A method for simultaneous target association, identification and positioning

Abstract

The invention discloses a method for simultaneous target association recognition and positioning, relates to the field of artificial intelligence technology, and solves the technical problem that the prior art estimates the positioning of a vehicle through the Kalman pose based on probability data association of a laser radar, but does not consider that the Kalman filter has a large amount of calculation when processing complex models and observation data, which may cause target recognition errors or failure to recognize. The invention obtains a point cloud image; obtains clustering contours by extracting clusters of the point cloud; performs Kalman filtering on the clustering contours that are closest to each other; selects the optimal cutting point, and uses a multi-threading method to process the clustering contours to be cut; extracts the nearest clustering contour, traverses to obtain a normal plane, and averages the point cloud data in the normal plane to obtain a mean; sets a trapezoidal transformation threshold to perform a trapezoidal transformation to construct a trapezoidal recognition area; locates the target to be identified according to the trapezoidal recognition area; identifies the target through a target association mapping method, and improves the positioning accuracy.

Description

Technical Field

[0001] The invention belongs to the field of artificial intelligence and relates to identification and positioning technology, in particular to a target simultaneous association identification and positioning method. Background Art

[0002] Target recognition and positioning usually require high real-time performance, and can quickly and accurately identify the target and determine its location, which is crucial for tasks that require rapid response; simultaneous associative recognition and positioning can reduce the errors that may be introduced in step-by-step processing; traditional step-by-step processing methods may identify the target first and then locate it, and the errors between these two steps may accumulate with each other, while simultaneous processing can effectively avoid this problem; for complex and changing environments, target recognition and positioning can help the system quickly adapt and respond to changes, and improve the robustness and stability of the system; in intelligent monitoring systems, target recognition and positioning can achieve rapid detection and early warning of abnormal behavior, and improve the safety and reliability of monitoring systems.

[0003] The prior art (the invention patent application with publication number CN114527480A) discloses a method for accurately positioning ground handling vehicles in complex airport environments; the prior art uses a roadside camera to identify a vehicle according to the vehicle's autonomous positioning information and motion status when the vehicle enters a non-ideal area; a roadside lidar locks the target, and the Kalman filter pose estimation based on probabilistic data association eliminates interference from non-detected targets to locate the vehicle; a collaborative fault-tolerant fusion algorithm is used according to environmental characteristics to achieve high-precision positioning; the prior art uses a lidar to estimate the vehicle's position based on the Kalman pose of probabilistic data association, but does not take into account that the Kalman filter has a large amount of computation when processing complex models and observation data, which may lead to target recognition errors or failure to recognize the target.

[0004] The present invention provides a target simultaneous association identification and positioning method to solve the above technical problems. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a method for simultaneous target association, identification and positioning, which is used to solve the technical problem that the prior art estimates the positioning of a vehicle through the Kalman pose based on probability data association by a laser radar, but does not take into account the large amount of calculation of the Kalman filter when processing complex models and observation data, which may lead to target recognition errors or unrecognition.

[0006] To achieve the above object, a first aspect of the present invention provides a method for simultaneously associating, identifying and locating a target, comprising:

[0007] S100: acquiring a point cloud image by a visual method; performing multi-source coupling clustering analysis on the point cloud in the point cloud image to obtain clusters of the point cloud; obtaining cluster contours by extracting the clusters of the point cloud; performing Kalman filtering on the cluster contours that are closest to each other; adding all cluster contour points to a set of to-be-selected cutting points to obtain clusters to be cut, wherein the point cloud includes: a laser radar point cloud and a visual point cloud;

[0008] S200: Select the optimal cutting point and use a multi-threading method to process the cluster outline to be cut;

[0009] S300: After extracting the nearest cluster outline, traverse to obtain the normal plane, average the point cloud data in the normal plane to obtain the mean; set the trapezoidal transformation threshold to perform trapezoidal transformation to construct a trapezoidal recognition area; dynamically locate the target to be identified according to the trapezoidal recognition area; and identify the target through the target association mapping method.

[0010] Preferably, performing multi-source coupling clustering analysis on the point cloud in the point cloud image to obtain clustering of the point cloud includes:

[0011] Retrieve the point cloud in the point cloud image, number the point clouds from different sources and mark them as i; mark the position accuracy of the point cloud as Wi, and the density as Mi, where i = 1, 2, 3, ..., n; n is an integer greater than 0;

[0012] The data coupling weight is calculated by the formula SQi=α×ln(Wi)+β×ln(Mi); wherein α and β are proportional coefficients greater than 0 and less than 1;

[0013] The point cloud is clustered by forming a specific adaptive clustering analysis method based on the data coupling weights.

[0014] By considering the data coupling weight, the present invention can optimize the characteristics of a specific data set to form a more targeted clustering method; this specificity makes the clustering result closer to the real structure of the data, and improves the accuracy and effectiveness of clustering; the data coupling weight reflects the degree of association between data points, and clustering using the data coupling weight can more accurately divide the boundaries between data points.

[0015] Preferably, the step of extracting cluster outlines from clusters of point clouds comprises:

[0016] Retrieve the point cloud image, pre-process the point cloud image, and obtain the key points from the point cloud image; perform contour outlining based on the key points, find the nearest cluster based on the number of refined neighboring points, and determine whether the nearest cluster disappears after Gaussian filtering; if yes, re-contour the point cloud based on the key points; if not, continue to determine.

[0017] The present invention first performs cluster analysis on the nearest point cloud, and then repeats the cluster analysis on other point clouds according to the number of neighboring points after refinement. If the cluster disappears after filtering, the cluster outlines of all point clouds are gradually obtained. The cyclic repetitive process can perform cluster analysis on all point clouds without omission. A large amount of data is processed through cluster analysis, which lays the foundation for the subsequent acquisition of the outlines of all point clouds.

[0018] Preferably, the preprocessing of the point cloud image includes:

[0019] S111: using a Gaussian filtering algorithm to perform a convolution operation on the operator and the point cloud image, performing a smoothing filter on the original image, and obtaining an output image with Gaussian noise eliminated;

[0020] S112: performing a convolution operation on the output image of S111 and the gradient operator to obtain a gradient magnitude and a gradient direction;

[0021] S113: Perform non-maximum suppression processing on the output image, set a segmentation line with the pixel in the point cloud as the origin, and determine whether the pixel value of the origin is greater than the pixel values ​​of two adjacent pixels along the segmentation line; if yes, retain the pixel value of the origin; if not, return to zero.

[0022] It should be noted that the dividing lines are bidirectional dividing lines that are respectively 45 degrees and 135 degrees and pass through the origin.

[0023] The present invention can effectively reduce image noise through the Gaussian filtering algorithm, and can also retain the edge information of the image to a certain extent, providing a clearer image basis for subsequent processing; by calculating the gradient amplitude and gradient direction of the image, the edge information in the image can be highlighted, and the strength and direction of the image edge can be better represented.

[0024] Preferably, adding all clustering profiles to the set of cut points to be selected comprises:

[0025] Retrieve pixel values ​​and set high and low threshold calculation factors; mark the high threshold calculation factor as Th and the low threshold calculation factor as Tl; mark pixel points corresponding to pixel values ​​higher than Th as strong edge points, set pixel points corresponding to pixel values ​​less than Tl to 0, and set pixel points corresponding to pixel values ​​between the high and low threshold calculation factors and not connected to the strong edge points to 0, and finally obtain the cluster contour points at the edge of the point cloud, and record the cluster contour points in the set of cut points to be selected.

[0026] The present invention can effectively remove most of the noise edges by selecting a high threshold, significantly reducing the impact of noise on edge detection results; the selection of a low threshold helps to retain more edge information, ensuring that edge points with weak edge strength but still important features of the image are not missed; the optimal cutting point is accurately located by combining the Euclidean distance with the dynamic threshold, and recognition is performed through a deep learning model, which can avoid the complex process of a series of curvature calculations.

[0027] Preferably, the selecting the optimal cutting point includes:

[0028] Retrieve a set of candidate cutting points, divide the candidate cutting points into multiple subsets according to the clustering algorithm, count the number of candidate cutting points, calculate the Euclidean distance from the center point of each subset, and select the candidate cutting point corresponding to the minimum Euclidean distance as the optimal cutting point of the overlapping point cloud;

[0029] Determine whether there is more than one optimal cutting point; if yes, reselect the minimum distance value; if no, slice the point cloud according to the optimal cutting point.

[0030] It should be noted that the number of slices obtained by slicing the point cloud according to the optimal cutting point is equal to the number of threads used to process the cluster contours to be cut using the multi-threading method.

[0031] By selecting the point corresponding to the minimum Euclidean distance in each subset as the candidate optimal cutting point, the present invention can ensure that the selected cutting point is optimal within its local range, thereby improving the overall accuracy and efficiency; when processing point cloud data of complex scenes, cluster analysis can better distinguish potential cutting points or overlapping areas, and select the optimal cutting point for each area.

[0032] Preferably, the dynamically positioning the target to be identified according to the trapezoidal identification area includes:

[0033] Retrieve the trapezoidal recognition area, set the trapezoidal area scanning area for the target to be identified; set the threshold in the trapezoidal area scanning area through the sliding window; retrieve the mean, perform D2C transformation on the mean to convert it into the distance mean; determine whether the distance mean is less than the threshold; if yes, use Yolov8 to dynamically locate the target to be identified entering the trapezoidal recognition area; if no, continue to determine.

[0034] It should be noted that the projected pixel area of ​​the actual three-dimensional point cloud area in the image plane is a nearly trapezoidal area; the trapezoidal area parameters are dynamically calculated values.

[0035] By setting a trapezoidal recognition area, the present invention can limit the detection range, reduce background noise and interference from irrelevant objects, thereby improving the accuracy of target detection, and greatly reduce the amount of data that needs to be processed, thereby improving processing efficiency; compared with indiscriminate detection of the entire image, this method of setting a trapezoidal recognition area can locate the target object more quickly.

[0036] Preferably, the target identification by target association mapping method comprises:

[0037] The nearest cluster and the second nearest cluster are obtained through the target recognition area threshold, and the similarity between the nearest cluster and the second nearest cluster and the target to be identified is compared to obtain the category to which the specific target belongs; the target to be identified is semantically annotated to construct a target semantic triple, and the characteristics of the target to be identified at the semantic level are determined through the target semantic triple, and the target to be identified is recognized based on the characteristics.

[0038] It should be noted that the target semantic triple format is <category, relationship, feature>, where category is the target classification, relationship is the relationship between the target to be identified and the feature, and feature is the set of associated features of the target to be identified; relationship types are divided into: inclusion, unidirectional association, bidirectional association, dependency, and nested relationship.

[0039] The present invention can clearly specify the characteristics of the target at the semantic level by semantically annotating the target to be identified and establishing the target semantic triple. The semantic triple clearly defines the semantics of the subject, predicate and object, making the association relationship between the target and the characteristics clearer and more accurate.

[0040] Preferably, obtaining the closest cluster and the second closest cluster by using the target identification area threshold comprises:

[0041] The threshold of the target recognition area is determined through deep learning, the mean is retrieved and the variance of the point cloud data in the normal plane is calculated based on the mean; it is judged whether the variance of the point cloud data in the normal plane is less than the threshold of the target recognition area. If so, the target to be recognized is recognized through the target association mapping method; if not, the nearest cluster is reselected for secondary filtering to obtain the next nearest cluster.

[0042] The present invention determines the threshold of the target identification area by utilizing a deep learning algorithm, which can fully consider the complexity and diversity of the data, thereby setting the threshold range more accurately; by calculating the variance of the point cloud data in the normal plane, the discreteness and distribution of the data can be evaluated; wherein, if the variance is less than the threshold of the target identification area, it means that the data points are more concentrated and have a greater probability of belonging to the same target or area, thereby improving the accuracy of target identification.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. The present invention combines lidar point cloud and visual point cloud for multi-source coupling clustering analysis, which can make full use of the advantages of different data sources and improve the accuracy and robustness of point cloud clustering; lidar point cloud has high precision and high density, while visual point cloud may contain more color and texture information. The combination of the two can generate more comprehensive point cloud data; by extracting cluster contours, the target area can be clearly defined, and the interference of background noise and other non-target objects can be reduced, thereby improving the accuracy of target recognition.

[0045] 2. The present invention can smooth the noise and outliers in the point cloud data and improve the stability and reliability of the data by performing Kalman filtering on the cluster contours that are closest to each other; this filtering method is particularly effective when processing dynamic scenes and complex environments; using a multi-threaded method to process the cluster contours to be cut can significantly improve the processing speed while reducing delays and waiting time in the processing process; this is especially important for application scenarios with high real-time requirements; dynamically positioning the target to be identified based on the trapezoidal recognition area and identifying the target through the target association mapping method can clearly define the position and state of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A schematic diagram of the identification and positioning method of the present invention;

[0048] Figure 2 A schematic diagram of the target association processing flow of the present invention;

[0049] Figure 3 This is a schematic diagram of the specific process of identifying and positioning the present invention. DETAILED DESCRIPTION

[0050] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] See also Figure 1 The first aspect of the present invention provides a method for simultaneously associating, identifying and locating a target, including:

[0052] S100: acquiring a point cloud image by a visual method; performing multi-source coupling clustering analysis on the point cloud in the point cloud image to obtain clusters of the point cloud; obtaining cluster contours by extracting the clusters of the point cloud; performing Kalman filtering on the cluster contours that are closest to each other; adding all cluster contour points to a set of to-be-selected cutting points to obtain clusters to be cut, wherein the point cloud includes: a laser radar point cloud and a visual point cloud;

[0053] S200: Select the optimal cutting point and use a multi-threading method to process the cluster outline to be cut;

[0054] S300: After extracting the nearest cluster outline, traverse to obtain the normal plane, average the point cloud data in the normal plane to obtain the mean; set the trapezoidal transformation threshold to perform trapezoidal transformation to construct a trapezoidal recognition area; dynamically locate the target to be identified according to the trapezoidal recognition area; and identify the target through the target association mapping method.

[0055] See also Figure 2 , target association processing flow, obtain point cloud images through visual methods; retrieve point clouds in point cloud images, number point clouds from different sources and mark them as i; mark the position accuracy of point clouds as Wi, and mark the density as Mi, where i = 1, 2, 3, ..., n; n is an integer greater than 0;

[0056] It should be noted that visual methods generally use two or more cameras to capture images of the same scene from different angles, calculate the depth information of the object by comparing the parallax in the images, and then generate a three-dimensional point cloud.

[0057] The data coupling weight is calculated by the formula SQi=α×ln(Wi)+β×ln(Mi); wherein α and β are proportional coefficients greater than 0 and less than 1; a specific adaptive clustering analysis method is formed according to the data coupling weight to obtain the clustering of the point cloud;

[0058] The point cloud image is preprocessed, and the operator is convolved with the point cloud image using the Gaussian filtering algorithm, and the original image is smoothed and filtered to obtain an output image with Gaussian noise eliminated; the output image is convolved with the gradient operator to obtain the gradient amplitude and gradient direction;

[0059] Perform non-maximum suppression processing on the output image, set the segmentation line with the pixel in the point cloud as the origin, and judge whether the pixel value of the origin is greater than the pixel values ​​of the two adjacent pixels along the segmentation line; if yes, keep the pixel value of the origin; if not, return it to zero;

[0060] Obtain key points from the point cloud image; perform contour outlining based on the key points, find the nearest cluster based on the number of refined neighboring points, and determine whether the nearest cluster disappears after Gaussian filtering; if yes, re-contour the point cloud based on the key points; if no, continue to determine;

[0061] Retrieve pixel values ​​and set high and low threshold calculation factors; mark the high threshold calculation factor as Th and the low threshold calculation factor as Tl; mark pixel points corresponding to pixel values ​​higher than Th as strong edge points, set pixel points corresponding to pixel values ​​lower than Tl to 0, and set pixel points corresponding to pixel values ​​between the high and low threshold calculation factors and not connected to strong edge points to 0, and finally obtain cluster contour points at the edge of the point cloud, and record the cluster contour points in the set of cut points to be selected;

[0062] Retrieve a set of candidate cutting points, divide the candidate cutting points into multiple subsets according to the clustering algorithm, count the number of candidate cutting points, calculate the Euclidean distance from the center point of each subset, and select the candidate cutting point corresponding to the minimum Euclidean distance as the optimal cutting point of the overlapping point cloud;

[0063] Determine whether there is more than one optimal cutting point; if yes, reselect the minimum distance value; if no, slice the point cloud according to the optimal cutting point.

[0064] For example, point cloud 1 and point cloud 2 are obtained by visual method and lidar method, and the position accuracy of point cloud 1 is marked as 10mm and the density is marked as 15ppm. 3 ; The position accuracy of point cloud 2 is marked as 8mm, and the density is marked as 20ppm 3 , set α and β to 0.1; calculate the data coupling weight of point cloud 1 SQ1 = 0.1 × ln10 + 0.1 × ln15 = 0.501 through the formula SQi = α × ln(Wi) + β × ln(Mi); the data coupling weight of point cloud 2 SQ1 = 0.1 × ln8 + 0.1 × ln20 = 0.506; form a specific adaptive clustering analysis method based on the data coupling weight to obtain the clustering of point clouds;

[0065] Get 5 pixel values ​​and record them as 8, 13, 12, 5, and 0.5 respectively; set the high threshold calculation factor to 10 and the low threshold calculation factor to 1; mark the pixel points corresponding to pixel values ​​higher than 10 as strong edge points, that is, the pixel points corresponding to pixel values ​​13 and 12 are marked as strong edge points; set the pixel points corresponding to pixel values ​​less than 1 to 0, that is, the pixel points corresponding to the pixel value of 0.5 are set to 0; set the pixel points corresponding to the pixel values ​​between the high and low threshold calculation factors and the points not connected to the strong edge points to 0, that is, the pixel points corresponding to the pixel values ​​8 and 5 are set to 0; finally, obtain the cluster contour points at the edge of the point cloud, and record the cluster contour points in the set of selected cutting points.

[0066] See also Figure 3, the specific process of identification and positioning, call the trapezoidal identification area, set the trapezoidal area scanning area for the target to be identified; set the threshold in the trapezoidal area scanning area through the sliding window; call the mean, and perform D2C transformation on the mean to the distance mean; judge whether the distance mean is less than the threshold; if yes, use Yolov8 to dynamically locate the target to be identified entering the trapezoidal identification area; if no, continue to judge;

[0067] The nearest cluster and the second nearest cluster are obtained by using the target recognition area threshold, and the similarity between the nearest cluster and the second nearest cluster and the target to be recognized is compared to obtain the category to which the specific target belongs; the target to be recognized is semantically annotated to construct a target semantic triple, and the features of the target to be recognized at the semantic level are determined by the target semantic triple, and the target to be recognized is recognized based on the features;

[0068] The threshold of the target recognition area is determined through deep learning, the mean is retrieved and the variance of the point cloud data in the normal plane is calculated based on the mean; it is judged whether the variance of the point cloud data in the normal plane is less than the threshold of the target recognition area. If so, the target to be recognized is recognized through the target association mapping method; if not, the nearest cluster is reselected for secondary filtering to obtain the next nearest cluster.

[0069] For example, semantic annotation is performed on the target to be identified to construct a target semantic triple; among which, the relationship types in the target semantic triple are divided into: inclusion, unidirectional association, bidirectional association, dependency, and nested relationship;

[0070] A: Example of nested relationship, <goal, relationship type, <sub-goal <grandchild goal>>>;

[0071] B: Example of association relationship, <target = stairs, relationship = <contains, associates>, feature = <steps <steps, contains, outlines>, handrails <handrails, associates, edge grids>>>;

[0072] The actual form of the triple is <stairs, contains, steps>, <stairs, one-way association, handrail>; the one-way association is a possible association, and the staircase has steps but no handrails, that is, the staircase may be associated with handrails, and the handrail feature is not necessarily associated with the staircase target;

[0073] C: Examples of inclusion relations, <glass door, including, door frame>, <door frame, including, glass>.

[0074] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0075] The working principle of the present invention is as follows: the present invention obtains a point cloud image through a visual method; performs multi-source coupling clustering analysis on the point cloud in the point cloud image to obtain the clustering of the point cloud; obtains the clustering contour by extracting the clustering of the point cloud; performs Kalman filtering on the clustering contours that are closest to each other; adds all clustering contour points to a set of to-be-selected cutting points to obtain clusters to be cut, wherein the point cloud includes: a lidar point cloud and a visual point cloud; selects the optimal cutting point, and uses a multi-threading method to process the clustering contours to be cut; extracts the nearest clustering contour, traverses to obtain a normal plane, and averages the point cloud data in the normal plane to obtain a mean; sets a trapezoidal transformation threshold to perform a trapezoidal transformation to construct a trapezoidal recognition area; dynamically locates the target to be identified according to the trapezoidal recognition area; and identifies the target through a target association mapping method.

[0076] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for simultaneous target association, identification and positioning, characterized in that: S100: acquiring a point cloud image by a visual method; performing multi-source coupling clustering analysis on the point cloud in the point cloud image to obtain clusters of the point cloud; and obtaining cluster outlines by extracting clusters of the point cloud; Perform Kalman filtering on the closest cluster outlines; Add all cluster contour points to the set of to-be-selected cutting points to obtain the clusters to be cut, where the point cloud includes: lidar point cloud and visual point cloud; S200: Select the optimal cutting point and use a multi-threading method to process the cluster outline to be cut; S300: After extracting the nearest cluster outline, traverse to obtain the normal plane, average the point cloud data in the normal plane to obtain the mean value; set the trapezoidal transformation threshold to perform trapezoidal transformation to construct a trapezoidal recognition area; dynamically locate the target to be identified according to the trapezoidal recognition area; identify the target through the target association mapping method; The step of performing multi-source coupling clustering analysis on the point cloud in the point cloud image to obtain clustering of the point cloud includes: Retrieve the point cloud in the point cloud image, number the point clouds from different sources and mark them as i; mark the position accuracy of the point cloud as Wi, and the density as Mi, where i=1,2,3,…,n; n is an integer greater than 0; The data coupling weight is calculated by the formula SQi=α×ln(Wi)+β×ln(Mi); where α and β are proportional coefficients greater than 0 and less than 1; The point cloud is clustered by forming a specific adaptive clustering analysis method based on data coupling weights; The selecting the optimal cutting point comprises: Retrieve a set of candidate cutting points, divide the candidate cutting points into multiple subsets according to the clustering algorithm, count the number of candidate cutting points, calculate the Euclidean distance from the center point of each subset, and select the candidate cutting point corresponding to the minimum Euclidean distance as the optimal cutting point of the overlapping point cloud; Determine whether there is more than one optimal cutting point; if yes, reselect the minimum distance value; if no, slice the point cloud according to the optimal cutting point; The dynamically positioning the target to be identified according to the trapezoidal identification area includes: Retrieve the trapezoidal recognition area, set the trapezoidal area scanning area for the target to be identified; set the threshold in the trapezoidal area scanning area through the sliding window; retrieve the mean, perform D2C transformation on the mean to convert it into the distance mean; determine whether the distance mean is less than the threshold; if yes, use Yolov8 to dynamically locate the target to be identified that enters the trapezoidal recognition area; if no, continue to determine; The target identification is performed by the target association mapping method, including: The nearest cluster and the second nearest cluster are obtained through the target recognition area threshold, and the similarity between the nearest cluster and the second nearest cluster and the target to be identified is compared to obtain the category to which the specific target belongs; the target to be identified is semantically annotated to construct a target semantic triple, and the characteristics of the target to be identified at the semantic level are determined through the target semantic triple, and the target to be identified is recognized based on the characteristics.

2. A method for simultaneous target association identification and positioning according to claim 1, characterized in that: The step of extracting cluster outlines from clusters of point clouds includes: Retrieve the point cloud image, pre-process the point cloud image, and obtain the key points from the point cloud image; perform contour outlining based on the key points, find the nearest cluster based on the number of refined neighboring points, and determine whether the nearest cluster disappears after Gaussian filtering; if yes, re-contour the point cloud based on the key points; if not, continue to determine.

3. A method for simultaneous target association identification and positioning according to claim 2, characterized in that: The preprocessing of the point cloud image includes: S111: using a Gaussian filtering algorithm to perform a convolution operation on the operator and the point cloud image, performing a smoothing filter on the original image, and obtaining an output image with Gaussian noise eliminated; S112: performing a convolution operation on the output image of S111 and the gradient operator to obtain a gradient magnitude and a gradient direction; S113: Perform non-maximum suppression processing on the output image, set a segmentation line with the pixel in the point cloud as the origin, and determine whether the pixel value of the origin is greater than the pixel values ​​of two adjacent pixels along the segmentation line; if yes, retain the pixel value of the origin; if not, return to zero.

4. The method for simultaneous target association identification and positioning according to claim 1, characterized in that: The step of adding all cluster contour points to the set of cutting points to be selected includes: Retrieve pixel values ​​and set high and low threshold calculation factors; mark the high threshold calculation factor as Th and the low threshold calculation factor as Tl; mark pixel points corresponding to pixel values ​​higher than Th as strong edge points, set pixel points corresponding to pixel values ​​less than Tl to 0, and set pixel points corresponding to pixel values ​​between the high and low threshold calculation factors and not connected to the strong edge points to 0, and finally obtain the cluster contour points at the edge of the point cloud, and record the cluster contour points in the set of cut points to be selected.

5. The method for simultaneous target association identification and positioning according to claim 1, characterized in that: The obtaining of the nearest cluster and the second nearest cluster by the target identification area threshold comprises: The threshold of the target recognition area is determined through deep learning, the mean is retrieved and the variance of the point cloud data in the normal plane is calculated based on the mean; it is judged whether the variance of the point cloud data in the normal plane is less than the threshold of the target recognition area. If so, the target to be recognized is recognized through the target association mapping method; if not, the nearest cluster is reselected for secondary filtering to obtain the next nearest cluster.

Images