] Perform element-wise multiplication on the confidence-weighted matrix and the normalized joint feature tensor to obtain the confidence-weighted feature tensor;

[0063] The difference between the normalized joint feature tensor and the confidence-weighted feature tensor is used as the residual feature tensor. One-dimensional convolution and GELU activation are then performed on the residual feature tensor to obtain the residual adjustment tensor.

[0064] The confidence-weighted feature tensor and the residual adjustment tensor are fused element-wise by addition and then layer normalization is performed to obtain the fused feature tensor.

[0065] The Euclidean distance is used to calculate the 3D coordinate difference between all pairs of filtered point cloud data items in two adjacent frames of filtered point cloud datasets to obtain the inter-frame residual; the confidence score difference between all pairs of filtered point cloud data items in two adjacent frames of filtered point cloud datasets is calculated.

[0066] Set inter-frame distance threshold and confidence tolerance threshold, and construct geometric consistency mask and confidence consistency mask;

[0067] If the inter-frame residual is less than or equal to the inter-frame distance threshold, the geometric consistency mask is set to 1; otherwise, the geometric consistency mask is set to 0. If the confidence score difference is less than or equal to the confidence tolerance threshold, the confidence consistency mask is set to 1; otherwise, the confidence consistency mask is set to 0.

[0068] Perform a point-by-point logical AND operation between the geometric consistency mask and the confidence consistency mask to obtain the temporal consistency mask;

[0069] Based on the temporal consistency mask, consistency filtering is performed on the fused feature tensor: only the fused features corresponding to the filtered point cloud data items with a temporal consistency mask of 1 are retained to generate the enhanced feature tensor.

[0070] Optionally, step seven specifically includes:

[0071] The enhanced feature tensor is processed through two layers of one-dimensional convolution and ReLU activation function to generate a coordinate residual matrix. Each row of the coordinate residual matrix represents the three-dimensional coordinate residual value of the corresponding point cloud. The coordinate residual matrix is ​​then added element-wise to the three-dimensional coordinate matrix of the filtered point cloud dataset to generate the enhanced coordinate matrix.

[0072] The enhanced feature tensor is passed through a one-dimensional convolution and a sigmoid activation function to generate a reflection enhancement value vector. The reflection enhancement value vector is then weighted and fused with the reflection intensity vector of the filtered point cloud dataset to obtain the enhanced reflection intensity vector.

[0073] The enhanced feature tensor is used to generate a confidence prediction vector through a one-layer one-dimensional convolution and Softmax normalization operation.

[0074] The enhanced coordinate matrix, the enhanced reflection intensity vector, and the confidence prediction vector are combined point by point to generate an enhanced point cloud dataset.

[0075] The beneficial effects of this invention are:

[0076] First, this invention introduces an improved adaptive Kalman filter algorithm to perform point-by-point denoising and reflection intensity enhancement on a structured point cloud dataset. It dynamically updates the system noise covariance matrix using the local density values ​​of the point cloud and introduces an adaptive confidence adjustment mechanism to generate confidence-gated Kalman gain. This enables differentiated filtering strategies in different density regions, effectively improving the smoothness and structural continuity of the filtered point cloud dataset and enhancing the ability to express reflection intensity in boundary regions.

[0077] Secondly, to address the issues of non-uniformity and scale variation in point cloud distribution, this invention constructs an improved PointNeXt network. The improved PointNeXt network introduces a density multi-scale sampling module, which performs density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset based on a multi-scale neighborhood radius and local density inverse weighting strategy. During the feature construction process, the spatial relative position offset and reflection intensity difference are fused to form a multi-scale fused feature tensor.

[0078] Furthermore, this invention introduces reflection intensity information into the residual feature extraction structure of the improved PointNeXt network to construct a reflection feature tensor, which is then concatenated with the backbone geometric feature tensor to form a joint geometric-reflection feature tensor. Channel weighting and residual fusion are performed through a confidence adjustment module, enhancing the feature identification capability in low-confidence regions. A temporal consistency module is introduced to construct a geometric consistency mask and a confidence consistency mask by combining the inter-frame 3D coordinate difference and the confidence score difference, thereby generating a temporal consistency mask. This mask is used to perform inter-frame filtering on the fused feature tensor, improving the cross-frame stability and semantic continuity of the enhanced feature tensor.

[0079] In summary, this invention achieves high-precision noise suppression, structure restoration, and reflection attribute enhancement for radar point clouds through an improved adaptive filtering algorithm and an improved PointNeXt network. Furthermore, it enhances the reliability and stability of point cloud modeling results through confidence-guided and temporal consistency modeling, significantly improving the structural integrity, spatial consistency, and semantic expressiveness of point cloud data in complex scenarios. Attached Figure Description

[0080] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0081] Figure 1 This is a schematic diagram of a radar point cloud denoising and enhancement method based on adaptive filtering proposed in this invention;

[0082] Figure 2 This is a flowchart of the improved adaptive Kalman filter algorithm in the radar point cloud denoising and enhancement method based on adaptive filtering proposed in this invention;

[0083] Figure 3 This is a diagram of the improved PointNeXt network structure in the radar point cloud denoising and enhancement method based on adaptive filtering proposed in this invention. Detailed Implementation

[0084] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0085] refer to Figures 1-3 A radar point cloud denoising and enhancement method based on adaptive filtering includes the following steps:

[0086] Step 1: Collect the raw echo signals from the radar sensor to generate an initial point cloud dataset;

[0087] Step 2: Perform spatial coordinate alignment, reflection intensity normalization, and outlier removal on the initial point cloud dataset to generate a structured point cloud dataset;

[0088] Step 3: Perform point-by-point denoising and reflection intensity enhancement on the structured point cloud dataset using the improved adaptive Kalman filter algorithm to generate a filtered point cloud dataset;

[0089] Step 4: Construct an improved PointNeXt network. The improved PointNeXt network includes a density multi-scale sampling module, a residual feature extraction structure, a confidence adjustment module, a temporal consistency module, and a multi-head reconstruction module. The density multi-scale sampling module performs density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset to generate a multi-scale fused feature tensor.

[0090] Step 5: Extract joint features from the multi-scale fusion feature tensor using a residual feature extraction structure to generate a joint geometry-reflection feature tensor. The residual feature extraction structure incorporates reflection intensity information to construct the reflection feature tensor.

[0091] Step 6: Perform channel weighting and residual fusion on the joint geometry-reflection feature tensor through the confidence adjustment module, and perform inter-frame geometric constraints and confidence consistency constraints through the temporal consistency module to generate the enhanced feature tensor;

[0092] Step 7: Use the multi-head reconstruction module to predict the coordinate residuals, reflection enhancement values, and confidence scores of the enhanced feature tensor to generate an enhanced point cloud dataset.

[0093] In this embodiment, step one specifically includes:

[0094] Set the transmission parameters of the radar sensor and determine the transmission time. The transmission parameters include the transmission waveform type, pulse repetition frequency, bandwidth, and sampling rate.

[0095] The original echo signal received by the radar sensor is converted from analog to digital to obtain an echo discrete sequence. The echo discrete sequence consists of several sampling points, and each sampling point corresponds to a discrete time sampling value.

[0096] Waveform demodulation and down-conversion processing are performed on the discrete echo sequence to obtain a complex baseband echo sequence. The complex amplitude of each sampling point in the complex baseband echo sequence represents the echo amplitude and phase information.

[0097] Multiply the difference between the arrival time and the transmission time of each sampling point by the speed of electromagnetic wave propagation in air, and divide by two to calculate the polar coordinate distance of each sampling point.

[0098] Set a minimum distance threshold and a maximum distance threshold, perform distance threshold filtering on the polar coordinate distance of all sampling points, and retain only sampling points whose polar coordinate distance is greater than or equal to the minimum distance threshold and less than or equal to the maximum distance threshold to obtain a valid echo sampling point sequence;

[0099] Based on the azimuth and elevation angles of the radar array, the polar coordinate distance of each effective echo sampling point is converted into spatial rectangular coordinates to obtain the three-dimensional spatial coordinates of each effective echo sampling point.

[0100] Based on the complex baseband echo sequence, the square of the complex amplitude of each valid echo sampling point is taken as the reflection intensity;

[0101] The initial point cloud data for each valid echo sampling point is composed of three-dimensional spatial coordinates and reflection intensity. All initial point cloud data are aggregated to form an initial point cloud dataset.

[0102] In this embodiment, step two specifically includes:

[0103] A rigid transformation matrix is ​​set to convert the 3D spatial coordinates of all points in the initial point cloud dataset into homogeneous coordinates to obtain aligned spatial coordinates.

[0104] The reflection intensity values ​​of all points are normalized using the min-max normalization method to obtain normalized reflection intensity values.

[0105] Set a spatial neighborhood radius and a threshold for the number of neighborhood points, remove outliers from the initial point cloud dataset, and retain only valid echo sampling points within the spatial neighborhood radius whose number of neighboring points is greater than or equal to the threshold for the number of neighboring points, to obtain a sequence of valid neighborhood points;

[0106] The aligned spatial coordinates and normalized reflection intensity values ​​of all valid neighboring points are combined to form a structured point cloud dataset.

[0107] In this embodiment, step three specifically includes:

[0108] Setting the initial state of the filter specifically includes: initializing the predicted state vector, state covariance matrix, system noise covariance matrix, state transition matrix, and observation matrix;

[0109] Calculate the local density value of each neighborhood valid point in the structured point cloud dataset. The local density value is obtained by counting the number of neighbor points contained in a spatial neighborhood of a set radius and dividing by the volume of the spatial neighborhood.

[0110] Based on local density values, the system noise covariance matrix is ​​dynamically updated using an exponential decay method, so that the larger the local point density, the smaller the adjustment range of the system noise covariance.

[0111] Define a system noise term, which follows a multivariate Gaussian distribution with a mean of zero and a covariance equal to the updated system noise covariance matrix;

[0112] Update the predicted state vector and the predicted covariance matrix. The predicted state vector is the product of the state transition matrix and the predicted state vector at the previous time step, plus the system noise term. The predicted covariance matrix is ​​the product of the state transition matrix and the state covariance matrix at the previous time step, plus the system noise covariance matrix.

[0113] Calculate the product of the predicted state vector and the observation matrix to obtain the predicted observation vector, and use the vector difference between the current observation vector and the predicted observation vector as the observation residual vector.

[0114] The square norm of the observed residual vector is mapped to the confidence score of the current neighborhood valid point through a negative exponential function. The confidence score represents the estimation confidence level of the current neighborhood valid point in the current filtering period.

[0115] The confidence gating function is set using the Sigmoid gating function, mapping the confidence scores of the current neighborhood valid points to confidence gating coefficients:

[0116] ;

[0117] in, This represents the confidence gating coefficient of the current neighborhood valid points. Represents an exponential function. This is the magnification factor. This represents the confidence score of the current neighborhood valid points. The set confidence threshold value;

[0118] Multiply the standard Kalman gain by the confidence gating coefficient to generate the confidence-gated Kalman gain;

[0119] The confidence-gated Kalman gain is multiplied by the observation residual vector and added to the predicted state vector to obtain the filtered state vector of the current neighborhood effective point.

[0120] Based on the filtered state vector, three-dimensional spatial coordinates and reflection intensity are extracted, and the three-dimensional spatial coordinates, reflection intensity and confidence score are combined to form a filtered point cloud data item. All filtered point cloud data items are aggregated to generate a filtered point cloud dataset.

[0121] This invention introduces a system noise dynamic adjustment mechanism based on local density adaptation and a Kalman gain adjustment strategy with confidence gating. This enables the filtering process to flexibly adjust the filtering weights according to the spatial density and estimation confidence of each neighboring point when facing scenarios such as uneven spatial distribution of point clouds and large variations in noise intensity. This effectively improves the local adaptive capability and reflection intensity recovery accuracy of point cloud filtering. At the same time, the confidence gating function, combined with the norm of the observation residual, constructs a filtering confidence interval, further reducing the interference of outliers on state estimation. This improves the overall geometric continuity, edge structure preservation capability, and reflection intensity consistency of the filtered point cloud dataset, achieving more robust point cloud denoising and enhancement in complex environments.

[0122] In this embodiment, the step of performing density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset through the density multi-scale sampling module to generate a multi-scale fused feature tensor specifically includes:

[0123] Define a multi-scale neighborhood set, which includes neighborhood radii of M scales;

[0124] Calculate the local density value of each filtered point cloud data item in the multi-scale neighborhood set;

[0125] For each filtered point cloud data item, the reciprocal of the local density value at each scale is taken to obtain the reciprocal density value at each scale, and the reciprocal density values ​​at all scales are normalized to obtain the neighborhood sampling weight at each scale.

[0126] Each filtered point cloud data item is used as the center point, and weighted K-nearest neighbor sampling is performed based on the neighborhood radius at each scale to obtain the neighborhood point set of M scales for each filtered point cloud data item.

[0127] For each scale's neighborhood point set, the spatial relative position offset and reflection intensity difference between the neighborhood points and the center point are calculated, where the position offset is the difference in three-dimensional spatial coordinates and the intensity difference is the difference in reflection intensity. The position offset and intensity difference are then concatenated along the feature dimension to generate a local fusion feature matrix for each scale.

[0128] The local structural features of the M-scale local fusion feature matrices are extracted by passing them through one-dimensional convolutional units to obtain M-scale local structural feature matrices.

[0129] The local structural feature matrix is ​​weighted and fused based on the neighborhood sampling weight to generate a multi-scale fused feature matrix for each filtered point cloud data item. The multi-scale fused feature matrices of all filtered point cloud data items are then aggregated to form a multi-scale fused feature tensor.

[0130] In this embodiment, step five specifically includes:

[0131] The multi-scale fused feature tensor is input into the residual feature extraction structure of the improved PointNeXt network, which includes a backbone feature extraction branch and a reflection feature enhancement branch.

[0132] The main feature extraction branch includes L layers of PointNeXtBlock. The input features of the first layer of PointNeXtBlock are multi-scale fused feature tensors, and the output feature tensor of the current layer of PointNeXtBlock is the input feature tensor of the next layer of PointNeXtBlock. Specifically:

[0133] In the current layer PointNeXtBlock, layer normalization is performed on the input feature tensor to obtain a normalized feature tensor; the normalized feature tensor is passed through two fully connected layers and generated into a nonlinear mapping feature tensor through the GELU activation function and DropPath technique; the input feature tensor and the nonlinear mapping feature tensor are added by residual addition to obtain the output feature tensor of the current layer PointNeXtBlock.

[0134] In this invention, the DropPath technique randomly discards some residual branch paths during the training phase, causing slight changes in the network structure in different training iterations. This avoids all paths relying on the same feature propagation channel simultaneously, reducing the risk of overfitting. At the same time, the DropPath technique enables the network to form diverse feature representations and path selections during feature propagation, improving the stability and convergence of deep networks in complex point cloud scenarios and ensuring the preservation of structural details in multi-scale and high-noise environments.

[0135] The output feature tensor of the Lth layer PointNeXtBlock is used as the backbone geometric feature tensor;

[0136] In the reflection feature enhancement branch, the reflection intensity data is extracted from the filtered point cloud dataset to obtain the reflection intensity channel vector;

[0137] The reflection intensity channel vector is modeled for local intensity changes and global feature aggregation in the feature dimension through two layers of one-dimensional convolution, and an intermediate feature tensor is output. The reflection intensity channel vector is a single-channel scalar in the feature dimension, and after passing through two layers of one-dimensional convolution, the feature dimension is mapped to a high-dimensional channel.

[0138] Global average pooling and max pooling are performed on the intermediate feature tensor in the point cloud dimension to obtain average pooling vector and max pooling vector respectively. The average pooling vector and max pooling vector are then concatenated in the channel dimension to obtain the channel description vector. The channel description vector is then input into a fully connected network and a sigmoid activation function to obtain the channel attention weight vector.

[0139] The channel attention weight vector and the intermediate feature tensor are weighted by attention to obtain the reflection feature tensor;

[0140] The core geometric feature tensor and the reflection enhancement feature tensor are concatenated to generate a joint geometric-reflection feature tensor.

[0141] In this embodiment, step six specifically includes:

[0142] Layer normalization is performed on the joint geometry-reflection feature tensor to obtain the normalized joint feature tensor;

[0143] Extract confidence score data from the filtered point cloud dataset to form a confidence feature vector;

[0144] Softmax normalization is performed on the confidence feature vectors, and the channel dimension is expanded to be consistent with the normalized joint feature tensor to obtain the confidence weighted matrix;

[0145] Perform element-wise multiplication on the confidence-weighted matrix and the normalized joint feature tensor to obtain the confidence-weighted feature tensor;

[0146] The difference between the normalized joint feature tensor and the confidence-weighted feature tensor is used as the residual feature tensor. One-dimensional convolution and GELU activation are then performed on the residual feature tensor to obtain the residual adjustment tensor.

[0147] The confidence-weighted feature tensor and the residual adjustment tensor are fused element-wise by addition and then layer normalization is performed to obtain the fused feature tensor.

[0148] The Euclidean distance is used to calculate the 3D coordinate difference between all pairs of filtered point cloud data items in two adjacent frames of filtered point cloud datasets to obtain the inter-frame residual; the confidence score difference between all pairs of filtered point cloud data items in two adjacent frames of filtered point cloud datasets is calculated.

[0149] Set inter-frame distance threshold and confidence tolerance threshold, and construct geometric consistency mask and confidence consistency mask;

[0150] If the inter-frame residual is less than or equal to the inter-frame distance threshold, the geometric consistency mask is set to 1; otherwise, the geometric consistency mask is set to 0. If the confidence score difference is less than or equal to the confidence tolerance threshold, the confidence consistency mask is set to 1; otherwise, the confidence consistency mask is set to 0.

[0151] Perform a point-by-point logical AND operation between the geometric consistency mask and the confidence consistency mask to obtain the temporal consistency mask;

[0152] Based on the temporal consistency mask, consistency filtering is performed on the fused feature tensor: only the fused features corresponding to the filtered point cloud data items with a temporal consistency mask of 1 are retained to generate the enhanced feature tensor.

[0153] In this invention, a temporal consistency mask based on geometric residuals and confidence fluctuations is constructed to perform a consistency screening operation on the fused feature tensor, effectively improving the stability and physical consistency of the enhanced feature tensor across consecutive frames. This operation can suppress inconsistent features caused by sensor jitter, motion occlusion, or sporadic noise in dynamic environments, thereby improving the continuity and reliability of point cloud augmentation results, and is particularly suitable for high-precision mapping and target tracking tasks. For data items with a temporal consistency mask value of 0, they are not directly deleted, but their corresponding fused features are set to all-zero tensors to preserve their spatial position and avoid point cloud structure breaks. This also prevents erroneous features from affecting subsequent augmentation and reconstruction modules, effectively ensuring the robustness and structural continuity of the overall model.

[0154] In this embodiment, step seven specifically includes:

[0155] The enhanced feature tensor is processed through two layers of one-dimensional convolution and ReLU activation function to generate a coordinate residual matrix. Each row of the coordinate residual matrix represents the three-dimensional coordinate residual value of the corresponding point cloud. The coordinate residual matrix is ​​then added element-wise to the three-dimensional coordinate matrix of the filtered point cloud dataset to generate the enhanced coordinate matrix.

[0156] The enhanced feature tensor is passed through a one-dimensional convolution and a sigmoid activation function to generate a reflection enhancement value vector. The reflection enhancement value vector is then weighted and fused with the reflection intensity vector of the filtered point cloud dataset to obtain the enhanced reflection intensity vector.

[0157] The enhanced feature tensor is used to generate a confidence prediction vector through a one-layer one-dimensional convolution and Softmax normalization operation.

[0158] The enhanced coordinate matrix, the enhanced reflection intensity vector, and the confidence prediction vector are combined point by point to generate an enhanced point cloud dataset.

[0159] Example 1

[0160] To verify the feasibility of this invention in practice, it was applied to a 3D point cloud inspection system in a large-scale intelligent manufacturing workshop for automotive parts. This workshop uses millimeter-wave radar as one of the inspection sensors for the production line equipment, used to monitor the shape integrity and assembly status of metal components such as brake calipers and steering knuckles on the conveyor belt in real time. Due to high-temperature radiation, oil mist reflection, and mechanical vibration in the production environment, the original echo signals acquired by the radar exhibit significant multipath interference and non-uniform noise. This results in traditional point cloud filtering algorithms performing poorly in edge feature recovery and reflection intensity consistency, especially in areas with smooth surfaces where the reflected signal saturates and the noise point density is much higher than that of structural points. This leads to a high misjudgment rate in the detection model, making accurate geometric reconstruction and defect identification impossible.

[0161] In practical applications, this invention uses an industrial radar sensor to acquire raw echo signals at a sampling frequency of 40MHz and a pulse repetition frequency of 20kHz, resulting in approximately 85,000 point cloud points per frame. First, waveform demodulation and distance threshold filtering are performed to remove invalid sampling points with distances less than 50cm or greater than 3m. After spatial coordinate alignment and reflection intensity normalization, a structured point cloud dataset is generated. For this structured point cloud dataset, an improved adaptive Kalman filter algorithm is used to denoise and enhance the reflection intensity of each point cloud data item point by point. During the filtering process, the system noise covariance is dynamically adjusted based on the local density of the point cloud. When the local density is greater than 120 points / m³, the covariance attenuation coefficient is set to 0.3; when the density is less than 40 points / m³, the covariance attenuation coefficient is increased to 0.8 to maintain noise suppression capability in sparse regions.

[0162] The filtered point cloud dataset is input into an improved PointNeXt network for point cloud enhancement and reconstruction. The density multi-scale sampling module sets three scale neighborhood radii: 5mm, 10mm, and 20mm, and performs weighted K-nearest neighbor sampling on regions of different densities. After structural modeling through residual feature extraction, the output joint geometry-reflection feature tensor shows significant improvements in both spatial boundaries and reflection continuity. The confidence adjustment module adaptively enhances uncertain regions through Softmax normalization and channel weighting, raising the confidence of feature points with a confidence level below 0.4 to above 0.7 after enhancement. Simultaneously, the temporal consistency module constrains inter-frame geometric deviations across 10 consecutive frame point cloud sequences, reducing the average spatial drift from 2.4mm to 0.6mm and the confidence standard deviation from 0.15 to 0.05, significantly improving cross-frame stability.

[0163] To further verify the effectiveness of the present invention, the method of the present invention was compared and analyzed with traditional mean filtering, statistical feature-based filtering algorithms, and PointNet++ network enhancement methods. The experimental results are shown in Table 1.

[0164] Table 1. Comparison of point cloud denoising and enhancement performance of different methods

[0165] Indicator Item Traditional mean filtering method Filtering algorithms based on statistical features PointNet++ Network Enhancement Methods Method of the present invention Signal-to-noise ratio improvement (dB) 8.7 11.2 14.8 17.6 Edge point retention rate (%) 82.3 86.5 90.4 96.5 Mean square error of reflection intensity (MSE) 0.094 0.071 0.056 0.032 Geometric reconstruction accuracy (mm) 2.7 2.3 1.6 1.0 Confidence stability (σ) 0.21 0.17 0.13 0.05 Inter-frame consistency index 0.73 0.79 0.86 0.94 Processing time (ms / frame) 41.6 58.2 72.5 79.8

[0166] As shown in Table 1, the present invention significantly outperforms the comparative methods in multiple metrics. The signal-to-noise ratio of the present invention is improved by 17.6 dB, approximately 57% higher than the statistical feature-based filtering algorithm; the edge point preservation rate of the present invention is 96.5%, effectively avoiding the structural breakage problem caused by traditional filtering; the mean square error of reflection intensity is reduced to 0.032, indicating that the enhanced reflection intensity distribution is smoother and more continuous, possessing better surface restoration capability and local detail fidelity; in terms of geometric reconstruction accuracy, the average reconstruction error of the present invention is only 1.0 mm, a 37.5% reduction compared to the PointNet++ network enhancement method; the standard deviation of confidence stability is only 0.05, indicating that the confidence distribution of the enhanced point cloud is more stable across multiple frames; the inter-frame consistency index reaches 0.94, proving that the temporal consistency module effectively suppresses point cloud drift in dynamic scenes. Although the present invention introduces a more complex density adaptive sampling and confidence adjustment mechanism, resulting in a slightly higher single-frame processing time of 79.8 ms / frame, it significantly improves point cloud quality and structural accuracy while ensuring real-time performance.

[0167] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A radar point cloud denoising and enhancement method based on adaptive filtering, characterized in that, Includes the following steps: Step 1: Collect the raw echo signals from the radar sensor to generate an initial point cloud dataset; Step 2: Perform spatial coordinate alignment, reflection intensity normalization, and outlier removal on the initial point cloud dataset to generate a structured point cloud dataset; Step 3: Perform point-by-point denoising and reflection intensity enhancement on the structured point cloud dataset using the improved adaptive Kalman filter algorithm to generate a filtered point cloud dataset; Step 4: Construct an improved PointNeXt network. The improved PointNeXt network includes a density multi-scale sampling module, a residual feature extraction structure, a confidence adjustment module, a temporal consistency module, and a multi-head reconstruction module. The density multi-scale sampling module performs density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset to generate a multi-scale fused feature tensor. Step 5: Extract joint features from the multi-scale fusion feature tensor using a residual feature extraction structure to generate a joint geometry-reflection feature tensor. The residual feature extraction structure incorporates reflection intensity information to construct the reflection feature tensor. Step 6: Perform channel weighting and residual fusion on the joint geometry-reflection feature tensor through the confidence adjustment module, and perform inter-frame geometric constraints and confidence consistency constraints through the temporal consistency module to generate the enhanced feature tensor; Step 7: Use the multi-head reconstruction module to predict the coordinate residuals, reflection enhancement values, and confidence scores of the enhanced feature tensor to generate an enhanced point cloud dataset.

2. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, Step one specifically includes: Set the transmission parameters of the radar sensor and determine the transmission time. The transmission parameters include the transmission waveform type, pulse repetition frequency, bandwidth, and sampling rate. The original echo signal received by the radar sensor is converted from analog to digital to obtain an echo discrete sequence. The echo discrete sequence consists of several sampling points, and each sampling point corresponds to a discrete time sampling value. Waveform demodulation and down-conversion processing are performed on the discrete echo sequence to obtain a complex baseband echo sequence. The complex amplitude of each sampling point in the complex baseband echo sequence represents the echo amplitude and phase information. Multiply the difference between the arrival time and the transmission time of each sampling point by the speed of electromagnetic wave propagation in air, and divide by two to calculate the polar coordinate distance of each sampling point. Set a minimum distance threshold and a maximum distance threshold, perform distance threshold filtering on the polar coordinate distance of all sampling points, and retain only sampling points whose polar coordinate distance is greater than or equal to the minimum distance threshold and less than or equal to the maximum distance threshold to obtain a valid echo sampling point sequence; Based on the azimuth and elevation angles of the radar array, the polar coordinate distance of each effective echo sampling point is converted into spatial rectangular coordinates to obtain the three-dimensional spatial coordinates of each effective echo sampling point. Based on the complex baseband echo sequence, the square of the complex amplitude of each valid echo sampling point is taken as the reflection intensity; The initial point cloud data for each valid echo sampling point is composed of three-dimensional spatial coordinates and reflection intensity. All initial point cloud data are aggregated to form an initial point cloud dataset.

3. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, Step two specifically includes: A rigid transformation matrix is ​​set to convert the 3D spatial coordinates of all points in the initial point cloud dataset into homogeneous coordinates to obtain aligned spatial coordinates. The reflection intensity values ​​of all points are normalized using the min-max normalization method to obtain normalized reflection intensity values. Set a spatial neighborhood radius and a threshold for the number of neighborhood points, remove outliers from the initial point cloud dataset, and retain only valid echo sampling points within the spatial neighborhood radius whose number of neighboring points is greater than or equal to the threshold for the number of neighboring points, to obtain a sequence of valid neighborhood points; The aligned spatial coordinates and normalized reflection intensity values ​​of all valid neighboring points are combined to form a structured point cloud dataset.

4. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, Step three specifically includes: Setting the initial state of the filter specifically includes: initializing the predicted state vector, state covariance matrix, system noise covariance matrix, state transition matrix, and observation matrix; Calculate the local density value of each neighborhood valid point in the structured point cloud dataset. The local density value is obtained by counting the number of neighbor points contained in a spatial neighborhood of a set radius and dividing by the volume of the spatial neighborhood. Based on local density values, the system noise covariance matrix is ​​dynamically updated using an exponential decay method. Define a system noise term, which follows a multivariate Gaussian distribution with a mean of zero and a covariance equal to the updated system noise covariance matrix; Update the predicted state vector and the predicted covariance matrix. The predicted state vector is the product of the state transition matrix and the predicted state vector at the previous time step, plus the system noise term. The predicted covariance matrix is ​​the product of the state transition matrix and the state covariance matrix at the previous time step, plus the system noise covariance matrix. Calculate the product of the predicted state vector and the observation matrix to obtain the predicted observation vector, and use the vector difference between the current observation vector and the predicted observation vector as the observation residual vector. The square norm of the observed residual vector is mapped to the confidence score of the current neighborhood valid points through a negative exponential function; The confidence gating function is set using the Sigmoid gating function, which maps the confidence scores of the current neighborhood valid points to the confidence gating coefficients. Multiply the standard Kalman gain by the confidence gating coefficient to generate the confidence-gated Kalman gain; The confidence-gated Kalman gain is multiplied by the observation residual vector and added to the predicted state vector to obtain the filtered state vector of the current neighborhood effective point. Based on the filtered state vector, three-dimensional spatial coordinates and reflection intensity are extracted, and the three-dimensional spatial coordinates, reflection intensity and confidence score are combined to form a filtered point cloud data item. All filtered point cloud data items are aggregated to generate a filtered point cloud dataset.

5. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, The process of performing density adaptive sampling and multi-scale neighborhood feature grouping on the filtered point cloud dataset through a density multi-scale sampling module to generate a multi-scale fused feature tensor specifically includes: Define a multi-scale neighborhood set, which includes neighborhood radii of M scales; Calculate the local density value of each filtered point cloud data item in the multi-scale neighborhood set; For each filtered point cloud data item, the reciprocal of the local density value at each scale is taken to obtain the reciprocal density value at each scale, and the reciprocal density values ​​at all scales are normalized to obtain the neighborhood sampling weight at each scale. Each filtered point cloud data item is used as the center point, and weighted K-nearest neighbor sampling is performed based on the neighborhood radius at each scale to obtain the neighborhood point set of M scales for each filtered point cloud data item. For each scale's neighborhood point set, the spatial relative position offset and reflection intensity difference between the neighborhood points and the center point are calculated, where the position offset is the difference in three-dimensional spatial coordinates and the intensity difference is the difference in reflection intensity. The position offset and intensity difference are then concatenated along the feature dimension to generate a local fusion feature matrix for each scale. The local structural features of the M-scale local fusion feature matrices are extracted by passing them through one-dimensional convolutional units to obtain M-scale local structural feature matrices. The local structural feature matrix is ​​weighted and fused based on the neighborhood sampling weight to generate a multi-scale fused feature matrix for each filtered point cloud data item. The multi-scale fused feature matrices of all filtered point cloud data items are then aggregated to form a multi-scale fused feature tensor.

6. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, Step five specifically includes: The multi-scale fused feature tensor is input into the residual feature extraction structure of the improved PointNeXt network, which includes a backbone feature extraction branch and a reflection feature enhancement branch. The main feature extraction branch includes L layers of PointNeXtBlock. The input features of the first layer of PointNeXtBlock are multi-scale fused feature tensors, and the output feature tensor of the current layer of PointNeXtBlock is the input feature tensor of the next layer of PointNeXtBlock. Specifically: In the current layer PointNeXtBlock, layer normalization is performed on the input feature tensor to obtain a normalized feature tensor; the normalized feature tensor is passed through two fully connected layers and generated into a nonlinear mapping feature tensor through the GELU activation function and DropPath technique; the input feature tensor and the nonlinear mapping feature tensor are added by residual addition to obtain the output feature tensor of the current layer PointNeXtBlock. The output feature tensor of the Lth layer PointNeXtBlock is used as the backbone geometric feature tensor; In the reflection feature enhancement branch, the reflection intensity data is extracted from the filtered point cloud dataset to obtain the reflection intensity channel vector; The reflection intensity channel vector is modeled for local intensity changes and global feature convergence in the feature dimension through two layers of one-dimensional convolution, and the intermediate feature tensor is output. Global average pooling and max pooling are performed on the intermediate feature tensor in the point cloud dimension to obtain average pooling vector and max pooling vector respectively. The average pooling vector and max pooling vector are then concatenated in the channel dimension to obtain the channel description vector. The channel description vector is then input into a fully connected network and a sigmoid activation function to obtain the channel attention weight vector. The channel attention weight vector and the intermediate feature tensor are weighted by attention to obtain the reflection feature tensor; The core geometric feature tensor and the reflection enhancement feature tensor are concatenated to generate a joint geometric-reflection feature tensor.

7. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, Step six specifically includes: Layer normalization is performed on the joint geometry-reflection feature tensor to obtain the normalized joint feature tensor; Extract confidence score data from the filtered point cloud dataset to form a confidence feature vector; Softmax normalization is performed on the confidence feature vectors, and the channel dimension is expanded to be consistent with the normalized joint feature tensor to obtain the confidence weighted matrix; Perform element-wise multiplication on the confidence-weighted matrix and the normalized joint feature tensor to obtain the confidence-weighted feature tensor; The difference between the normalized joint feature tensor and the confidence-weighted feature tensor is used as the residual feature tensor. One-dimensional convolution and GELU activation are then performed on the residual feature tensor to obtain the residual adjustment tensor. The confidence-weighted feature tensor and the residual adjustment tensor are fused element-wise by addition and then layer normalization is performed to obtain the fused feature tensor. The Euclidean distance is used to calculate the 3D coordinate difference between all pairs of filtered point cloud data items in two adjacent frames of filtered point cloud datasets to obtain the inter-frame residual; the confidence score difference between all pairs of filtered point cloud data items in two adjacent frames of filtered point cloud datasets is calculated. Set inter-frame distance threshold and confidence tolerance threshold, and construct geometric consistency mask and confidence consistency mask; If the inter-frame residual is less than or equal to the inter-frame distance threshold, the geometric consistency mask is set to 1; otherwise, the geometric consistency mask is set to 0. If the confidence score difference is less than or equal to the confidence tolerance threshold, the confidence consistency mask is set to 1; otherwise, the confidence consistency mask is set to 0. Perform a point-by-point logical AND operation between the geometric consistency mask and the confidence consistency mask to obtain the temporal consistency mask; Based on the temporal consistency mask, consistency filtering is performed on the fused feature tensor: only the fused features corresponding to the filtered point cloud data items with a temporal consistency mask of 1 are retained to generate the enhanced feature tensor.

8. The radar point cloud denoising and enhancement method based on adaptive filtering according to claim 1, characterized in that, Step seven specifically includes: The enhanced feature tensor is processed through two layers of one-dimensional convolution and ReLU activation function to generate a coordinate residual matrix. Each row of the coordinate residual matrix represents the three-dimensional coordinate residual value of the corresponding point cloud. The coordinate residual matrix is ​​then added element-wise to the three-dimensional coordinate matrix of the filtered point cloud dataset to generate the enhanced coordinate matrix. The enhanced feature tensor is passed through a one-dimensional convolution and a sigmoid activation function to generate a reflection enhancement value vector. The reflection enhancement value vector is then weighted and fused with the reflection intensity vector of the filtered point cloud dataset to obtain the enhanced reflection intensity vector. The enhanced feature tensor is used to generate a confidence prediction vector through a one-layer one-dimensional convolution and Softmax normalization operation. The enhanced coordinate matrix, the enhanced reflection intensity vector, and the confidence prediction vector are combined point by point to generate an enhanced point cloud dataset.

Images