A Flight Trajectory Point Behavior Pattern Analysis Method Based on Deep Learning Image Methods

Abstract

This invention discloses a method for analyzing the behavior patterns of flight trajectory points based on deep learning image processing, comprising the following steps: Step 1, converting the latitude and longitude of the flight trajectory points into x and y coordinates on the image using the Universal Transversal Mercator method; Step 2, using the rol-traj-yolov5 algorithm to adaptively rotate bounding box detection to predict the position of a specific trajectory and identify its pattern; Step 3, analyzing the specific behavior patterns of the trajectory to determine the specific aircraft model that formed the trajectory, and the specific mission or flight route. This invention, by identifying the trajectory patterns of a specific region of the flight trajectory point, can analyze useful information such as the aircraft model and the currently executed mission or flight route.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method for analyzing flight trajectory point behavior patterns based on deep learning image methods. Background Technology

[0002] While my country's aerospace field is developing rapidly, the performance of drones, airplanes and other aircraft is constantly improving, which in turn accelerates the research on flight trajectories or spatiotemporal trajectories.

[0003] The study of spatiotemporal trajectories has benefited from the rapid development of spatial positioning technology and sensor networks. Spatial positioning sensors are widely used in aircraft, ships, automobiles, and handheld devices, generating and accumulating massive amounts of spatiotemporal trajectory data of moving targets.

[0004] In the military field, the analysis of flight path points is particularly important. While the massive amount of flight path point information may seem chaotic, it actually contains a wealth of information that can be analyzed in military contexts. In military missions, specific types of fighter jets typically exhibit unique flight paths, such as figure-eight, rectangular, and elliptical maneuvers. Each flight path may indicate a specific combat or exercise mission, thus, analyzing patterns in flight path points can yield more specific military information in certain military applications.

[0005] Currently, commonly used trajectory motion pattern recognition methods mainly rely on a combination of feature engineering and traditional machine learning algorithms. For example, global optimal solution theory algorithms are used to identify flight actions, and particle swarm optimization algorithms are combined to symbolize flight parameter features, thereby enabling the processing and analysis of symbols with lower dimensions, achieving intelligent extraction of flight action trajectory features, and thus performing classification and other analyses. Based on traditional genetic algorithms, flight parameter-related attributes are encoded, and genetic mutation and rough set theory are used to establish action recognition standards and extract data attributes. Finally, genetic algorithms are used to obtain the recognition features of flight action trajectories. A fuzzy support vector machine-based flight action recognition method combines traditional support vector machine methods with fuzzy logic theory, transforming them into multiple binary problems for solving and judging sample categories. Using the LS-SVM method can improve the computational efficiency of the model and effectively identify various flight action trajectories.

[0006] However, feature engineering and traditional machine learning algorithms rely on the accuracy of manually extracted features. In trajectory data processing, it is necessary to design and extract corresponding features, which is a rather cumbersome process. Moreover, when manually extracted features are not representative, such as relying only on shallow features like time, speed, and distance, the recognition accuracy will be limited. Furthermore, the quality of manually extracted features is difficult to guarantee, and the final detection result will greatly depend on the quality of manually extracted features.

[0007] Deep learning can effectively overcome the shortcomings of traditional machine learning, which relies heavily on manual feature extraction and has low recognition accuracy.

[0008] Deep learning algorithms are based on large datasets and use networks to automatically extract target features for training, which greatly simplifies the training process and effectively improves the accuracy of recognition. Therefore, using deep learning algorithms to detect and analyze trajectory points is bound to be a major trend in the future.

[0009] In the military field, current methods for pattern recognition of specific trajectories formed by trajectory points often involve manually extracting features and then combining them with traditional machine learning algorithms to classify the trajectories and detect their locations. However, this trajectory pattern recognition method, which relies on traditional machine learning, is dependent on manually extracted features and has insufficient accuracy, making it difficult to accurately identify trajectory points. For each mission, the amount of trajectory points generated is enormous; however, only a small fraction of these points are worth analyzing to obtain useful military information. This results in a large number of redundant and meaningless trajectory points. Summary of the Invention

[0010] The purpose of this invention is to provide a deep learning image method for recognizing behavioral patterns and detecting starting points of flight trajectory points using convolutional neural networks. This technology can analyze and process trajectory points formed by military aircraft and drones in flight. By using the Universal Transvers Mercator projection, the latitude and longitude of the flight trajectory points are converted into images. The generated trajectory point images are then processed, and a rotating box detection network of the rol-traj-yolov5 deep learning object detection algorithm is used to train specific trajectory patterns in the trajectory point images. Once training is complete, behavioral patterns of specific trajectories in the trajectory points can be recognized. By identifying the trajectory patterns of specific areas of the military flight trajectory points, the model of the fighter jet and the mission it is currently performing can be analyzed.

[0011] The purpose of this invention is to provide a method for analyzing flight trajectory point behavior patterns based on deep learning image methods, so as to solve the problems existing in the prior art.

[0012] To achieve the above objectives, this invention provides a method for analyzing flight trajectory point behavior patterns based on deep learning image methods, comprising the following steps:

[0013] Step 1: Convert the latitude and longitude of the flight trajectory points into x and y coordinates on the image using the Universal TransversMercator method;

[0014] Step 2: Use the rol-traj-yolov5 algorithm to adaptively rotate the bounding box detection algorithm to predict the position of a specific trajectory and perform trajectory pattern recognition;

[0015] Step 3: By analyzing the specific behavioral patterns of the trajectory, we can determine the specific aircraft model that formed the trajectory and the specific mission or flight route.

[0016] Furthermore, the latitude and longitude coordinates of the trajectory points are converted into x and y coordinates on the image using a general transverse Mercator transform algorithm and normalization processing. Then, OpenCV is used to plot the planar projection coordinates of the trajectory points on the image to generate the image.

[0017] Furthermore, the rol-traj-yolov5 algorithm in step two is a rotating target box algorithm based on the YOLOv5-DOTA-OBB algorithm. This algorithm adopts a one-stage algorithm model and modifies the yolov5x baseline model. The rol-traj-yolov5 algorithm consists of backbone, neck, and yolo_head.

[0018] Furthermore, the backbone is the main part of the rol-traj-yolov5 algorithm. The main body uses the CSPDarnet53 structure and the focus structure. The focus layer transforms the information on the wh plane to the channel dimension, and then extracts different features through convolution. This method can reduce the information loss caused by downsampling.

[0019] Furthermore, in the neck structure, SPP is replaced with SPPF. Both serve the same purpose, but the latter greatly reduces the computational load of the model while maintaining the same feature extraction effect, thus accelerating the model's inference process.

[0020] The SPPF structure employs a receptive field mechanism, converting the three pooling layers of different sizes in the SPP structure into a single 5x5 pooling layer. The SPPF structure obtains the same receptive field size as the SPP structure through the 5x5, 9x9, and 13x13 pooling layers by connecting three 5x5 pooling layers in series, and reduces the pooling layer size, thus reducing the computational load.

[0021] The neck structure also includes New CSP-PAN, which incorporates a CSP network residual module into the feature fusion module PAN structure to prevent redundant information when calculating feature gradients. Furthermore, CBAM attention mechanisms are added to the upsampling of the PAN, inferring the image's attention in both channel and spatial independent dimensions. The attention map is then multiplied by the input feature map, making the rol-traj-yolov5 algorithm more focused on the target region of interest when extracting features and better filtering background noise.

[0022] Furthermore, the yolo_head is:

[0023] The detection head part of the network uses a Double-Head module. This module has a fully connected head focused on classification and a convolutional head for bounding box regression. Because the fc-head has higher spatial sensitivity than the conv-head, the fc-head has a stronger ability to distinguish between complete targets and partial targets, but it is not robust to regressing the entire target. Therefore, using the Double-Head to predict coordinates and categories separately can make the detection results more robust.

[0024] Furthermore, step two specifically involves:

[0025] During the network prediction phase, the network predicts the coordinates, category, and confidence level of the target information to be detected in the input data.

[0026] The coordinate information is the coordinates x, y, w, h, θ of the prediction box of the target by the network. Here, x, y, w, and h are the coordinates of the center point of the prediction box and the width and height of the prediction box, respectively, and θ is the rotation angle of the bottom edge of the prediction box relative to the horizontal coordinate axis, which is within the range of 0-180°.

[0027] During post-prediction processing, the predicted coordinate information is converted into the absolute coordinates of the four corners of the target box. The final prediction result is the category of the predicted target trajectory of the input trajectory and the position information of the four coordinates of the predicted box.

[0028] Furthermore, step three includes:

[0029] By analyzing the state of the aircraft trajectory points in the input data, we can obtain the target location information, target behavior category, and target confidence level of this trajectory point;

[0030] By pre-labeling the relationship between the behavior categories of the target, the aircraft model, and the mission, the specific model of the aircraft and the specific mission can be deduced.

[0031] After detection, the detection results are processed, and the location information of the detected target is fed into the starting point detection algorithm to calculate the start and end points of the trajectory behavior.

[0032] The technical effects of this invention are as follows:

[0033] The advantages of the flight trajectory point behavior pattern analysis and starting point detection technology based on deep learning image methods of the present invention are as follows:

[0034] (1) Faced with massive trajectory point datasets, images are more sensitive to behavioral pattern analysis than traditional machine learning algorithms. Trajectory point to image detection can use better learning of behavioral patterns such as trajectory point shape than neural networks.

[0035] (2) Compared with other latitude and longitude coordinate projection methods, the universal transverse Mercator projection has conformal properties. Angles on the sphere are projected onto the plane without changing, which can more accurately represent the two-dimensional plane coordinates of trajectory points.

[0036] (3) The rol-traj-yolov5 algorithm uses Double-Head for the yolo_head detection head. The Double-Head method has a fully connected head focused on classification and a convolutional head for bounding box regression. Predicting coordinates separately from categories can make the detection results more robust.

[0037] (4) The rol-traj-yolov5 algorithm uses a rotating box detection, which can be closer to the location of a specific trajectory than the traditional YOLO series target detection algorithms. The analysis and location detection of trajectory points are more accurate, and the calculation of the approximate start and end points of a specific trajectory is more accurate. It also filters out a large number of useless trajectory points for calculating the start point of a specific trajectory. Attached Figure Description

[0038] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0039] Figure 1 This is a schematic diagram of the flight trajectory point pattern recognition and detection module in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of the trajectory point to image conversion algorithm 1 in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of target detection algorithm 2 in an embodiment of the present invention;

[0042] Figure 4This is a schematic diagram of the starting point detection algorithm 3 in this embodiment of the invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] Example 1

[0045] like Figure 1-4 As shown, this embodiment provides a method for analyzing flight trajectory point behavior patterns based on deep learning image methods:

[0046] The present invention provides a flight trajectory point behavior pattern analysis and start-point detection technology based on deep learning image methods, such as... Figure 1 As shown, it consists of trajectory point to image algorithm 1, target detection algorithm 2, and starting point detection algorithm 3.

[0047] The trajectory point to image algorithm module 1 consists of a general transverse Mercator (UTM) algorithm and global normalization. For example... Figure 1 In the target recognition and detection technology, the flight trajectory points are converted into a trajectory image by the trajectory point to image algorithm module 1. Then, the target detection algorithm 2 uses the ROL-TRAJ-YOLOv5 algorithm to detect the specific trajectory pattern, confidence level, and position information of the target in the trajectory. Finally, the position information of the detected trajectory points is sent to the starting point detection algorithm 3 to calculate the starting point of the target trajectory in the trajectory. Finally, the output is the behavior pattern, confidence level, coordinates, and starting point of the target trajectory in the trajectory.

[0048] like Figure 2 As shown, the trajectory point to image algorithm module 1 includes a general transverse Mercator algorithm (UTM) and a global normalization algorithm. The global normalization algorithm uses the leftmost and rightmost edges of the trajectory points with x and y coordinates formed by the UTM algorithm as boundaries, calculates the normalization value for all trajectory points using these boundaries, and finally multiplies it by the image size to transform the trajectory point coordinates onto the image. Then, it uses OpenCV's polylines algorithm to convert the trajectory point coordinates into a trajectory point image.

[0049] like Figure 3 As shown, target detection algorithm 2 extracts three feature layers from the backbone network. After upsampling, the three feature layers are concatenated with other feature layers to perform target detection, outputting candidate boxes for three scales of large, medium and small targets.

[0050] like Figure 4As shown, the starting point detection algorithm 3 traverses all trajectory points in this trajectory using a traversal algorithm. The is_in_poly algorithm is used to determine whether a point is inside a polygon by using a ray method. Simply draw a ray horizontally to the right from the point. If the number of intersections with the polygon is odd, the point is inside the polygon; otherwise, it is outside the polygon. This process is used to extract all trajectory points within the predicted target box. Then, the maximum and minimum times of the time taken by traversing each trajectory point are obtained from the filtered trajectory points to extract the starting point of this specific trajectory.

[0051] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for analyzing flight trajectory point behavior patterns based on deep learning image methods, characterized in that, Includes the following steps: Step 1: Convert the latitude and longitude of the flight trajectory points into x and y coordinates on the image using the Universal Transvers Mercator method; Step 2: Use the rol-traj-yolov5 algorithm to adaptively rotate the bounding box detection algorithm to predict the position of a specific trajectory and perform trajectory pattern recognition; Step 3: Analyze the specific behavioral patterns of the trajectory to determine the specific aircraft model, mission, and flight route information that formed the trajectory; The rol-traj-yolov5 algorithm in step two is a rotating bounding box algorithm based on the YOLOv5-DOTA-OBB algorithm. This algorithm adopts a one-stage algorithm model and modifies the yolov5x model. The rol-traj-yolov5 algorithm consists of backbone, neck, and yolo_head. The backbone is the main part of the rol-traj-yolov5 algorithm. The main body uses the CSPDarnet53 structure and the focus structure. The focus layer transforms the information on the wh plane to the channel dimension, and then extracts different features through convolution. This method can reduce the information loss caused by downsampling. In the aforementioned neck structure, SPP is replaced with SPPF. Both serve the same purpose, but the latter greatly reduces the computational load of the model while maintaining the same feature extraction effect, thus accelerating the model's inference process. The SPPF structure employs a receptive field mechanism, converting the three pooling layers of different sizes in the SPP structure into a single 5x5 pooling layer. The SPPF structure obtains the same receptive field size as the SPP structure through the 5x5, 9x9, and 13x13 pooling layers by connecting three 5x5 pooling layers in series, and reduces the pooling layer size, thus reducing the computational load. The neck structure also includes New CSP-PAN, which incorporates a CSP network residual module into the feature fusion module PAN structure to prevent redundant information when calculating feature gradients. Furthermore, CBAM attention mechanisms are added to the upsampling of the PAN, inferring the image's attention in two independent dimensions: channel and space. The attention map is then multiplied by the input feature map, making the rol-traj-yolov5 algorithm more focused on the target region of interest when extracting features and better filtering background noise. The yolo_head is: The detection head part of the network uses a Double-Head module, which has a fully connected head focused on classification and a convolutional head for bounding box regression. Because the fc-head has higher spatial sensitivity than the conv-head, the fc-head has a stronger ability to distinguish between complete targets and partial targets, but it is not robust to regressing the entire target. Therefore, using the Double-Head to predict coordinates and categories separately can make the detection results more robust.

2. The method for analyzing flight trajectory point behavior patterns based on deep learning image methods according to claim 1, characterized in that, Step one specifically involves: The latitude and longitude coordinates of the trajectory points are converted into x and y coordinates on the image using a general transverse Mercator transform algorithm and normalization. Then, OpenCV is used to plot the planar projection coordinates of the trajectory points on the image to generate the image.

3. The method for analyzing flight trajectory point behavior patterns based on deep learning image methods according to claim 1, characterized in that, Step two specifically involves: During the network prediction phase, the network predicts the coordinates, category, and confidence level of the target information to be detected in the input data. The coordinate information is the coordinates x, y, w, h, ɵ of the prediction box of the target by the network. Here, x, y, w, h are the coordinates of the center point of the prediction box and the width and height of the prediction box, respectively, and ɵ is the rotation angle of the bottom edge of the prediction box relative to the horizontal coordinate axis, which is within the range of 0-180°. During post-prediction processing, the predicted coordinate information is converted into the absolute coordinates of the four corners of the target box. The final prediction result is the category of the predicted target trajectory of the input trajectory and the position information of the four coordinates of the predicted box.

4. The method for analyzing flight trajectory point behavior patterns based on deep learning image methods according to claim 1, characterized in that, Step three includes: By analyzing the state of the aircraft trajectory points in the input data, the target location information, target behavior category, and target confidence level of this trajectory point can be obtained. By pre-labeling the relationship between the behavior categories of the target, the aircraft model, and the mission, the specific model of the aircraft and the specific mission can be inferred. After detection, the detection results are processed, and the location information of the detected target is fed into the starting point detection algorithm to calculate the start and end points of the trajectory behavior.

Images