Laser radar point cloud enhancement processing method, device and equipment and storage medium

Abstract

The invention discloses a laser radar point cloud enhancement processing method, device and equipment and a storage medium, and the method is characterized in that the method comprises the steps: carrying out the space-time registration, elevation filtering and clustering density filtering processing of an initial point cloud collected by a laser radar, and obtaining an initial point cloud point; removing the outliers to obtain to-be-enhanced point cloud data in a target format; setting a classification label, and labeling the to-be-enhanced point cloud data based on the classification label to obtain a labeled data set; the annotation data set comprises a plurality of data points; inputting the annotation data set into an initial neural network model for training to obtain a multi-task neural network model; and performing point cloud enhancement processing on to-be-enhanced point cloud initial data by using the multi-task neural network model, and outputting enhanced point cloud data. According to the invention, the problems of low point cloud recognition precision, low effective point cloud retention rate, high target detection false detection rate and the like in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of point cloud enhancement, and more particularly to a laser radar point cloud enhancement processing method and device, equipment and a storage medium. BACKGROUND

[0002] With the rapid development of automatic driving technology, the data processing capability of vehicle-mounted laser radar as a core perception sensor directly affects the accuracy and real-time performance of environmental perception. High-beam laser radar has become a key hardware for improving perception performance due to its higher point cloud density and angular resolution. However, how to efficiently process high-density point cloud data and maintain robustness in complex environments (such as rain, fog, and night) has become a core challenge restricting the performance release of laser radar.

[0003] The existing vehicle-mounted laser radar system generally adopts a single-mode data processing architecture, that is, the laser radar is responsible for echo signal acquisition, and point cloud processing relies on an independent computing unit to realize basic point cloud cleaning through a traditional filtering algorithm and complete preliminary data optimization based on a fixed parameter signal processing logic. At present, multi-sensor is also introduced in some schemes to improve the adaptability of the environment by fusing multiple sensors.

[0004] However, the traditional filtering algorithm relies on preset thresholds, and in dynamic interference scenes such as rain, fog, and dust, the fixed thresholds cannot be adaptively adjusted, resulting in a significant decrease in effective point cloud retention rate; the existing scheme lacks targeted processing capability for complex interference scenes such as high-reflection ghosting and multipath reflection, and traditional time-domain filters cannot separate real targets from false reflections, resulting in an increase in target detection false detection rate; hardware-side predefined algorithms are difficult to adapt to new deep learning models, and software-side AI models are limited by the computing power of embedded computing units and cannot deploy large-parameter networks, restricting further improvement of perception accuracy. SUMMARY

[0005] In view of at least one defect or improvement demand of the prior art, the present application provides a laser radar point cloud enhancement processing method, device, equipment and storage medium, which is used to solve the problem that the point cloud processing method in the prior art is easily affected by dynamic environmental interference, cannot adaptively adjust thresholds, results in low point cloud recognition accuracy, low effective point cloud retention rate, cannot separate real targets from false reflections, and results in an increase in target detection false detection rate.

[0006] To achieve the above-mentioned purpose, according to a first aspect of the present application, a laser radar point cloud enhancement processing method is provided, comprising: performing space-time registration, elevation filtering and cluster density filtering processing on the initial point cloud collected by the laser radar, and removing outliers to obtain target format point cloud data to be enhanced; Set a classification label, label the to-be-enhanced point cloud data based on the classification label, and obtain a labeled data set; the labeled data set contains a plurality of data points; Input the labeled data set into an initial neural network model for training, and obtain a multi-task neural network model; Perform point cloud enhancement processing on the to-be-enhanced point cloud initial data by using the multi-task neural network model, and output enhanced point cloud data.

[0007] In a possible implementation, the multi-task neural network model is used to perform point cloud enhancement processing on the to-be-enhanced point cloud initial data, and output enhanced point cloud data, and the method further includes: Extract the geometric feature, intensity feature and local neighborhood texture feature of each point in the labeled data set, output the noise point confidence of each point through the first detection head, and mark and remove the noise points; Extract the reflection intensity variation feature and spatial discreteness feature of each point in the labeled data set, output the rain and fog probability map through the second detection head, and mark and remove all rain and fog points; Extract the delay time feature and intensity attenuation feature of each point in the labeled data set, output the ghosting / ghosting probability map through the third detection head, and mark and remove the ghosting / ghosting points; Extract the angle feature, distance difference feature and intensity feature of each point in the labeled data set, output the multipath reflection correction label through the fourth detection head, and retain the target reflection points and correct other reflection points; Based on the fifth detection head, extract the geometric feature of the boundary region in the labeled data set, and generate the completed point cloud of the sparse region.

[0008] In a possible implementation, the geometric feature, intensity feature and local neighborhood texture feature of each point in the labeled data set are extracted, the noise point confidence of each point is output through the first detection head, and the noise points are marked and removed, and the method further includes: Extract the three-dimensional coordinates and laser reflection intensity of each point from the labeled data set, and calculate the geometric feature of each point based on the three-dimensional coordinates; Determine the neighboring points within a preset range of each point, and extract the local texture feature according to the covariance matrix of the neighboring points; The first detection head calculates the noise point confidence of each point according to the geometric feature, laser reflection intensity and local neighborhood texture feature of each point; According to the preset confidence threshold and the noise point confidence of each point, all noise points are marked and removed.

[0009] In a possible implementation, the reflection intensity variation feature and spatial discreteness feature of each point in the labeled data set are extracted, the rain and fog probability map is output through the second detection head, and all rain and fog points are marked and removed, and the method further includes: Extract the Euclidean distance between each point and the neighboring points from the labeled data set; calculate the spatial discreteness feature according to the Euclidean distance; determine the reflection intensity variation feature according to the laser reflection intensity; The second detection head infers and outputs the rain and fog probability map according to the reflection intensity variation feature and the spatial discreteness feature of each point; Based on the rain and fog probability map and the preset rain and fog probability threshold, all rain and fog points in the labeled data set are marked and removed.

[0010] In a possible implementation, the delay time feature and the intensity attenuation feature of each point in the labeled data set are extracted, and the ghosting / ghosting probability map is output by the third detection head, the ghosting / ghosting points are marked and removed, and the method further includes: extract the echo delay time of each point from the labeled data set; determine the delay time feature of each point based on the echo delay time and the expected delay time; extract the intensity attenuation feature according to the attenuation of the laser reflection intensity and the distance; The third detection head infers and outputs the ghosting / ghosting probability map according to the delay time feature and the intensity attenuation feature of each point; Based on the ghosting / ghosting probability map and the preset ghosting / ghosting probability threshold, all ghosting / ghosting points in the labeled data set are marked and removed.

[0011] In a possible implementation, the angle feature, distance difference feature and intensity feature of each point in the labeled data set are extracted, and the multi-path reflection correction label is output by the fourth detection head, the target reflection point is retained and other reflection points are corrected, and the method further includes: extract the perception center angle and distance information of each point from the labeled data set; determine the angle relationship between the emission direction of the laser beam and the target point according to the perception center angle to obtain the angle feature; determine the distance difference feature according to the distance difference of the distance information and the adjacent points; The fourth detection head generates a multi-path reflection correction label according to the angle feature, the distance difference feature and the intensity feature of each point; Based on the multi-path reflection correction label, the target reflection point is retained and other reflection points are corrected.

[0012] In a possible implementation, the geometric feature of the boundary region in the labeled data set is extracted based on the fifth detection head, and the completed point cloud of the sparse region is generated, and the method further includes: identify the boundary points according to the spatial distribution density and the curvature variation of the labeled data set to obtain the boundary region; extract the geometric feature of the boundary region from the labeled data set based on the fifth detection head; identify the sparse region in the boundary region according to the geometric feature of the boundary region; The regression task is established to predict the coordinates of the completion points in the sparse region, and a completion point cloud is generated to complete the sparse region.

[0013] According to a second aspect of the present application, a laser radar point cloud enhancement processing device is also provided, comprising: The acquisition processing module is configured to perform space-time registration, elevation filtering and cluster density filtering processing on the initial point cloud acquired by the laser radar, and remove outliers to obtain target format point cloud data to be enhanced; The classification labeling module is configured to set a classification label, label the point cloud data to be enhanced based on the classification label, and obtain a labeled data set; the labeled data set contains a plurality of data points; The model training module is configured to input the labeled data set into an initial neural network model for training, and obtain a multi-task neural network model; The point cloud enhancement module is configured to perform point cloud enhancement processing on the initial point cloud data to be enhanced by using the multi-task neural network model, and output enhanced point cloud data.

[0014] According to a third aspect of the present application, a laser radar point cloud enhancement processing device is also provided, comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit executes the steps of any one of the above laser radar point cloud enhancement processing methods.

[0015] According to a fourth aspect of the present application, a storage medium is also provided, which stores a computer program executable by a laser radar point cloud enhancement processing device, and when the computer program runs on the laser radar point cloud enhancement processing device, the laser radar point cloud enhancement processing device executes the steps of any one of the above laser radar point cloud enhancement processing methods.

[0016] Overall, compared with the prior art, the above technical solutions conceived by the present application can achieve the following beneficial effects: The application provides a laser radar point cloud enhancement processing method, which effectively removes outliers and reduces noise interference by performing space-time registration, height filtering and cluster density filtering processing on the initial point cloud, thereby providing a cleaner and more accurate data basis for subsequent processing and helping to improve the overall accuracy of point cloud recognition. The multi-task neural network model can adaptively learn point cloud features in different environments, including noise points, rain and fog points, ghosting / virtual image, multipath reflection, etc., and can maintain a high effective point cloud retention rate under different environmental conditions (such as rain and fog weather), solving the problem of low effective point cloud retention rate in a traditional method that relies on a preset threshold in a harsh environment. The point cloud enhancement processing can accurately identify and remove false point clouds caused by reflection, refraction, etc., while retaining the point cloud data of real targets, thereby reducing the false detection rate of target detection. By setting a classification label and labeling the to-be-enhanced point cloud data, a labeled data set is constructed, providing a rich variety of data resources for model training. This modular design makes it easy to extend new point cloud enhancement functions or adapt to new application scenarios, and only the labeled data set and model training strategy need to be adjusted. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0018] Figure 1 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 2 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 1 The flowchart of an embodiment of step S103 in the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 3 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 2 The flowchart of an embodiment of step S201 in the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 4 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 2 The flowchart of an embodiment of step S202 in the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 5 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 2 The flowchart of an embodiment of step S203 in the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 6 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 2 The flowchart of an embodiment of step S204 in the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 7 The flowchart of an embodiment of the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 2 The flowchart of an embodiment of step S205 in the laser radar point cloud enhancement processing method provided by the present application is shown in the figure. Figure 8 Structure diagram of an embodiment of a laser radar point cloud enhancement processing device provided by the present application; Figure 9 Structure diagram of a laser radar point cloud enhancement processing device provided by an embodiment of the present application; Figure 10 Framework diagram of an embodiment of a laser radar point cloud enhancement processing method provided by the present application. DETAILED DESCRIPTION

[0019] In order to make the objects, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0020] The terms "first", "second", "third" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish different objects and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed or can optionally include other steps or units inherent to the process, method, product or device.

[0021] The present application provides a laser radar point cloud enhancement processing method, device, equipment and storage medium, which are described below respectively.

[0022] Please refer to Figure 1 , Figure 1 Flow diagram of an embodiment of a laser radar point cloud enhancement processing method provided by the present application, in one specific embodiment of the present application, a laser radar point cloud enhancement processing method is disclosed, which comprises: S101, performing space-time registration, elevation filtering and cluster density filtering processing on the initial point cloud collected by the laser radar, removing outliers to obtain target format point cloud data to be enhanced; S102, setting a classification label, labeling the point cloud data to be enhanced based on the classification label to obtain a labeled data set; the labeled data set contains a plurality of data points; S103, inputting the labeled data set into an initial neural network model for training to obtain a multi-task neural network model; S104, using the multi-task neural network model to perform point cloud enhancement processing on the initial point cloud data to be enhanced, and outputting enhanced point cloud data.

[0023] In the above embodiments, first, the initial point cloud data collected by the laser radar is subjected to spatio-temporal registration processing, ensuring that the data collected at different time points or by different sensors can be accurately aligned in the same coordinate system, eliminating errors caused by motion or time differences. Then, the elevation filtering technique is applied, and according to the height characteristics of the terrain or objects, the point cloud data that deviates significantly from the expected height range, such as ground clutter or low-altitude birds, is removed, thereby reducing unnecessary computational burden and interference. Through cluster density analysis, outliers, i.e., abnormal points that exist in space in isolation and have a density significantly lower than the surrounding point cloud, are identified and removed, thereby obtaining target format point cloud data to be enhanced.

[0024] According to the characteristics of the point cloud data and application requirements, detailed classification labels are set, such as noise points, rain and fog points, ghost / virtual points, multi-path reflection points, and edge points. Based on the set classification labels, the point cloud data to be enhanced is labeled point by point to form a labeled data set. During the labeling process, an automatic label generation combined with manual verification method can be used to improve the labeling efficiency and accuracy. The labeled data set contains a plurality of data points, each of which is attached with corresponding class labels and possible confidence information.

[0025] A multi-task neural network model (such as a CNN or Transformer model) is used to perform point cloud enhancement processing on the labeled data set. The model uses a shared backbone network and multi-task detection head design, which can simultaneously process multiple point cloud enhancement tasks in a single model, such as noise point identification, rain and fog point removal, ghost / virtual point removal, multi-path reflection identification, and edge completion.

[0026] The multi-task neural network model is used to extract features from the point cloud, and a special detection head is designed to identify and classify noise points, so that the model can learn the geometric features, intensity features, and local neighborhood texture information of noise points, thereby accurately identifying discrete noise points in space and effectively filtering them out.

[0027] For point cloud data in rainy and foggy weather, the model learns the "weak and discontinuous point band" feature by identifying the reflection intensity variation and spatial discreteness characteristics of the points, thereby accurately labeling and removing rain or fog points, and improving the quality of point cloud data in adverse weather conditions.

[0028] Using the CNN backbone network and the multi-path feature modeling detection head, the model can identify and remove ghost or virtual points caused by multiple reflections by learning the multiple reflection points and delay time characteristics in the same direction, effectively solving the problem of separating real targets from false reflections in traditional methods.

[0029] By combining the angle, distance difference, and intensity features between the laser point and the sensing center, the model corrects and classifies suspicious reflection points in the point cloud. Through joint feature encoding of delay time and path deviation, the model can accurately identify and repair erroneous points caused by multipath reflection, thereby improving the accuracy of target detection.

[0030] For sparse edge regions in point cloud data, the model utilizes the spatial location, curvature, and other geometric features of points in the boundary region, and employs an interpolation detection head based on the UNet structure to generate and complete the point cloud, thereby enhancing the integrity and continuity of the point cloud data and providing a better data foundation for subsequent tasks such as object detection and semantic segmentation.

[0031] Compared with existing technologies, the lidar point cloud enhancement processing method provided in this embodiment effectively removes outliers and reduces noise interference by performing spatiotemporal registration, elevation filtering, and cluster density filtering on the initial point cloud. This provides a cleaner and more accurate data foundation for subsequent processing, helping to improve the overall accuracy of point cloud recognition. The multi-task neural network model can adaptively learn point cloud features under different environments, including noise, rain / fog points, ghosting / phantom images, and multipath reflections. It maintains a high effective point cloud retention rate under various environmental conditions (such as rain and fog), solving the problem of traditional methods relying on preset thresholds and having low effective point cloud retention rates in harsh environments. Point cloud enhancement processing can accurately identify and remove false point clouds caused by reflection and refraction, while retaining the point cloud data of real targets, reducing the false detection rate of targets. By setting classification labels and annotating the point cloud data to be enhanced, a labeled dataset is constructed, providing rich and diverse data resources for model training. This modular design allows the system to easily expand to new point cloud enhancement functions or adapt to new application scenarios; only the labeled dataset and model training strategy need to be adjusted.

[0032] Please see Figure 2 , Figure 2 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S103. In some embodiments of the present invention, a multi-task neural network model is used to perform point cloud augmentation processing on the initial data of the point cloud to be augmented, outputting augmented point cloud data, and further includes: S201. Extract the geometric features, intensity features and local neighborhood texture features of each point in the labeled dataset, output the noise confidence of each point through the first detection head, and mark and remove noise points. S202. Extract the reflection intensity variation features and spatial discrete features of each point in the labeled dataset, output the rain and fog probability map through the second detection head, and mark and remove all rain and fog points; S203, extract the delay time feature and the intensity attenuation feature of each point in the labeled data set, output the ghosting / shadowing probability map through the third detection head, mark and remove the ghosting / shadowing points; S204, extract the angle feature, distance difference feature and intensity feature of each point in the labeled data set, output the multipath reflection correction label through the fourth detection head, keep the target reflection points and correct other reflection points; S205, based on the fifth detection head, extract the geometric feature of the boundary region in the labeled data set, and generate the completed point cloud of the sparse region.

[0033] In the above embodiment, the labeled data set is a collection of point cloud data, which itself contains the three-dimensional coordinates, laser reflection intensity, echo delay time, perception center angle and distance information of each point cloud. From the labeled data set, the geometric feature, intensity feature and local neighborhood texture feature of each point are extracted. The first detection head (branch structure based on CNN) uses the extracted features to output the probability that each point belongs to a noise point through learned weights and biases. According to a preset confidence threshold (such as 0.7), the model marks points with confidence exceeding the threshold as noise points and removes them in subsequent processing, which can reduce random noise in the point cloud and improve data purity.

[0034] For point cloud data in rainy and foggy weather, the reflection intensity variation feature and spatial discreteness feature of each point are extracted. The second detection head uses the extracted features to generate a rain and fog probability map, which visually shows the likelihood that each point belongs to a rain or fog point, providing an intuitive basis for subsequent processing. According to the rain and fog probability map, points with a probability exceeding a preset threshold (such as 0.8) are marked as rain or fog points and removed in processing, reducing the interference of rain and fog weather on point cloud data and improving the usability of data in harsh environments.

[0035] The delay time feature and the intensity attenuation feature help to identify ghost or shadow points caused by multiple reflections. The third detection head uses the extracted features to generate a ghosting / shadowing probability map, which shows the likelihood that each point belongs to a ghost or shadow. According to the ghosting / shadowing probability map, points with a probability exceeding a preset threshold (such as 0.8) are marked as ghosting / shadowing points and removed in processing, thus solving the problem of difficult separation of real targets and false reflections in traditional methods and improving the accuracy of target detection.

[0036] Extracting the angle feature, distance difference feature and intensity feature of each point helps to identify false points caused by multipath reflection. The fourth detection head uses the extracted features to output a multipath reflection correction label, which labels whether each point belongs to a false point caused by multipath reflection and provides a correction suggestion. According to the correction label, false points caused by multipath reflection are corrected, keeping the reflection points of real targets and adjusting the positions or intensities of other reflection points to eliminate the influence of multipath reflection.

[0037] Based on the fifth detection head, the model extracts the geometric features of the boundary region in the labeled data set, which helps to identify the sparse points in the edge region and generate the completed point cloud. By using the extracted geometric features, the interpolation detection head generates a completed point cloud in the edge sparse region, which is seamlessly fused with the original point cloud data, enhancing the integrity and continuity of the point cloud data. The final output point cloud data not only contains high-quality point clouds after noise elimination, rain and fog point elimination, ghosting / ghosting elimination and multipath reflection correction, but also contains complete point clouds after edge region completion.

[0038] It can be understood that the enhancement processing of the point cloud data can also include other processing methods. Additional detection heads can be added to increase the functionality of other enhancement processing, or the above detection heads can be reduced according to actual use. That is, the point cloud enhancement processing in the present application includes but is not limited to noise points, rain and fog points, ghost points, multipath reflection points, environmental interference points and other non-real target generated point cloud data.

[0039] Please refer to Figure 3 , Figure 3 The flowchart of an embodiment of step S201 in the present application is provided. Figure 2 In some embodiments of the present application, the geometric features, intensity features and local neighborhood texture features of each point in the labeled data set are extracted, the noise point confidence of each point is output by the first detection head, and the noise points are marked and removed, and further comprising: S301, extracting the three-dimensional coordinates and laser reflection intensity of each point from the labeled data set, and calculating the geometric features of each point based on the three-dimensional coordinates; S302, determining the neighboring points within a predetermined range of each point, and extracting the local texture features according to the covariance matrix of the neighboring points; S303, the first detection head calculates the noise point confidence of each point according to the geometric features, laser reflection intensity and local neighborhood texture features of each point; S304, marking and removing all noise points according to the preset confidence threshold and the noise point confidence of each point.

[0040] In the above embodiment, the three-dimensional coordinates provide accurate position information of the point in space, and the laser reflection intensity reflects the reflection ability of the target surface to the laser. Based on the extracted three-dimensional coordinates, the geometric features of each point are calculated, including but not limited to the distance from the point to the origin of the laser radar, the vertical distance from the point to a certain reference plane, the normal direction of the point and the local curvature of the point, etc.

[0041] For each point, determine its neighboring points within a preset range, which can be adjusted according to the density and distribution characteristics of the point cloud data, usually a fixed radius or the number of neighboring points is determined based on the k-nearest neighbor algorithm. According to the determined neighboring points, the covariance matrix of these points is calculated, and the eigenvalues and eigenvectors of the covariance matrix can reflect the spatial distribution shape and direction of the neighboring points. Based on the eigenvalues and eigenvectors of the covariance matrix, local texture features are extracted, including but not limited to linearity, planarity, scattering degree, and anisotropy, etc.

[0042] The first detection head receives the geometric features, laser reflection intensity and local neighborhood texture features of each point as input, and fuses them through a fully connected layer or a convolutional layer inside the detection head to extract higher-level feature representations. Based on the fused feature representations, the first detection head outputs the noise point confidence of each point through a sigmoid activation function or a softmax function.

[0043] According to the actual application requirements and the characteristics of the point cloud data, a confidence threshold is preset for determining whether a point belongs to a noise point. Compare the noise point confidence of each point with the preset confidence threshold, if the confidence exceeds the threshold, mark the point as a noise point; otherwise, keep the point as a real target point. According to the marking result, all points marked as noise points are removed from the point cloud data, which can reduce random noise and outliers in the point cloud, and improve the purity of the data and the accuracy of subsequent processing.

[0044] Please refer to Figure 4 , Figure 4 The flowchart of an embodiment of step S202 in the Figure 2 In some embodiments of the present application, the reflection intensity variation feature and the spatial dispersion feature of each point in the labeled data set are extracted, and the rain and fog probability map is output by the second detection head to mark and remove all rain and fog points, which also includes: S401, extract the Euclidean distance between each point and its neighboring points from the labeled data set; S402, calculate the spatial dispersion feature according to the Euclidean distance; S403, determine the reflection intensity variation feature according to the laser reflection intensity; S404, the second detection head infers and outputs the rain and fog probability map according to the reflection intensity variation feature and the spatial dispersion feature of each point; S405, based on the rain and fog probability map and the preset rain and fog probability threshold, mark and remove all rain and fog points in the labeled data set.

[0045] In the above embodiment, the Euclidean distance between the target point and all neighboring points is calculated to form a distance set D={d1, d2, d i ,..., d n}, i=1,2,...,n, where d i This represents the distance between the target point and the i-th neighboring point. The distance set D is stored in association with other attributes of the target point.

[0046] Based on the distance set D, the mean distance, standard deviation of distance, and coefficient of variation of distance are calculated respectively. The mean distance reflects the overall density of points within a neighborhood, the standard deviation of distance measures the dispersion of the distance distribution, and the coefficient of variation of distance eliminates the influence of dimensions, directly comparing the dispersion of different neighborhoods. Finally, the statistics are combined into a spatial dispersion feature vector, which is used as one of the inputs of the second detection head.

[0047] For the target point and its neighboring points, extract the laser reflection intensity values ​​to form an intensity sequence I={I0, I1, I... i , ..., I n}, where I0 is the target point intensity, I i Let be the intensity of the i-th neighboring point. Further calculate the intensity difference, intensity gradient, and intensity entropy, and perform feature encoding. Combine the quantization results of the change pattern into a reflection intensity change feature vector, which serves as another input to the second detection head.

[0048] The second detection head employs a deep learning model (such as a Convolutional Neural Network (CNN) or a Graph Neural Network (GNN)) as input, using F_spatial and F_intensity as inputs. It learns the feature differences between rain / fog points and normal points through multiple nonlinear transformations. The detection head outputs the probability that each point belongs to the rain / fog point, forming a rain / fog probability map, where the pixel value corresponds to the rain / fog probability of that point.

[0049] Set a rain / fog probability threshold (e.g., 0.7) based on the actual application scenario for binary classification. Iterate through the rain / fog probability map; if the rain / fog probability of a point is greater than the set threshold, mark it as a rain / fog point; otherwise, retain it as a normal point. Remove all marked rain / fog points from the labeled dataset to generate purified point cloud data.

[0050] Please see Figure 5 , Figure 5 Provided by the present invention Figure 2 A flowchart illustrating an embodiment of step S203. In some embodiments of the present invention, the delay time features and intensity attenuation features of each point in the labeled dataset are extracted, a ghost / phantom probability map is output through a third detection head, and ghost / phantom points are marked and removed. The method further includes: S501. Extract the echo delay time of each point from the labeled dataset; S502. Determine the delay time characteristics of each point based on the echo delay time and the expected delay time; S503. Based on the analysis of laser reflection intensity and the attenuation over distance, extract the intensity attenuation characteristics; S504, the third detection head infers and outputs a ghost / shadow probability map according to the delay time feature and the intensity attenuation feature of each point; S505, based on the ghost / shadow probability map and a preset ghost / shadow probability threshold, all ghost / shadow points in the labeled data set are marked and removed.

[0051] In the above embodiment, the laser emission timestamp and the echo reception timestamp of each point are read from the labeled data set, and the echo delay time is calculated. If the point cloud data supports multiple echoes, the delay times of each echo are extracted respectively, and the echo sequence number is recorded. The delay time is associated with the spatial coordinates, reflection intensity and other attributes of the point to construct a structured data unit containing time-space-intensity information.

[0052] According to the ranging principle of the laser radar, the model is corrected in combination with environmental factors to improve the accuracy of the expected delay time. The deviation between the actual delay time and the expected delay time is calculated, and the absolute value, relative deviation and sign of the deviation are counted to reflect the degree of time domain anomaly of the point. If the data contains multiple echoes, the difference between the delay times of each echo is analyzed to capture the time layering phenomenon caused by multiple reflections. The standard deviation and coefficient of variation of the delay time are calculated to quantify the dispersion of the time distribution. The above features are combined to obtain a delay time feature vector.

[0053] According to the laser radar ranging equation, a theoretical attenuation relationship between reflection intensity and distance is established, the actual reflection intensity of the point is extracted from the data, the actual attenuation rate is calculated, and the difference between the actual attenuation rate and the theoretical attenuation rate is compared. In the local neighborhood, the standard deviation and coefficient of variation of the intensity of the neighboring points are calculated to capture the randomness of the intensity distribution. Finally, the attenuation features are combined into a vector to obtain the intensity attenuation feature.

[0054] The third detection head adopts a double-branch network structure. The time branch: taking the delay time feature as input, extracting the time domain feature through 2D convolution or 3D convolution or transformer architecture; the intensity branch: taking the intensity attenuation feature as input, extracting the intensity domain feature through similar structure. After fusing the features of the two branches, the probability of each point belonging to ghost / shadow is output. The ghost / shadow probabilities of all points are mapped to a two-dimensional plane to generate a ghost / shadow probability map, in which the pixel value corresponds to the probability of the point.

[0055] According to the ghost / shadow probability threshold set according to the actual application scenario (such as 0.8), a self-adaptive threshold method can be used to automatically determine the best threshold according to the probability map distribution. Traverse the probability map, if the probability of a point is greater than the set ghost / shadow probability threshold, mark it as a ghost / shadow point; otherwise, keep it as a normal point. Remove all marked points from the labeled data set to generate purified point cloud data.

[0056] Please see Figure 6 , Figure 6 Provided by the present invention Figure 2 A flowchart illustrating an embodiment of step S204. In some embodiments of the present invention, the included angle features, distance difference features, and intensity features of each point in the labeled dataset are extracted. A multipath reflection correction label is output through a fourth detection head, retaining the target reflection point and correcting other reflection points. The method also includes: S601. Extract the sensory center angle and distance information of each point from the labeled dataset; S602. Determine the angular relationship between the laser beam emission direction and the target point based on the angle between the sensing center, and obtain the angle characteristics; S603. Determine the distance difference characteristics based on the distance information and the distance difference with neighboring points; S604, the fourth detection head generates a multipath reflection correction label based on the included angle characteristics, distance difference characteristics and intensity characteristics of each point; S605, Based on multipath reflection correction label, retain the target reflection point and correct other reflection points.

[0057] In the above embodiments, using the lidar coordinate system as a reference, each reflection point is spatially located, and the angle between the line connecting it to the origin of the laser beam emission and the direction of the laser beam emission is calculated. This angle is used as the sensing center angle feature, reflecting the degree of deviation of the reflection point relative to the laser beam. Simultaneously, neighboring points of each point are quickly retrieved using a KD-tree or spatial hash algorithm, and the mean and standard deviation of the distance between each point and its neighbors are calculated to form distance information.

[0058] Based on the size of the angle between the sensory centers, the reflection points are divided into three categories: Principal reflection point: The included angle is less than half of the laser beam divergence angle, indicating that the reflection point is directly located on the main path of the laser beam; First-path reflection point: The included angle is between half and 1.5 times the divergence angle, and it may be that the laser beam enters the receiver after one surface reflection; High-order multipath reflection points: The included angle is greater than 1.5 times the divergence angle, usually caused by multiple reflections, resulting in lower signal reliability.

[0059] By leveraging the statistical characteristics of the angle distribution among neighboring points, the angle classification result of the current point is further validated, thereby improving feature accuracy.

[0060] Based on neighborhood distance information, the set of distance differences between each point and its neighbors is calculated, and the median, maximum value, and distribution pattern of the distance differences are statistically analyzed. Based on the distribution characteristics of the distance differences, the reflecting points are divided into: Uniform distribution: The distance difference is concentrated in a small area, indicating that the reflection points come from the same surface; Bimodal distribution: The distance difference has two distinct peaks, which may be caused by a mixture of primary reflection and primary multipath reflection; Long-tailed distribution: The distance difference range is large, which may include high-order multipath reflections or noise points.

[0061] It is understood that the ranges in the above distribution can be set according to the actual situation, and the present invention does not impose further restrictions on this.

[0062] The included angle feature, distance difference feature, and intensity feature are concatenated into a comprehensive feature vector, which is then input into a fourth detection head (such as a Transformer-based temporal feature aggregation network). Through a multi-task learning framework, the reflection point type and the corrected offset are predicted simultaneously. For primary reflection points, they are directly marked as "retained"; for primary multipath reflection points, their true position offset is predicted and a corrected label is generated; for higher-order multipath reflection points, they are marked as "filtered". A confidence threshold (e.g., 0.8) is used to filter high-confidence labels, reducing the risk of incorrect corrections.

[0063] All points marked as "primary reflection" are retained as the true reflection signals of the target object; for points marked as "first-order multipath", the predicted offset is applied to correct their position so that they are close to the true surface position; all points marked as "higher-order multipath" are removed to eliminate their interference with target detection.

[0064] The correction effect is verified by geometric consistency checks (e.g., the angle between the normal vector of the corrected point and the neighboring main reflection point is less than 15°) and intensity consistency checks (e.g., the intensity difference is less than 20%), ensuring the spatial continuity and physical rationality of the point cloud data.

[0065] Please see Figure 7 , Figure 7 Provided by the present invention Figure 2 A flowchart illustrating an embodiment of step S205. In some embodiments of the present invention, based on the extraction of geometric features of the boundary regions in the labeled dataset by the fifth detection head, and the generation of a complete point cloud of sparse regions, the method further includes: S701. Identify boundary points and obtain the boundary region based on the spatial distribution density and curvature changes of the labeled dataset; S702. Extract geometric features of the boundary region from the labeled dataset based on the fifth detection head; S703. Identify sparse regions in the boundary region based on the geometric features of the boundary region; S704. Establish the coordinates of the completion points for the sparse region predicted by the regression task and generate the completion point cloud to complete the sparse region.

[0066] In the above embodiment, first, the point cloud in the labeled data set is subjected to spatial density analysis, the number of neighboring points within a preset radius range around each point is counted, and if the number is less than 30% of the global average, it is marked as a potential boundary point. At the same time, the local curvature of each point is calculated, and if the curvature value is higher than 1.5 times of the global curvature median, it is further confirmed as a boundary point. Based on the spatial clustering of the boundary points, the continuously distributed boundary points are divided into independent boundary regions, and each region corresponds to the local edge or interface of an object.

[0067] The boundary region is input into the fifth detection head, and for each point in the boundary region, the neighborhood points are sampled with a radius of 0.2 meters, the normal vector distribution, spatial density gradient and curvature change pattern of the neighborhood points are calculated, the local features of all points in the boundary region are aggregated through an attention mechanism, and a global feature vector reflecting the overall geometric shape is generated. The local features and global features are combined to form comprehensive geometric features containing shape, curvature, density and direction information of the boundary region.

[0068] Based on the geometric features of the boundary region, if the point density of a sub-region (radius 0.2 meters) in the boundary region is less than 20% of the average density of the global boundary region, it is marked as a sparse candidate region; the standard deviation of the normal vector direction of the points in the sparse candidate region is calculated, and if the standard deviation is greater than 15°, it indicates that the normal vector distribution is dispersed, and the geometric information may be lost due to sparsity; if the curvature difference between the edge points of the sparse candidate region and the curvature of the adjacent non-sparse region exceeds 30%, the region is confirmed as a sparse region. For the identified sparse region, its boundary contour is further extracted, and the geometric features of the contour line are strengthened as input conditions for the completion task.

[0069] The geometric features of the boundary region and the contour features of the sparse region are input, a regression task is established to predict the coordinates of the completed points, the contour features of the sparse region and the global features of the boundary region are spliced, and a fully connected network is input to generate feature embedding; the three-dimensional coordinates of the completed points are predicted by a multilayer perception regression, and the spatial distribution constraint of the adjacent points of the sparse region is combined during prediction; the completed points are generated according to the predicted coordinates, and repeated points are filtered by non-maximum suppression; Laplace smoothing is performed on the completed point cloud to eliminate local noise, and the geometric consistency of the completed points and the original boundary points is verified to ensure that the completion result conforms to the real object surface distribution.

[0070] In order to better implement the laser radar point cloud enhancement processing method in the embodiment of the present application, on the basis of the laser radar point cloud enhancement processing method, please refer to Figure 8 , Figure 8 The structure diagram of an embodiment of the laser radar point cloud enhancement processing device provided by the present application is shown in the figure. The embodiment of the present application provides a laser radar point cloud enhancement processing device 800, which comprises: The acquisition processing module 810 is configured to perform space-time registration, height filtering and cluster density filtering processing on the initial point cloud collected by the laser radar, remove outliers to obtain target format point cloud data to be enhanced; The classification labeling module 820 is configured to set a classification label, label the point cloud data to be enhanced based on the classification label, and obtain a labeled data set; the labeled data set contains a plurality of data points. The model training module 830 is configured to input the labeled data set into an initial neural network model for training, and obtain a multi-task neural network model. The point cloud enhancement module 840 is configured to perform point cloud enhancement processing on the point cloud initial data to be enhanced by using the multi-task neural network model, and output enhanced point cloud data.

[0071] It should be noted that the device 800 provided in the above embodiments can implement the technical solutions described in the above method embodiments, and the principles of the implementation of the above modules or units can be referred to the corresponding content in the above method embodiments, which will not be described here.

[0072] Please refer to Figure 9 , Figure 9 The structure diagram of the laser radar point cloud enhancement processing device provided in the embodiments of the present application. Based on the laser radar point cloud enhancement processing method described above, the present application also correspondingly provides a laser radar point cloud enhancement processing device. The laser radar point cloud enhancement processing device can be a mobile terminal, a desktop computer, a notebook computer, a palm computer and a server, etc. The laser radar point cloud enhancement processing device 900 includes a processor 910, a memory 920 and a display 930. Figure 9 Only part of the components of the laser radar point cloud enhancement processing device are shown, but it should be understood that all the shown components are not required to be implemented, and more or less components can be alternatively implemented.

[0073] The memory 920 may, in some embodiments, be an internal storage unit of the lidar point cloud enhancement processing device 900, such as a hard disk or a memory of the lidar point cloud enhancement processing device 900. The memory 920 may, in other embodiments, also be an external storage device of the lidar point cloud enhancement processing device 900, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the lidar point cloud enhancement processing device 900. Further, the memory 920 may also include both an internal storage unit and an external storage device of the lidar point cloud enhancement processing device 900. The memory 920 is used to store application software installed on the lidar point cloud enhancement processing device 900 and various types of data, such as program codes installed on the lidar point cloud enhancement processing device 900. The memory 920 may also be used to temporarily store data that has been output or is to be output. In an embodiment, the memory 920 stores a lidar point cloud enhancement processing program 940, which can be executed by the processor 910 to implement the lidar point cloud enhancement processing method of the embodiments of the present application.

[0074] The processor 910 may, in some embodiments, be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program codes stored in the memory 920 or process data, such as executing the lidar point cloud enhancement processing method, etc.

[0075] The display 930 may, in some embodiments, be an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode) touch, etc. The display 930 is used to display information of the lidar point cloud enhancement processing device 900 and to display a visualized user interface. The components 910-930 of the lidar point cloud enhancement processing device 900 communicate with each other through a system bus.

[0076] In an embodiment, when the processor 910 executes the lidar point cloud enhancement processing program 940 in the memory 920, the steps in the above lidar point cloud enhancement processing method are implemented.

[0077] The embodiment also provides a computer readable storage medium having a lidar point cloud enhancement processing program stored thereon, which, when executed by a processor, implements the following steps: The initial point cloud collected by the lidar is subjected to space-time registration, elevation filtering, and cluster density filtering processing, and outliers are removed to obtain target format point cloud data to be enhanced; setting a classification label, labeling the to-be-enhanced point cloud data based on the classification label to obtain a labeled data set; the labeled data set contains a plurality of data points; inputting the labeled data set into an initial neural network model for training to obtain a multi-task neural network model; performing point cloud enhancement processing on the to-be-enhanced point cloud initial data by using the multi-task neural network model, and outputting enhanced point cloud data.

[0078] Please refer to Figure 10 , Figure 10 The framework schematic diagram of an embodiment of the laser radar point cloud enhancement processing method provided by the application. In combination with Figure 10 , in the specific process of the laser radar point cloud enhancement processing method provided by the application, the initial input laser radar data contains three-dimensional coordinates, reflection intensity and echo time data of the point cloud, at this time, the input is to-be-enhanced point cloud initial data, and at this time, the neural network model is also a trained multi-task neural network model, which can perform corresponding enhancement processing through five detection heads, according to Figure 10 The specific content of the corresponding detection head can also be known, and the application does not make too much repetition.

[0079] In summary, the laser radar point cloud enhancement processing method provided by the application effectively removes outliers and reduces noise interference by performing space-time registration, elevation filtering and cluster density filtering processing on the initial point cloud, provides a cleaner and more accurate data basis for subsequent processing, and helps to improve the overall accuracy of point cloud recognition. The multi-task neural network model can adaptively learn the point cloud features in different environments, including noise points, rain and fog points, ghost / virtual shadow, multi-path reflection, etc., and can maintain a high effective point cloud retention rate in different environmental conditions (such as rain and fog weather), solving the problem of low effective point cloud retention rate in harsh environments by traditional methods which rely on preset thresholds. The point cloud enhancement processing can accurately identify and remove false point clouds caused by reflection, refraction, etc., while retaining the point cloud data of real targets, reducing the false detection rate of target detection. By setting a classification label and labeling the to-be-enhanced point cloud data, a labeled data set is constructed, providing rich and diverse data resources for model training. This modular design makes it easy for the system to extend new point cloud enhancement functions or adapt to new application scenarios, only by adjusting the labeled data set and model training strategy.

[0080] The application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement the steps of the method. The computer readable storage medium can include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a micro drive, a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nanosystem (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.

[0081] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all described as a series of action combinations, but those skilled in the art should know that the application is not limited to the order of the actions described, because according to the application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the application.

[0082] In the above embodiments, the description of each embodiment is focused on, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0083] In several embodiments provided by the present application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely schematic. The division of the units is merely a logical function division. In actual implementation, another division mode can be adopted. For example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some service interface, device or unit, and can be electrical or other forms.

[0084] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0085] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically independently, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0086] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable memory. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a memory and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned memory includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0087] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by a program instructing relevant hardware, and the program can be stored in a computer readable memory, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0088] The above is only exemplary embodiments of the present disclosure, and cannot limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Those skilled in the art will easily think of embodiments of the present disclosure after considering the specification and practicing the disclosure herein. The present application is intended to cover any variations, uses or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and examples are only considered as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

[0089] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, however, as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present disclosure.

[0090] Those skilled in the art readily understand that the above only describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for laser radar point cloud enhancement processing, characterized in that, The method comprises the following steps: spatially and temporally registering, height filtering and cluster density filtering the initial point cloud collected by the laser radar, removing outliers to obtain target format point cloud data to be enhanced; setting a classification label, labeling the point cloud data to be enhanced based on the classification label, and obtaining a labeled data set; the labeled data set contains a plurality of data points; inputting the labeled data set into an initial neural network model for training to obtain a multi-task neural network model; using the multi-task neural network model to perform point cloud enhancement processing on the initial point cloud data to be enhanced, and outputting enhanced point cloud data.

2. The LIDAR point cloud enhancement processing method of claim 1, wherein, The method of using the multi-task neural network model to perform point cloud enhancement processing on the initial point cloud data to be enhanced and outputting enhanced point cloud data further comprises: extracting the geometric features, intensity features and local neighborhood texture features of each point in the labeled data set, outputting the noise point confidence of each point through the first detection head, and marking and removing noise points; extracting the reflection intensity variation feature and spatial discreteness feature of each point in the labeled data set, outputting the rain and fog probability map through the second detection head, and marking and removing all rain and fog points; extracting the delay time feature and intensity attenuation feature of each point in the labeled data set, outputting the ghosting / ghosting probability map through the third detection head, and marking and removing ghosting / ghosting points; extracting the angle feature, distance difference feature and intensity feature of each point in the labeled data set, outputting the multipath reflection correction label through the fourth detection head, and retaining target reflection points and correcting other reflection points; based on the fifth detection head, extracting the geometric features of the boundary region in the labeled data set, and generating the completed point cloud of the sparse region.

3. The LIDAR point cloud enhancement processing method of claim 2, wherein, The method of extracting the geometric features, intensity features and local neighborhood texture features of each point in the labeled data set, outputting the noise point confidence of each point through the first detection head, and marking and removing noise points further comprises: extracting the three-dimensional coordinates and laser reflection intensity of each point from the labeled data set, and calculating the geometric features of each point based on the three-dimensional coordinates; determining the neighboring points within a predetermined range of each point, and extracting the local texture features according to the covariance matrix of the neighboring points; the first detection head calculates the noise point confidence of each point according to the geometric features, laser reflection intensity and local neighborhood texture features of each point; according to the preset confidence threshold and the noise point confidence of each point, all noise points are marked and removed.

4. The LIDAR point cloud enhancement processing method of claim 3, wherein, The method of extracting the reflection intensity variation feature and spatial discreteness feature of each point in the labeled data set, outputting the rain and fog probability map through the second detection head, and marking and removing all rain and fog points further comprises: extracting the Euclidean distance between each point and its neighboring points from the labeled data set; calculating the spatial discreteness feature according to the Euclidean distance; determining the reflection intensity variation feature according to the laser reflection intensity; the second detection head infers and outputs the rain and fog probability map according to the reflection intensity variation feature and the spatial discreteness feature of each point; based on the rain and fog probability map and the preset rain and fog probability threshold, all rain and fog points in the labeled data set are marked and removed.

5. The LIDAR point cloud enhancement processing method of claim 3, wherein, The method further comprises: extracting echo delay time of each point from the labeled data set; determining delay time feature of each point based on the echo delay time and expected delay time; extracting intensity decay feature according to the analysis of laser reflection intensity and distance decay; the third detection head infers and outputs ghosting / ghosting probability map according to the delay time feature and intensity decay feature of each point; based on the ghosting / ghosting probability map and the preset ghosting / ghosting probability threshold, all ghosting / ghosting points in the labeled data set are marked and removed.

6. The LIDAR point cloud enhancement processing method of claim 3, wherein, The method further comprises: extracting the angle feature, distance difference feature and intensity feature of each point in the labeled data set, outputting the multipath reflection correction label through the fourth detection head, retaining the target reflection point and correcting other reflection points, and further comprising: extracting the perception center angle and distance information of each point from the labeled data set; determining the angle relationship between the emission direction of the laser beam and the target point according to the perception center angle to obtain the angle feature; determining the distance difference feature according to the distance difference between the distance information and the adjacent points; the fourth detection head generates the multipath reflection correction label according to the angle feature, distance difference feature and intensity feature of each point; 7. The LIDAR point cloud enhancement processing method of claim 2, wherein, based on the multipath reflection correction label, the target reflection point is retained and other reflection points are corrected. The method further comprises: identifying the boundary points according to the spatial distribution density and curvature change of the labeled data set to obtain the boundary region; the fifth detection head extracts the geometric feature of the boundary region from the labeled data set; identifying the sparse region in the boundary region according to the geometric feature of the boundary region; 8. A lidar point cloud enhancement processing apparatus, characterized by, establishing a regression task to predict the coordinates of the completed points in the sparse region and generating a completed point cloud to complete the sparse region. The method comprises: a collection processing module configured to perform temporal and spatial registration, elevation filtering and cluster density filtering processing on the initial point cloud collected by the laser radar, remove outliers to obtain target format to-be-enhanced point cloud data; a classification labeling module configured to set a classification label, label the to-be-enhanced point cloud data based on the classification label, and obtain a labeled data set; the labeled data set contains a plurality of data points; a model training module configured to input the labeled data set into an initial neural network model for training to obtain a multi-task neural network model; 9. A lidar point cloud enhancement processing device, comprising: a point cloud enhancement module configured to perform point cloud enhancement processing on to-be-enhanced point cloud initial data using the multi-task neural network model, and output enhanced point cloud data. The method comprises at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit executes the steps of the laser radar point cloud enhancement processing method in any one of claims 1-7.

10. A storage medium, characterized by It stores a computer program executable by the laser radar point cloud enhancement processing device, which, when running on the laser radar point cloud enhancement processing device, causes the laser radar point cloud enhancement processing device to execute the steps of the laser radar point cloud enhancement processing method of any one of claims 1-7.

Images