A method for intelligent fusion and completion of target information under perception data loss

Abstract

The application discloses a kind of target information intelligent fusion completion method under the perception data missing, belong to multi-sensor information fusion field.The application includes the following main steps: multi-sensor data acquisition and pre-processing, dimension cutting and sub data set construction, feature sub data set clustering for each feature dimension, position information construction position data set based on each feature clustering result, based on maximum clustering quantity clustering again and feature information summary and mean completion.The application can solve the problem of incomplete information fusion data caused by target feature information missing in multi-sensor perception process, complete and fuse information to form complete target information, and the fusion completion process has high degree of automation, with engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of multi-sensor information fusion, and in particular, it is a method for intelligent fusion and completion of target information under the condition of missing sensing data. Background Technology

[0002] With the widespread application of multi-sensor fusion technology in complex collaborative and game-theoretic fields such as autonomous driving, intelligent monitoring, drone swarming, industrial IoT, and situational awareness, how to efficiently and accurately integrate heterogeneous sensor data has become a research hotspot. However, due to limitations in sensor types, performance, and environmental conditions, target perception data often faces the problem of missing feature dimensions in practical applications. That is, some sensors cannot obtain complete feature information of the target (such as speed, attitude, size, scattering characteristics, radio frequency fingerprints, etc.), resulting in incomplete data or information bias, which in turn affects the accuracy of target feature association, information fusion, and individual identification.

[0003] In the paper "Research on Adaptive Clustering Algorithm for Missing Data Completion in Asynchronous Sensors" published in the *Journal of Sensor Technology* by Feng Dongyan et al., an adaptive clustering algorithm was proposed to address the need for missing data completion in asynchronous sensors. This algorithm combines a sliding window model, matrix factorization, and adaptive fruit fly optimization to optimize the clustering process and achieve adaptive clustering results for missing data completion. However, this method involves multiple algorithmic models, leading to a complex system structure. Furthermore, the fruit fly optimization algorithm requires iteration, and factors such as convergence condition settings affect the optimization time. In the paper "Time Series Missing Data Completion Method Based on Generalized Central Clustering" published in the *Journal of Jilin University (Natural Science Edition)* by Yu Yanpeng et al., a generalized central clustering method was proposed to fill missing values ​​in time series data and reduce errors. With the support of data spatial relationships, combined with information bottleneck calculations, the cluster radius was calculated and a fluctuation threshold was set to achieve high-accuracy time series missing data completion. However, in multi-sensor scenarios, the relationship between target features is relatively weak, and multiple sensors will perceive the target's feature information, making it difficult to effectively allocate incomplete sensing information to the target attribution in the initial stage. In the patent application "A Long-Term Tracking Method with Dual Threshold Control" (CN114332157B) filed by Beijing Institute of Technology, the MDNet verification network utilizes information from subsequent frames through online training to compensate for the lack of target information in the tracking algorithm. However, this method relies on supervised learning and requires a significant amount of prior information. In multi-sensor perception information fusion scenarios, the applicability of this method is limited because sensors cannot fully perceive all feature values ​​of the target, and the incomplete information from multiple targets adds to the problem. Furthermore, the existence of unknown targets further complicates the application. In the patent application "A Multi-Target Radiation Reconnaissance Method Based on Path Optimization" (CN114115237B) filed by the PLA Army Chemical Defense Academy, the reconnaissance targets are directly clustered using the K-means algorithm. However, performing unsupervised clustering directly without sufficient data leads to significant clustering errors. Summary of the Invention

[0004] In view of this, the present invention discloses an intelligent fusion and completion method for target information under conditions of missing sensing data. This invention can solve the problem of incomplete information fusion data caused by missing target feature information during multi-sensor sensing, complete and fuse information to form complete target information, and the fusion and completion process has a high degree of automation and has engineering application value.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for intelligent fusion and completion of target information under the condition of missing perception data includes the following steps:

[0007] Step 1: Collect target perception data from multiple sensors, including location information and multiple target feature information, preprocess the perception data, and summarize them to form a multi-sensor dataset;

[0008] Step 2: Divide the multi-sensor dataset into multiple subsets based on the target feature dimensions, retaining the two-dimensional location information and each feature dimension to achieve feature dimension separation and form each feature subset;

[0009] Step 3: Cluster each feature subset to obtain the clustering results of each feature subset, i.e., the centroids of clusters for each feature dimension, and record the corresponding number of clusters;

[0010] Step 4: Extract the location information from the clustering results of each feature subset, construct a location information dataset, and perform re-clustering based on the maximum number of clusters to obtain the centroids of the location clusters;

[0011] Step 5: Construct the target space based on the kernel coordinates of each location cluster, and fill in the corresponding location cluster centroids at the coordinate positions, and fill in the empty column list for each feature position; traverse all perceived target information in the multi-sensor dataset, match the nearest cluster result according to the two-dimensional location information of the target, and summarize the feature information into the corresponding feature dimension list in the target space to form a complete data summary; calculate the feature value by averaging the list corresponding to the features of each target to achieve the fusion and completion of target information.

[0012] Furthermore, the specific method of step 1 is as follows:

[0013] Based on N sensors, K feature dimensions of information are perceived from M targets. The first two feature dimensions are two-dimensional normalized coordinate positions, and the information of the m-th target perceived by the n-th sensor is... for:

[0014]

[0015] in, This represents the value of the k-th feature of the m-th target perceived by the n-th sensor. and All are two-dimensional normalized coordinate position features; for feature information that cannot be obtained, the corresponding feature dimension data will be used. Marked as zero;

[0016] By aggregating the data from N sensors, a multi-sensor dataset is formed. :

[0017]

[0018] Wherein, the amount of target information data sensed by each sensor is L, and L does not exceed M, and the target information sensed by the nth sensor for the lth target is... .

[0019] Furthermore, the specific method for step 2 is as follows:

[0020] The k-th feature, the first piece of information, and the second piece of information of the l-th target sensed by the n-th sensor are assembled to form the target feature information. :

[0021]

[0022] in, and ;

[0023] Data is aggregated based on target feature information to obtain the multi-sensor dataset of the k-th feature. :

[0024]

[0025] There are K-2 multi-sensor feature subsets.

[0026] Furthermore, the specific method for step 3 is as follows:

[0027] Step 3-1: Use the KMeans unsupervised clustering algorithm to cluster each feature subset. Cluster them one by one;

[0028] Step 3-2: For each feature subset Set the maximum possible number of clusters P, iterate through the range from 1 to P as the number of clusters, and perform a feature subset analysis. Clustering is performed to obtain the sum of squared intra-cluster errors for each cluster number; second-order differencing is applied to the sum of squared intra-cluster errors for each cluster number, and the cluster number with the largest difference value is selected. The number of clusters in this feature subset is used as the basis for unsupervised clustering; based on this number of clusters, unsupervised clustering is performed to obtain the clustering results. ,in The coordinates of the centroids of each cluster are 3.

[0029] Furthermore, step 4 is specifically implemented as follows:

[0030] Step 4-1: Obtain each feature subset dataset Maximum number of clusters obtained ,Right now The clustering results of all feature subsets are summarized, and the first two dimensions of the centroid coordinates of each cluster are taken to form the location dataset. The dimension of S is Where D is the centroid coordinate of each cluster. It is constructed by taking the first two position dimensions, and has a dimension of 2;

[0031] Step 4-2: Use the KMeans unsupervised clustering algorithm and set the number of clusters to 1. Clustering is performed on the location dataset S to obtain the clustering results. , where e represents the centroids of each type of location clustering result.

[0032] Furthermore, step 5 is specifically implemented as follows:

[0033] Step 5-1: Construct the target space based on the kernel coordinates of each location cluster. The target space is... A matrix of ×K, where each row corresponds to the target's feature information and each column corresponds to K target features; the centroids of each type of location clustering result are filled into the first and second columns of the corresponding rows in order, representing the target's location features, and the other columns are placed in an empty list;

[0034] Step 5-2: Traverse all perceived target information in the multi-sensor dataset, match the centroid of the nearest location clustering result based on the two-dimensional location information of the target; summarize the perceived target feature information into the corresponding feature dimension list in the target space to form a complete data summary;

[0035] Step 5-3: Calculate the feature values ​​by taking the mean of the list of features corresponding to each target, thereby achieving the fusion and completion of target information.

[0036] The beneficial effects of this invention are as follows:

[0037] 1) Addressing the issue of missing dimensions to improve data integrity. Traditional methods often suffer from information loss or incorrect associations due to missing feature dimensions when processing multi-sensor data. This invention avoids fusion bias caused by missing dimensions by segmenting, secondary clustering, and assembly.

[0038] 2) Dynamic clustering mechanism adapts to complex scenarios. Traditional clustering methods typically rely on a fixed number of clusters, making it difficult to adapt to changes in data distribution under different scenarios. This invention dynamically adjusts the number of clusters based on the maximum number of clusters, which can flexibly cope with differences in target distribution under complex environments and improve the robustness of clustering results.

[0039] 3) Hierarchical clustering mechanism increases data utilization. In cases where missing data significantly impacts estimation, this invention aggregates and clusters data according to feature dimensions, and then performs secondary clustering on distance information in the clustering results. This avoids clustering errors caused by missing data, leading to a hierarchical clustering mechanism that effectively improves data utilization.

[0040] 4) Intelligent completion of missing features for efficient fusion. By calculating the mean, missing feature information is summarized into the most recent clustering results. This invention achieves intelligent completion of missing features, while avoiding the subjective errors of traditional interpolation zero-padding methods, ensuring the objectivity and reliability of the fusion results. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0043] A method for intelligent fusion and completion of target information under conditions of missing perception data is proposed. This method involves multi-sensor data acquisition and preprocessing, dimensional segmentation and subset construction, clustering of feature subsets for each feature dimension, constructing a location dataset based on the location information of each feature clustering result, and further clustering based on the maximum number of clusters, as well as feature information summarization and mean completion. This method utilizes target feature information hierarchically, estimates the number of targets and location features through secondary clustering, and then summarizes the sensor perception information of the target based on the location features and performs mean processing to obtain the completed and fused target result. This method can solve the problem of target information fusion and completion under conditions of missing target perception information.

[0044] like Figure 1 As shown, the method specifically includes the following steps:

[0045] Step 1: Multi-sensor data acquisition and preprocessing. Acquire sensor data from multiple sensors to obtain the target's perception data, including location information and multiple target feature information. For features that cannot be acquired, mark the corresponding feature dimension data as zero. Summarize the data to form a multi-sensor dataset.

[0046] Multiple sensors are used to acquire target perception data, including location information and multiple target feature information. Based on N=10 sensors, K=5 feature dimensions of M=5 targets are perceived, where the first two feature information are two-dimensional normalized coordinate positions. Therefore, the m-th target information perceived by the n-th sensor... for:

[0047] ,

[0048] in, This represents the value of the k-th feature of the m-th target perceived by the n-th sensor. and All are two-dimensional normalized coordinate position features. For feature information that cannot be obtained, the corresponding feature dimension data is used. Marked as zero.

[0049] By aggregating the data from N sensors, a multi-sensor dataset is formed. There are instances where the sensors fail to detect the target, resulting in no corresponding target information data. Therefore, the amount of target information data sensed by each sensor is L=5, and at most M. Thus, the target information sensed by the nth sensor for the lth target is... Therefore, multi-sensor datasets It can be represented as:

[0050] ,

[0051] Step 2: Dimensional Segmentation and Subset Construction. The original data is divided into multiple subsets based on the target feature dimensions, retaining the two-dimensional location information and each feature dimension to achieve feature dimension separation and form each feature subset.

[0052] The k-th feature, the first piece of information, and the second piece of information of the l-th target perceived by the n-th sensor are assembled to form target feature information. ,in and .

[0053] ,

[0054] Data is aggregated based on target feature information to obtain the multi-sensor dataset of the k-th feature. That is, there are a total of K-2=3 multi-sensor feature subsets.

[0055] ,

[0056] Step 3: Clustering of feature subsets for each feature dimension. Cluster each feature subset, iterating through the possible number of clusters. Based on the maximum second-order difference of the sum of squared intra-cluster errors for each cluster, obtain the number of clusters and then cluster each feature subset to obtain the clustering results for each feature subset, i.e., the centroids of clusters for each feature dimension, and record the corresponding number of clusters.

[0057] Step 3-1: Use the KMeans unsupervised clustering algorithm to cluster each feature subset. Cluster them one by one, for a total of K-2=3, namely , and .

[0058] Step 3-2: For each feature subset Set the maximum possible number of clusters P, iterate through the range from 1 to P as the number of clusters, and perform a feature subset analysis. Clustering is performed to obtain the sum of squared intra-cluster errors for each cluster. Second-order differencing is then applied to the sum of squared intra-cluster errors for each cluster, and the cluster with the largest difference value is selected. The number of clusters in this feature subset is used as the basis for unsupervised clustering to obtain the clustering results. , ,in The coordinates of the centroids of each cluster are 3.

[0059] For feature subsets The number of clusters is 4, and the clustering result is obtained. The values ​​are: [[0.66542593,0.50719811, 0.59866149], [0.20948443, 0.15592717, 0.47885669], [0.12885785,0.36238831, 0.23094887], [0.39389786, 0.36317684, 0.8822696]].

[0060] For feature subsets The number of clusters is 3, and the clustering result is obtained. The values ​​are: [[0.21127202,0.18528502, 0.35572624], [0.65890774, 0.41019448, 0.63776507], [0.09646931,0.41458087, 0.2118899]].

[0061] For feature subsets The number of clusters is 5, and the clustering result is obtained. The values ​​are: [[0.23253867,0.19183833, 0.4147403], [0.52102451, 0.49538409, 0.70850217], [0.09731895,0.33368539, 0.28088706], [0.70309127, 0.25707588, 0.63547396], [0.34177299,0.40591277, 0.80946265]].

[0062] Step 4: Construct a location dataset based on the location information of the clustering results for each feature, and then cluster again based on the maximum number of clusters. Extract the location information from the clustering results of each feature subset, construct a location information dataset, and perform clustering based on the maximum number of clusters to obtain the centroids of the location clusters.

[0063] Step 4-1: Obtain each feature subset dataset Maximum number of clusters obtained ,Right now Therefore =5. Summarize the clustering results of all feature subsets, and take the first two dimensions of the centroid coordinates of each cluster to form the location dataset. Its dimensions are Where D is the centroid coordinate of each cluster. It is constructed by taking the first two position dimensions, with a dimension of 2.

[0064] Step 4-2: Use the KMeans unsupervised clustering algorithm and set the number of clusters to 1. Clustering the location dataset S yields clustering result E. , where e represents the centroids of each cluster in the location clustering result. The clustering result E is:

[0065] [[0.21776504, 0.17768351], [0.59322522, 0.5012911], [0.1075487, 0.37021819], [0.68099951, 0.33363518], [0.36783542, 0.38454481]]

[0066] Step 5: Feature Information Summarization and Mean Completion. Construct a target space based on the kernel coordinates of each location cluster, filling in the corresponding location cluster centroids at the coordinate positions, and filling empty columns for each feature location. Traverse all perceived target information in the multi-sensor dataset, matching the nearest cluster result based on the target's two-dimensional location information. Summarize the feature information into the corresponding feature dimension list in the target space, forming a complete data summary. Calculate the feature values ​​for each target's feature list using the mean, achieving target information fusion and completion.

[0067] Step 5-1: Construct the target space based on the kernel coordinates of each cluster location, and fill in the corresponding cluster centroids at the coordinate locations, while filling in empty columns for each feature location. The target space is... A matrix of ×K, each row corresponds to the feature information of the target, and each column corresponds to the centroid of each clustering result. Each column corresponds to K target features. The centroids of each clustering result are filled into the first and second columns of the corresponding rows in order to represent the location features of the target. The other columns are placed in an empty list.

[0068] Step 5-2: Traverse all perceived target information in the multi-sensor dataset and match the centroid of the nearest clustering result based on the two-dimensional position information of the target. The perceived target feature information is summarized into the corresponding feature dimension list in the target space, forming a complete data summary as follows: [[0.21776503876980116, 0.17768350616171308,[0, 0.4868499288443328, 0, 0, 0.4708634445583699, 0, 0, 0, 0],[0.32934361011979363, 0, 0, 0, 0, 0.3313142550925384, 0.3312595232844876,0.43098757967018786, 0], [0, 0, 0.47525713421341664, 0.43301362987174985, 0,0, 0, 0, 0.3359501474378612]], [0.5932252195002019, 0.501291098291453, [0, 0,0, 0.6491267531650358, 0, 0.6160289450649772, 0.6326651204733129, 0,0.6267879752906718, 0, 0], [0.5926755209162776, 0.6165372137332404,0.544601330661887, 0, 0.565568210633918, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0, 0, 0.6618994324128419, 0, 0.6754941853905096, 0.7416658500421928]],[0.10754870313531253, 0.3702181911089827, [0, 0, 0.2915951565798394, 0, 0,0.1703025849014362, 0, 0, 0, 0, 0], [0.26867060342378013, 0, 0,0.19883873741850916, 0, 0, 0.21134887323131657, 0.16870140193604743, 0, 0,0], [0, 0.28753832780723076, 0, 0, 0.22029187269424688, 0, 0, 0,0.24467680657132784, 0.27554074456151706, 0.37638756118750233]],[0.6809995085718705, 0.33363517783864677, [0, 0.6000488512958105, 0,0.5575225819853915, 0.5084502158243499, 0, 0, 0, 0, 0], [0.5687661151910591,0, 0.6093283098763093, 0, 0, 0.6938177474097458, 0, 0.6496644641110986,0.6197433022820952, 0.633661010001673], [0, 0, 0, 0, 0, 0,0.6354739609816293, 0, 0, 0]], [0.36783542477810677, 0.38454480731201685, [0,0.8707778706054602, 0, 0.8998926575508095, 0, 0, 0.8761382707358422, 0, 0],[0, 0, 0, 0, 0, 0.7706008472283485, 0, 0, 0.7882167600475474],[0.7352615941481594, 0, 0.8089032714665304, 0, 0.884223098221551, 0, 0,0.7549492097159609, 0]]].

[0069] Step 5-3: Calculate the feature values ​​from the list of features corresponding to each target using the mean, thus achieving the fusion and completion of target information, resulting in [[0.21776503876980116, 0.17768350616171308, 0.4788566867013514, 0.35572624204175185, 0.4147403038410092], [0.5932252195002019, 0.501291098291453, 0.6311521984984994, 0.5798455689863308, 0.6930198226151815], [0.10754870313531253,0.3702181911089827, 0.2309488707406378, 0.21188990400241334,0.28088706256436496], [0.6809995085718705, 0.33363517783864677,0.5553405497018505, 0.6291634914786635, 0.6354739609816293],[0.36783542477810677, 0.38454480731201685, 0.8822695996307041, 0.779408803637948, 0.7958342933880505]]

[0070] Traditional methods directly interpolate and pad missing data with zeros before performing KMeans unsupervised clustering, resulting in a target number of 3. The centroid coordinates of each class are: [[4.56892912e-01, 3.66089863e-01, -1.11022302e-16, 4.96182271e-01, -8.32667268e-17], [4.95385012e-01, 4.05865635e-01, 6.22057521e-01, -5.55111512e-17, -5.55111512e-17], [2.98416089e-01, 3.58858087e-01, 1.00177991e-02, -5.55111512e-17, 5.02736872e-01]].

[0071] The actual number of targets is 5, and the feature information of each target is [[0.1, 0.4, 0.1, 0.7, 0.2], [0.2, 0.2, 0.3, 0.3, 0.4], [0.6, 0.6, 0.5, 0.3, 0.6], [0.8, 0.3, 0.7, 0.7, 0.6], [0.4, 0.4, 0.9, 0.7, 0.8]].

[0072] By comparing our proposed method with traditional interpolation and zero-padding clustering methods, it is evident that our proposed method correctly analyzes the number of targets. Furthermore, we match the centroid coordinates obtained by our proposed method and the traditional interpolation and zero-padding method with the feature information of each real target. The matching principle is based on minimizing the L2 norm of the centroid coordinates and feature information. The evaluation index for quantization bias is the average of the absolute value of the estimated feature value minus the real target feature value, divided by the real target feature value. The bias of our proposed method is 0.2157567096308116, while the bias of the traditional method is 1.0022980903286258. This shows that the quantization bias of our proposed method is significantly smaller than that of the traditional method, indicating that our proposed method has a better fusion and completion effect when perceptual data is missing.

Claims

1. A method for intelligent fusion and completion of target information under conditions of missing perception data, characterized in that, Includes the following steps: Step 1: Collect target perception data from multiple sensors, including location information and multiple target feature information, preprocess the perception data, and summarize them to form a multi-sensor dataset; Step 2: Divide the multi-sensor dataset into multiple subsets based on the target feature dimensions, retaining the two-dimensional location information and each feature dimension to achieve feature dimension separation and form each feature subset; Step 3: Cluster each feature subset to obtain the clustering results of each feature subset, i.e., the centroids of clusters for each feature dimension, and record the corresponding number of clusters; Step 4: Extract the location information from the clustering results of each feature subset, construct a location information dataset, and perform re-clustering based on the maximum number of clusters to obtain the centroids of the location clusters; Step 5: Construct the target space based on the kernel coordinates of each location cluster, and fill in the corresponding location cluster centroids at the coordinate positions, and fill in the empty column list for each feature position; traverse all perceived target information in the multi-sensor dataset, match the nearest cluster result according to the two-dimensional location information of the target, and summarize the feature information into the corresponding feature dimension list in the target space to form a complete data summary; calculate the feature value by averaging the list corresponding to the features of each target to achieve the fusion and completion of target information.

2. The intelligent fusion and completion method for target information under missing perception data as described in claim 1, characterized in that, The specific method for step 1 is as follows: Based on N sensors, K feature dimensions of information are perceived from M targets. The first two feature dimensions are two-dimensional normalized coordinate positions, and the information of the m-th target perceived by the n-th sensor is... for: in, This represents the value of the k-th feature of the m-th target perceived by the n-th sensor. and All are two-dimensional normalized coordinate position features; for feature information that cannot be obtained, the corresponding feature dimension data will be used. Marked as zero; By aggregating the data from N sensors, a multi-sensor dataset is formed. : Wherein, the amount of target information data sensed by each sensor is L, and L does not exceed M, and the target information sensed by the nth sensor for the lth target is... .

3. The intelligent fusion and completion method for target information under missing perception data according to claim 2, characterized in that, The specific method for step 2 is as follows: The k-th feature, the first piece of information, and the second piece of information of the l-th target sensed by the n-th sensor are assembled to form the target feature information. : in, and ; Data is aggregated based on target feature information to obtain the multi-sensor dataset of the k-th feature. : There are K-2 multi-sensor feature subsets.

4. The intelligent fusion and completion method for target information under missing perception data according to claim 3, characterized in that, The specific method for step 3 is as follows: Step 3-1: Use the KMeans unsupervised clustering algorithm to cluster each feature subset. Cluster them one by one; Step 3-2: For each feature subset Set the maximum possible number of clusters P, iterate through the range from 1 to P as the number of clusters, and perform a feature subset analysis. Clustering is performed to obtain the sum of squared intra-cluster errors for each cluster number; second-order differencing is applied to the sum of squared intra-cluster errors for each cluster number, and the cluster number with the largest difference value is selected. The number of clusters in this feature subset; Unsupervised clustering is performed based on this number of clusters to obtain the clustering results. ,in The coordinates of the centroids of each cluster are 3.

5. The intelligent fusion and completion method for target information under missing perception data according to claim 4, characterized in that, The specific method for step 4 is as follows: Step 4-1: Obtain each feature subset dataset Maximum number of clusters obtained ,Right now The clustering results of all feature subsets are summarized, and the first two dimensions of the centroid coordinates of each cluster are taken to form the location dataset. The dimension of S is Where D is the centroid coordinate of each cluster. It is constructed by taking the first two position dimensions, and the dimension is 2; Step 4-2: Use the KMeans unsupervised clustering algorithm and set the number of clusters to 1. Clustering is performed on the location dataset S to obtain the clustering results. , where e represents the centroids of each type of location clustering result.

6. The intelligent fusion and completion method for target information under missing perception data according to claim 5, characterized in that, The specific method for step 5 is as follows: Step 5-1: Construct the target space based on the kernel coordinates of each location cluster. The target space is... A matrix of size ×K, where each row corresponds to the feature information of the target, and each column corresponds to K target features; Fill the centroids of each type of location clustering result into the first and second columns of the corresponding rows in order to represent the location features of the target, and put the other columns into an empty list; Step 5-2: Traverse all perceived target information in the multi-sensor dataset, match the centroid of the nearest location clustering result based on the two-dimensional location information of the target; summarize the perceived target feature information into the corresponding feature dimension list in the target space to form a complete data summary; Step 5-3: Calculate the feature values ​​by taking the mean of the list of features corresponding to each target, thereby achieving the fusion and completion of target information.

Images