A security unmanned aerial vehicle multi-target trajectory prediction method and system

Abstract

The present application relates to the technical field of unmanned aerial vehicle monitoring, and particularly relates to a multi-target trajectory prediction method and system for security unmanned aerial vehicles, which comprises the following steps: obtaining all trajectory segments in a preset time period in the video stream of the unmanned aerial vehicle; calculating the correlation similarity of each two trajectory segments that do not overlap on the time axis; constructing a similarity matrix based on the correlation similarity, using an improved inter-cluster distance algorithm to perform hierarchical clustering on the similarity matrix, and calculating the anomaly coefficient of the clustering cluster obtained by each merging; when the anomaly coefficient is greater than or equal to a preset threshold, the merging is stopped, and the final cluster is marked; the trajectory segments in the same final cluster are marked as the trajectory segments of the same target, and are spliced in time sequence; the spliced trajectory segments are input into a trained LSTM prediction model to predict the future trajectory of the target. The present application can improve the phenomenon of erroneous splicing of trajectory segments, improve the accuracy of trajectory prediction, and improve the estimation accuracy of the future position of the target by the unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) monitoring technology. In particular, it relates to a multi-target trajectory prediction method and system for security UAVs. Background Technology

[0002] In low-altitude security scenarios using drones, stable tracking and trajectory prediction of multiple targets presupposes obtaining complete and continuous motion trajectories for each target. However, due to the limited field of view of airborne cameras and the complexity of urban environments, targets are easily obscured by buildings, trees, utility poles, or other moving objects, and may also temporarily "disappear" due to sudden changes in lighting, motion blur, or a sharp drop in detection confidence. These intermittent events prevent the same physical target from being continuously observed in the video stream, causing its motion trajectory to be "cut" into multiple short fragments, increasing the difficulty of tracking and easily leading to mismatches of motion trajectories between multiple targets.

[0003] In recent years, deep learning-based multi-target tracking algorithms have made significant progress. These algorithms combine target detection and tracking algorithms, effectively handling problems such as small targets, complex target motion, and occlusion, while achieving a good balance between accuracy and real-time performance.

[0004] However, existing deep learning-based multi-target tracking algorithms still have some shortcomings: traditional feature extraction networks may struggle to extract sufficiently discriminative features, leading to mismatches; traditional IoU matching algorithms rely excessively on the geometric overlap between the detection box and the tracked target, and may fail to accurately reflect the matching degree for targets with complex or irregular shapes, or when the target is partially occluded or its shape is significantly deformed, resulting in incorrect associations or matching failures; traditional algorithms use linear Kalman filtering by default for target motion prediction, which is difficult to accurately describe the nonlinear motion of the target from the UAV's perspective, leading to tracking failures.

[0005] Therefore, there is a need in this field for a multi-target trajectory prediction method and system for security drones to solve the problem that when the above-mentioned tracking algorithm encounters trajectory gaps due to continuous observation interruptions, it is easy to cause errors in the splicing of individual target trajectories, thus affecting the accuracy of target trajectory prediction. Summary of the Invention

[0006] To address the technical problem that the above-mentioned tracking algorithm is prone to errors in splicing individual target trajectories when encountering trajectory gaps during continuous observation interruptions, thus affecting the accuracy of target trajectory prediction, this invention provides solutions in the following aspects.

[0007] In the first aspect, a multi-target trajectory prediction method for security drones includes:

[0008] Obtain the trajectory segments of all tracking numbers within a preset time period in the drone video stream. Each trajectory segment contains the position coordinates of the corresponding tracking number at each acquisition time.

[0009] Calculate the association similarity between any two trajectory segments that do not overlap on the time axis, and set the association similarity between any two trajectory segments that overlap on the time axis to 0;

[0010] A similarity matrix is ​​constructed based on association similarity. An improved inter-cluster distance algorithm is used to perform bottom-up hierarchical clustering on the similarity matrix. The anomaly coefficient of each merged cluster is calculated. If the anomaly coefficient is greater than or equal to a preset threshold, the cluster will no longer participate in the merge and will be marked as the final cluster.

[0011] Trajectory segments within the same final cluster are labeled as trajectory segments of the same target and spliced ​​together in chronological order. The spliced ​​trajectory segments are then input into a trained LSTM prediction model to predict the future trajectory of the target.

[0012] The improved inter-cluster distance algorithm includes: pairing each trajectory segment in one cluster with each trajectory segment in another cluster in turn, reading the correlation similarity between each pair of paired trajectory segments, and selecting the correlation similarity with the smallest value; reading the correlation similarity between every two trajectory segments in a single cluster, and calculating the average of all correlation similarities; and calculating the inter-cluster distance between the two clusters based on the correlation similarity with the smallest value and the average correlation similarity between the two clusters.

[0013] Preferably, the method for calculating the association similarity includes: calculating the motion feature similarity between two trajectory segments; calculating the shortest path distance between the two trajectory segments using the DTW matching algorithm, and performing negative correlation normalization on the shortest path distance; calculating the standard deviation of the time difference of all matching point pairs based on the matching point pairs obtained by the DTW matching algorithm; and multiplying the motion feature similarity, the normalized shortest path distance, and the standard deviation of the time difference of all matching point pairs to obtain the association similarity between the two trajectory segments.

[0014] By introducing a multi-dimensional association similarity calculation method based on motion feature similarity, DTW path distance and temporal rhythm consistency, this method effectively identifies broken trajectory segments due to occlusion, lighting changes or detection failures, significantly improving the matching accuracy of the same target trajectory segment and solving the problems of misjudgment and splicing errors at trajectory gaps in traditional methods.

[0015] Preferably, the method for calculating the motion feature similarity includes: for a single trajectory segment, calculating the mean and standard deviation of the motion speed between adjacent sampling points within the trajectory segment; calculating the ratio of the distance between the gaps between two trajectory segments to the time difference; calculating the absolute value of the difference between the mean of the motion speed of the two trajectory segments and the aforementioned ratio, and selecting the absolute value of the smallest difference; calculating the ratio of the standard deviation of the motion speed of the two trajectory segments, multiplying the absolute value of the difference between 1 and the ratio by the absolute value of the aforementioned smallest difference to obtain a product, and performing negative correlation normalization processing on the product to obtain the motion feature similarity.

[0016] Preferably, the method for calculating the distance and time difference of the gap between the two trajectory segments includes: distinguishing the two trajectory segments from one to the other according to the time sequence, obtaining the position coordinates of the end trajectory point of the previous trajectory segment and the position coordinates of the beginning trajectory point of the subsequent trajectory segment, and calculating the distance of the gap based on the two position coordinates; obtaining the acquisition time corresponding to the end trajectory point of the previous trajectory segment and the acquisition time corresponding to the beginning trajectory point of the subsequent trajectory segment, and calculating the time difference of the gap based on the two acquisition times.

[0017] Preferably, the step of calculating the inter-cluster distance between two clusters based on the minimum numerical correlation similarity and the average correlation similarity of the two clusters includes: performing negative correlation normalization on the minimum numerical correlation similarity; calculating the difference between the average correlation similarity of the two clusters, and calculating the ratio of the absolute value of the difference to the largest average value among the average values ​​of the two clusters; multiplying the aforementioned ratio by the normalized minimum numerical correlation similarity to obtain the inter-cluster distance between the two clusters.

[0018] By employing an improved inter-cluster distance algorithm for hierarchical clustering, combined with a minimum similarity penalty mechanism between trajectory segments and intra-cluster similarity consistency constraints, the erroneous merging of time-overlapping trajectory segments is effectively avoided, improving the adaptability of clustering results to complex scenarios and ensuring that each cluster more accurately corresponds to a single physical target.

[0019] Preferably, the method for calculating the anomaly coefficient includes: calculating the mean inter-cluster distance between the merged cluster and all other clusters, and performing positive correlation normalization on the mean inter-cluster distance; calculating the standard deviation of the correlation similarity of all pairwise trajectory segment combinations in the merged cluster; and multiplying the normalized mean inter-cluster distance by the standard deviation of the correlation similarity to obtain the anomaly coefficient of the merged cluster.

[0020] Preferably, the same target contains at least one trajectory segment corresponding to a tracking number.

[0021] Secondly, a multi-target trajectory prediction system for security drones includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned multi-target trajectory prediction method for security drones is implemented.

[0022] The present invention has the following effects:

[0023] 1. This invention completely eliminates the possibility of overlapping time trajectory segments being mistakenly identified as the same target by constraining that "association similarity is only calculated when there is no overlap in time axis". This fundamentally suppresses the phenomenon of trajectory mis-splitting caused by time conflicts in traditional methods and significantly improves the consistency of trajectory identity.

[0024] 2. This invention calculates the anomaly coefficient in real time during the clustering process and sets a hard threshold termination condition, which can stop merging in time at the key point of "the difference between clusters becomes smaller and the difference within clusters becomes larger", avoiding the forced inclusion of trajectory segments of different targets into the same cluster, ensuring that the trajectory segments within the final cluster all belong to the same physical target, and improving the quality of input data for subsequent trajectory prediction.

[0025] 3. Compared to directly using fragmented trajectory segments as prediction input, this invention first splices the trajectory segments within the same final cluster into the longest possible continuous trajectory according to time sequence, and then feeds it into the LSTM model. This allows the network to utilize longer historical time sequence information, significantly reducing prediction errors caused by data truncation, improving the estimation accuracy of the UAV for the future position of the target, overcoming the shortcomings of traditional linear models in describing nonlinear motion, and enhancing the UAV's ability to predict target behavior in security scenarios. Attached Figure Description

[0026] Figure 1 This is a flowchart of steps S1-S4 in a multi-target trajectory prediction method for security drones according to an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] Reference Figure 1 A multi-target trajectory prediction method for security drones includes steps S1-S4, as detailed below:

[0030] S1: Obtain the trajectory segments of all tracking numbers within a preset time period in the drone video stream. Each trajectory segment contains the position coordinates of the corresponding tracking number at each acquisition time.

[0031] The drone continuously acquires visible light video streams via its onboard gimbal camera and feeds them into the YOLO (You Only LookOnce) lightweight detection network. The YOLO network outputs the target's bounding box, which is then fed into SORT (Simple Online and Realtime Tracker). SORT uses IoU (Intersection over Union) matching and Kalman filtering to predict the position and assign a tracking number to continuously visible targets. When a target is occluded or moves out of the frame, SORT determines that the target is "lost" and immediately terminates the tracking number. When the target reappears, even if it looks exactly the same, it will be assigned a new tracking number.

[0032] Therefore, the same physical target may generate many disconnected sub-trajectories within a preset time period (such as 10 minutes) of video window, and each sub-trajectory is called a trajectory segment.

[0033] Obtain the trajectory segments with all tracking numbers within a preset time period, where the first... Each trajectory segment is denoted as ,in To track the number at time The location coordinates.

[0034] Because a single target's trajectory can form different tracking numbers due to occlusion, loss, or other reasons during its movement, obtaining all the tracking numbered trajectory segments allows for analysis of all trajectory segments. By stitching together trajectory segments belonging to the same target, the target's movement trajectory can be predicted more accurately.

[0035] S2: Calculate the association similarity between any two trajectory segments that do not overlap on the time axis, and set the association similarity between any two trajectory segments that overlap on the time axis to 0.

[0036] When calculating the association similarity between any two trajectory segments, it is important to note that it is not necessary to calculate the association similarity between any two trajectory segments in all trajectory segments. Instead, it is necessary to calculate the association similarity between any two trajectory segments that do not overlap on the time axis. If two trajectory segments overlap on the time axis, then they must not belong to the same target.

[0037] When calculating the correlation similarity between two trajectory segments that do not overlap on two time axes, we must first consider whether they conform to the movement characteristics of the same target. Since the appearance features of the target captured by the UAV are not sufficient to determine whether two trajectory segments belong to the same target, we use the movement speed features of the two trajectory segments, as well as the gap length and time difference between the two trajectory segments, to quantify the similarity of the movement characteristics of the two target trajectory segments.

[0038] In one embodiment, the method for calculating the similarity of movement features between two trajectory segments includes: for a single trajectory segment, calculating the mean and standard deviation of the movement speed between adjacent sampling points within the trajectory segment; calculating the ratio of the distance between the gaps between the two trajectory segments to the time difference; calculating the absolute value of the difference between the mean of the movement speed of the two trajectory segments and the aforementioned ratio, and selecting the absolute value of the smallest difference; calculating the ratio of the standard deviation of the movement speed of the two trajectory segments, multiplying the absolute value of the difference between 1 and the ratio by the absolute value of the aforementioned smallest difference to obtain a product, and performing negative correlation normalization on the product to obtain the similarity of movement features.

[0039] The method for calculating the distance and time difference of the gap between two trajectory segments includes: distinguishing the two trajectory segments from one to the other according to the time sequence, obtaining the position coordinates of the end trajectory point of the previous trajectory segment and the position coordinates of the beginning trajectory point of the subsequent trajectory segment, and calculating the distance of the gap based on the two position coordinates; obtaining the acquisition time corresponding to the end trajectory point of the previous trajectory segment and the acquisition time corresponding to the beginning trajectory point of the subsequent trajectory segment, and calculating the time difference of the gap based on the two acquisition times.

[0040] The formula for calculating the similarity of moving features is as follows:

[0041]

[0042] In the formula, Represents trajectory segment With trajectory segment Similarity of movement features between them; Represents trajectory segment The average moving speed between every two adjacent sampling points. Represents trajectory segment The standard deviation of the moving speed between any two adjacent sampling points is calculated by taking the position coordinates of the sampling points, calculating the displacement between two adjacent sampling points based on the sampling time, and then subtracting the displacement from the time difference to obtain the moving speed between the two adjacent sampling points. This process is applied to the trajectory segment. The average moving speed between every two adjacent sampling points is calculated. and standard deviation ; Represents trajectory segment The average moving speed between every two adjacent sampling points. Represents trajectory segment The standard deviation of the moving speed between every two adjacent sampling points.

[0043] Represents trajectory segment The position coordinates of the end trajectory point and the trajectory segment The distance between the position coordinates of the first trajectory point; Represents trajectory segment The terminal trajectory point and trajectory segment The difference between the first trajectory point and the corresponding acquisition time; Represents trajectory segment With trajectory segment The ratio of the distance between the gaps to the time difference represents the speed of the target within the gap caused when the UAV cannot detect the target corresponding to the tracking number. Represented by natural numbers An exponential function with base 1.

[0044] After calculating the average moving speed of the two trajectory segments and the difference between the moving speed within the gap between the two trajectory segments, the absolute value of the smaller difference is used to quantify the similarity of the moving features. The smaller the difference, the smaller the difference between the moving speed within the gap between the two trajectory segments and the moving speed within the trajectory segments, and the higher the similarity of the moving features should be.

[0045] The ratio of the standard deviations of the moving speeds of two trajectory segments reflects the similarity of their moving speeds. The closer the ratio is to 1, the higher the similarity of the moving speeds of the two trajectory segments, and the higher the similarity of their moving characteristics. Conversely, the further the ratio is from 1, the greater the difference in the moving speeds of the two trajectory segments, and the lower the similarity of their moving characteristics.

[0046] By quantifying the speed difference between trajectory segments and trajectory gaps, and the speed consistency between trajectory segments, the similarity of movement features between two trajectory segments can be reflected. The smaller the speed difference and the higher the speed consistency, the greater the similarity of movement features. This helps to identify trajectory segments with high similarity of movement features as the movement trajectory of the same target, avoid misjudging unrelated trajectories as the same target, and improve the reliability of trajectory stitching.

[0047] After quantifying the similarity of the movement features of two trajectory segments, it is necessary to consider the movement routes of the two trajectory segments to comprehensively quantify the correlation similarity between the two trajectory segments. Here, the similarity of the movement routes of the two trajectory segments is calculated by the DTW (Dynamic Time Warping) matching algorithm.

[0048] The method for calculating the association similarity between two trajectory segments that do not overlap on the time axis includes: calculating the motion feature similarity between the two trajectory segments; calculating the shortest path distance between the two trajectory segments using the DTW matching algorithm, and performing negative correlation normalization on the shortest path distance; calculating the standard deviation of the time difference of all matching point pairs based on the matching point pairs obtained by the DTW matching algorithm; and multiplying the motion feature similarity, the normalized shortest path distance, and the standard deviation of the time difference of all matching point pairs to obtain the association similarity between the two trajectory segments.

[0049] The formula for calculating association similarity is as follows:

[0050]

[0051] In the formula, Represents trajectory segment With trajectory segment The correlation and similarity between them; Represents trajectory segment With trajectory segment Similarity of movement features between them; This represents the trajectory segment calculated using the DTW matching algorithm. With trajectory segment The shortest path distance between them; This represents the normalization function, which normalizes the shortest path distance. The shortest path distance is negatively correlated with the association similarity. The smaller the shortest path distance between two trajectory segments, the higher the association similarity. This represents the standard deviation of the time difference between all matching points after the matching results between two trajectory segment sequences are obtained using the DTW matching algorithm. The smaller the standard deviation, the closer the time differences are, indicating that although the two trajectory segments are broken, their internal rhythms are almost synchronized. The more similar the movement trends of the targets in the two trajectory segments, the greater the possibility that the two trajectory segments belong to the same target. The larger the standard deviation, the more it indicates that the rhythms do not match, and the two trajectory segments may belong to different targets.

[0052] By comprehensively considering the similarity of movement features, movement routes, and movement trends between two trajectory segments, the accuracy of determining whether trajectory segments belong to the same target is improved, which facilitates subsequent clustering and trajectory prediction.

[0053] After calculating the correlation similarity between two trajectory segments that do not overlap on the time axis based on the above analysis, the trajectory segments with high correlation similarity need to be spliced ​​together to obtain the complete trajectory segments of all targets within the preset time period. In this process, in order to avoid interference, the correlation similarity between any two target trajectory segments that overlap on the time axis is recorded as 0.

[0054] S3: Construct a similarity matrix based on association similarity, use the improved inter-cluster distance algorithm to perform bottom-up hierarchical clustering on the similarity matrix, and calculate the anomaly coefficient of each merged cluster. If the anomaly coefficient is greater than or equal to a preset threshold, it will no longer participate in the merge and will be marked as the final cluster.

[0055] Treating all trajectory segments as nodes, a similarity matrix is ​​constructed, where the matrix elements represent the correlation similarity between trajectory segments. A hierarchical clustering algorithm is used to perform bottom-up clustering on the similarity matrix, grouping the trajectory segments into groups, with each cluster representing a potential target.

[0056] The method for calculating the inter-cluster distance includes: pairing each trajectory segment in one cluster with each trajectory segment in another cluster in turn; reading the correlation similarity between each pair of paired trajectory segments; and selecting the correlation similarity with the smallest value for negative correlation normalization; reading the correlation similarity between every two trajectory segments in a single cluster and calculating the average of all correlation similarities; calculating the difference between the average values ​​corresponding to the two clusters, and calculating the ratio of the absolute value of the difference to the largest average value among the average values ​​corresponding to the two clusters; multiplying the aforementioned ratio by the correlation similarity with the smallest value after normalization to obtain the inter-cluster distance between the two clusters.

[0057] The formula for calculating inter-cluster distance is as follows:

[0058]

[0059] In the formula, Represents clusters With clusters The distance between them; Indicates let Traversing clusters All trajectory segments within, let Traversing clusters All trajectory segments within, for each pair Read the pre-calculated correlation similarity between them The minimum correlation similarity is selected. The smaller the minimum correlation similarity value, the greater the possibility that the two clusters contain trajectory segments that do not belong to the same target. The minimum correlation similarity is used for exponential penalty to ensure that as long as there is any pair of trajectory segments that are obviously dissimilar in the two clusters, the distance between the two clusters is increased to prevent the two clusters from merging. Represented by natural numbers An exponential function with base 0. Furthermore, by setting the similarity between two overlapping trajectory segments on the time axis to 0 beforehand, if any pair of trajectory segments between two clusters overlaps on the time axis, then... That is, it is 0, which causes the inter-cluster distance to be amplified exponentially, thus being equivalent to infinity, preventing them from being merged.

[0060] Represents clusters Within the cluster, the average similarity of all pairwise trajectory segment combinations reflects the average level of trajectory segment similarity in the cluster. Represents clusters Within, the average similarity of all pairwise trajectory segment combinations; if the average similarity of trajectory segments of two clusters differs greatly, then the distance between the two clusters is correspondingly greater, and the possibility of merging the two clusters is smaller.

[0061] By using the improved inter-cluster distance calculation method and hierarchical clustering algorithm to group trajectory segments, the merging of trajectory segments that overlap on the time axis is avoided, the possibility of merging unrelated trajectory segments is reduced, and each cluster corresponds more accurately to a single target, thereby improving the accuracy of subsequent predictions.

[0062] Since the target number within the preset time period cannot be determined, the number of clusters cannot be predetermined. Therefore, a clustering termination condition needs to be set. This involves calculating an anomaly coefficient for each merged cluster to characterize the probability that they belong to the same target, thereby determining whether clustering should be terminated.

[0063] The method for calculating the anomaly coefficient includes: calculating the mean distance between the merged cluster and all other clusters, and performing positive correlation normalization on the mean distance between the clusters; calculating the standard deviation of the correlation similarity of all pairwise trajectory segment combinations in the merged cluster; and multiplying the normalized mean distance between the clusters by the standard deviation of the correlation similarity to obtain the anomaly coefficient of the merged cluster.

[0064] The formula for calculating the anomaly coefficient is as follows:

[0065]

[0066] In the formula, Represents clusters The anomaly coefficient; Represents clusters The mean of the inter-cluster distances to all other clusters; the larger this mean, the stronger the cluster. The greater the distance between the cluster and other clusters, the stronger the cluster. The lower the similarity between the trajectory segments contained in a cluster and the trajectory segments in other clusters, the lower the probability that they belong to the same target, and they should not be merged further. Represents the normalization function; Represents clusters The standard deviation of the association similarity of all pairwise trajectory segment combinations in the cluster reflects the difference in the association similarity of all pairwise trajectory segment combinations in the cluster. The larger the difference, the more trajectory segments are contained in this cluster, and the greater the possibility of anomalies. It should not be merged further.

[0067] A threshold is pre-set for the anomaly coefficient of the clusters. In one embodiment, the threshold is 0.5. When the anomaly coefficient is greater than or equal to 0.5, the cluster will no longer participate in the merging operation and will be marked as the final cluster. The threshold of 0.5 is a hyperparameter and can be selected according to the actual situation.

[0068] By quantifying the anomaly degree of clusters, excessive merging of clusters can be prevented, thereby improving the robustness of clustering results and the integrity of target trajectories.

[0069] S4: Trajectory segments within the same final cluster are marked as trajectory segments of the same target and spliced ​​together in chronological order. The spliced ​​trajectory segments are then input into the trained LSTM prediction model to predict the future trajectory of the target.

[0070] All trajectory segments belonging to the same final cluster are marked as different trajectory segments of the same target within a preset time period. These trajectory segments are sorted and concatenated in chronological order. The concatenated trajectory segments are then input into a pre-trained LSTM (Long Short-Term Memory) prediction model to predict the future trajectory of the target. The mean squared error loss function is used during the training of the LSTM prediction model.

[0071] This application also discloses a multi-target trajectory prediction system for security drones. The system includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the multi-target trajectory prediction method for security drones according to the above embodiments of the present invention is implemented.

[0072] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0073] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A multi-target trajectory prediction method for security drones, characterized in that, include: Obtain the trajectory segments of all tracking numbers within a preset time period in the drone video stream. Each trajectory segment contains the position coordinates of the corresponding tracking number at each acquisition time. Calculate the association similarity between any two trajectory segments that do not overlap on the time axis, and set the association similarity between any two trajectory segments that overlap on the time axis to 0; A similarity matrix is ​​constructed based on association similarity. An improved inter-cluster distance algorithm is used to perform bottom-up hierarchical clustering on the similarity matrix. The anomaly coefficient of each merged cluster is calculated. If the anomaly coefficient is greater than or equal to a preset threshold, the cluster will no longer participate in the merge and will be marked as the final cluster. Trajectory segments within the same final cluster are labeled as trajectory segments of the same target and spliced ​​together in chronological order. The spliced ​​trajectory segments are then input into a trained LSTM prediction model to predict the future trajectory of the target. The improved inter-cluster distance algorithm includes: pairing each trajectory segment in one cluster with each trajectory segment in another cluster in turn; reading the correlation similarity between each pair of paired trajectory segments and selecting the correlation similarity with the smallest value; reading the correlation similarity between every two trajectory segments in a single cluster and calculating the average of all correlation similarities; performing negative correlation normalization on the correlation similarity with the smallest value; calculating the difference between the average correlation similarities of the two clusters and calculating the ratio of the absolute value of the difference to the largest average of the average values ​​of the two clusters; multiplying the aforementioned ratio by the normalized correlation similarity with the smallest value to obtain the inter-cluster distance between the two clusters.

2. The multi-target trajectory prediction method for security drones according to claim 1, characterized in that, The method for calculating the correlation similarity includes: calculating the motion feature similarity between two trajectory segments; calculating the shortest path distance between the two trajectory segments using the DTW matching algorithm, and performing negative correlation normalization on the shortest path distance; calculating the standard deviation of the time difference of all matching point pairs based on the matching point pairs obtained by the DTW matching algorithm; and multiplying the motion feature similarity, the normalized shortest path distance, and the standard deviation of the time difference of all matching point pairs to obtain the correlation similarity between the two trajectory segments.

3. The multi-target trajectory prediction method for security drones according to claim 2, characterized in that, The method for calculating the similarity of movement features includes: for a single trajectory segment, calculating the mean and standard deviation of the movement speed between adjacent sampling points within the trajectory segment; calculating the ratio of the distance between the gaps between two trajectory segments to the time difference; calculating the absolute value of the difference between the mean of the movement speed of the two trajectory segments and the aforementioned ratio, and selecting the absolute value of the smallest difference; calculating the ratio of the standard deviation of the movement speed of the two trajectory segments, multiplying the absolute value of the difference between 1 and the ratio by the absolute value of the aforementioned smallest difference to obtain a product, and performing negative correlation normalization on the product to obtain the similarity of movement features.

4. The multi-target trajectory prediction method for security drones according to claim 3, characterized in that, The method for calculating the distance and time difference of the gap between the two trajectory segments includes: distinguishing the front and back of the two trajectory segments according to the time sequence, obtaining the position coordinates of the end trajectory point of the front trajectory segment and the position coordinates of the beginning trajectory point of the back trajectory segment, and calculating the distance of the gap based on the two position coordinates; obtaining the acquisition time corresponding to the end trajectory point of the front trajectory segment and the acquisition time corresponding to the beginning trajectory point of the back trajectory segment, and calculating the time difference of the gap based on the two acquisition times.

5. The multi-target trajectory prediction method for security drones according to claim 1, characterized in that, The method for calculating the anomaly coefficient includes: calculating the mean inter-cluster distance between the merged cluster and all other clusters, and performing positive correlation normalization on the mean inter-cluster distance; calculating the standard deviation of the correlation similarity of all pairwise trajectory segment combinations in the merged cluster; and multiplying the normalized mean inter-cluster distance by the standard deviation of the correlation similarity to obtain the anomaly coefficient of the merged cluster.

6. The multi-target trajectory prediction method for security drones according to claim 1, characterized in that, The same target contains at least one trajectory segment corresponding to a tracking number.

7. A multi-target trajectory prediction system for security drones, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement the multi-target trajectory prediction method for security drones according to any one of claims 1-6.

Images