Aircraft position positioning method and system based on wake cloud age

Abstract

The application discloses a kind of aircraft position positioning method and system based on tail cloud age, it is related to aviation monitoring and computer vision technical field, the method includes collecting binocular image sequence containing target aircraft tail cloud;The apparent age of the tail cloud region in the image and its different segments along the extension direction of tail cloud is identified, and the three-dimensional position information of each point on the tail cloud region is calculated based on binocular vision geometric model, and then the age gradient vector reflecting the change direction of tail cloud new and old is fitted;According to the three-dimensional coordinates of the age gradient direction and tail cloud, the real-time position, heading and speed of the aircraft are calculated, the data of each observation station is received, target correlation and data fusion are carried out, and the fused aircraft track information is output;The method realizes pure passive positioning independent of target cooperation, significantly improves the continuity, accuracy and real-time of positioning.

Description

Technical Field

[0001] This invention relates to the fields of aviation surveillance and computer vision technology, specifically to a method and system for aircraft positioning based on contrail age. Background Technology

[0002] Continuous and effective surveillance of high-altitude and low-observable aircraft is a major challenge in aviation safety and airspace management. Currently, existing technologies, primarily active radar and direct electro-optical tracking, rely on signal reflection or radiation from the target itself, which has inherent limitations when facing stealth designs and complex weather conditions. This technological bottleneck has prompted research to focus on persistent contrails generated by aircraft, which are difficult to conceal. As a natural physical phenomenon strongly correlated with flight paths, contrails offer a new approach to passive surveillance.

[0003] In recent years, existing research on contrails has largely focused on assessing their impact on climate, with monitoring primarily relying on satellite remote sensing. However, the spatial resolution and temporal sampling frequency of satellite images are insufficient, making it difficult to detect contrails immediately after their formation and reliably associate specific contrails with the flights that generated them. To compensate for the limitations of satellite observations, ground-based observation networks have become an important research direction due to their high spatiotemporal resolution. For example, some scholars have used ground-based sky cameras to observe contrails and elucidate the microphysical formation process of contrail ice crystals. Simultaneously, deep learning-based target detection and segmentation techniques have been introduced into the automatic identification of contrails. For instance, an open-source dataset released in 2025 was specifically designed to train and validate contrail detection algorithms based on ground-based images. The dataset's annotations even categorize contrails as "young" or "old" based on their duration, reflecting an initial focus on the temporal dimension of contrail "age."

[0004] However, while existing technologies combining ground-based optical observation with deep learning-based intelligent contrail image recognition can automatically detect and segment contrail regions from ground-based images, and even perform preliminary classification of their duration, most of them remain at the level of two-dimensional existence detection, classification, or climate effect analysis of contrails. They fail to deeply integrate the morphological and age characteristics of contrails with their precise three-dimensional geospatial coordinates, thus making it impossible to locate and track aircraft through contrails. Summary of the Invention

[0005] To address the shortcomings of existing technologies that fail to deeply integrate the morphological age characteristics of contrails with their precise three-dimensional geospatial coordinates, thus hindering the location and tracking of aircraft using contrails, this invention proposes an aircraft location method and system based on contrail cloud age, thereby solving the problems existing in the prior art.

[0006] A ground-based observation method for locating an aircraft by measuring the age of its contrail, deployed at at least two ground observation stations in different geographical locations, includes the following steps: Each ground observation station simultaneously acquires a sequence of binocular images containing the contrails of the target aircraft; Based on binocular image sequences, the contrail cloud region in the image is identified; and based on the morphological characteristics of the contrail cloud region and its physical model of evolution over time, the apparent age of different segments along the contrail cloud extension direction is analyzed; and the three-dimensional position information of each feature point on the contrail cloud region is calculated based on the binocular visual geometric model; and based on the three-dimensional position information and apparent age, an age gradient vector reflecting the direction of change of the contrail cloud is fitted. The flight course of the aircraft is inferred from the direction of the age gradient vector, and the average speed of the aircraft in the corresponding time period is inferred by combining the three-dimensional position information of the contrail feature points with different age characteristics and their corresponding timestamps. Based on the latest observed contrail feature points, the aircraft's three-dimensional position at the current moment is predicted by extrapolating along the flight course and based on the average speed. The three-dimensional position, flight heading, and average speed predicted from different ground observation stations are correlated and fused to generate continuous three-dimensional flight path information for the aircraft.

[0007] Furthermore, the process of fitting an age gradient vector reflecting the direction of change in the wake cloud based on three-dimensional location information and apparent age specifically includes the following steps: The identified wake cloud region is divided into multiple continuous segments along its main axis. Based on the morphological characteristics of each segment of the contrail cloud, and combined with a thermodynamic model, a state age is assigned to each segment; the morphological characteristics include diffusion width, edge sharpness, and optical thickness; Linear fitting is performed based on the spatial location sequence of each segment and its corresponding state age to obtain the age gradient vector of the wake cloud along its extension direction.

[0008] Furthermore, the identification of the contrail cloud region in the image based on the binocular image sequence specifically includes the following steps: The acquired binocular image sequence is input into the YOLOv8-tiny-based wake cloud recognition model. Through the backbone network, multi-scale feature extraction is performed on the binocular images to obtain the initial feature map. By using a feature enhancement layer incorporating the CBAM attention mechanism, channel and spatial attention weighting is applied to the initial feature map to obtain an enhanced feature map focused on the wake cloud region. Multi-scale feature fusion is performed on the enhanced feature map using a feature pyramid network; Based on the fused feature map, the probability of each pixel belonging to the trail cloud category is calculated and output through the segmentation head; Based on the probability, the whet cloud region in the image is identified.

[0009] Furthermore, the calculation of the three-dimensional position information of each feature point on the wake cloud region based on the binocular vision geometric model specifically includes the following: Distortion and epipolar correction are performed on binocular images; Identify the wake cloud region as the region of interest in the corrected binocular image; Dense stereo matching is performed within the region of interest to generate a disparity map for each pixel. Based on the binocular vision geometric model, the three-dimensional coordinates of each point on the contrail are calculated by combining the camera's intrinsic and extrinsic parameters, and then the average altitude of the contrail relative to the observation station is calculated; wherein, the binocular vision geometric model is expressed as Z=(f×B) / d, where Z is the depth, f is the focal length, B is the baseline distance, and d is the parallax.

[0010] Furthermore, the step of using the latest observed contrail feature points as a reference, extrapolating along the flight path and based on the average velocity to predict the aircraft's three-dimensional position at the current moment also includes determining whether the aircraft body has left the current image field of view. Specifically, this includes: when it is determined that the aircraft body is within the current image field of view, using the direct growth vector method to locate the aircraft's three-dimensional position at the current moment; when it is determined that the aircraft body has left the current image field of view, using the age gradient extrapolation method to locate the aircraft's three-dimensional position at the current moment.

[0011] Furthermore, when it is determined that the aircraft body is within the current image field of view, the direct vector growth method is used to locate the three-dimensional position of the aircraft at the current moment, specifically including the following steps: Temporal 3D point cloud of the head of the wake cloud is extracted from a series of continuously acquired binocular images; By tracking the displacement of the centroid of the head point cloud in three-dimensional space between consecutive frames, the spatial vector representing the growth direction and rate of the wake cloud can be directly calculated. The direction of the spatial vector representing the growth direction and rate of the contrail cloud is inverted to the real-time heading of the aircraft, and its magnitude is divided by the time interval to invert the real-time speed of the aircraft. The real-time heading of the aircraft is combined with the three-dimensional coordinates of the contrail head, and the real-time three-dimensional position of the aircraft is directly calculated by compensating for a preset geometric offset between the aircraft and the contrail generation point in the opposite direction to the spatial vector; wherein the geometric offset is estimated by a preset model of aircraft type, atmospheric conditions and engine parameters.

[0012] Furthermore, when it is determined that the aircraft body has left the current image view, the age gradient extrapolation method is used to locate the three-dimensional position of the aircraft at the current moment, specifically including the following steps: The age gradient vector is transformed from the image coordinate system to the geographic coordinate system with the observation station as the origin, and the horizontal heading of the aircraft is initially estimated based on the opposite direction of the projection of the transformed age gradient vector onto the horizontal plane. Select two time intervals Δt=A k -A j The age points j and k; Based on the coordinates and time difference of age points j and k, calculate the historical average speed and vertical speed of the aircraft within the time period Δt; Using the youngest feature point on the contrail as the starting point, the displacement from the observation time corresponding to the starting point to the current time is extrapolated along the initially estimated horizontal heading of the aircraft and based on its historical average speed and vertical speed, to calculate the predicted three-dimensional position of the aircraft at the current time.

[0013] Furthermore, the predicted three-dimensional position of the aircraft at the current moment is calculated using the geodesic model. The calculation process is as follows: spherical distance s=V h *t ext / R; Latitude φ current =arcsin(sin(φ new )*cos(s)+cos(φ new )*sin(s)*cos(Heading_final)); Longitude difference Δλ = arctan2(sin(Heading_final)*sin(s)*cos(φ) new ),cos(s)-sin(φ new )*sin(φ current )); Longitude λ current =λ new +Δλ; Height H current =H new +V H *t ext ; Where R is the Earth's radius; V h The magnitude of the historical average speed; V H The historical average vertical velocity magnitude; the feature point P with the youngest age on the contrail. new The three-dimensional coordinates are (λ new ,φ new H new The observation time is T.new Heading_final is the final heading; t ext =T now -T new For extrapolation time, T now λ represents the current moment. new For P new longitude, φ new For P new latitude, H new For P new The height.

[0014] Furthermore, the process of correlating and fusing the predicted three-dimensional position, flight heading, and velocity calculated from different ground observation stations to generate continuous three-dimensional flight path information for the aircraft specifically includes the following steps: Receive target status data from multiple ground observation stations; the target status data includes at least the predicted position, heading, velocity, and the station's estimated confidence level. Based on the similarity of target location and motion state, determine whether the data reported by different observation stations belong to the same aircraft target; For multiple sets of data identified as the same target, a Kalman filter algorithm is used to fuse them, and the fused three-dimensional position, velocity, heading and unique tracking identifier of the aircraft are output. The fusion weight of the data from each observation station is related to the estimated confidence level and the geometric configuration between the station and the target.

[0015] The present invention also includes a ground-based observation system for locating aircraft positions based on contrail age, comprising: The acquisition module is used to simultaneously acquire binocular image sequences containing the contrails of the target aircraft at each ground observation station; The contrail age analysis module is used to identify contrail regions in images based on binocular image sequences; and to analyze the apparent age of different segments along the contrail extension direction based on the morphological characteristics of the contrail region and its physical model of evolution over time; and to calculate the three-dimensional position information of each feature point on the contrail region based on the binocular visual geometric model; and to fit an age gradient vector reflecting the direction of contrail aging based on the three-dimensional position information and apparent age. The aircraft position estimation module is used to infer the aircraft's flight heading based on the direction of the age gradient vector, and combine the three-dimensional position information of different age feature points on the contrail region and their corresponding timestamps to invert the aircraft's average speed within the corresponding time period; based on the latest observed contrail feature points, it extrapolates the motion along the flight heading and based on the average speed to predict the aircraft's three-dimensional position at the current moment. The generation module is used to correlate and fuse the three-dimensional position, flight heading, and average speed predicted from different ground observation stations to generate continuous three-dimensional flight path information of the aircraft.

[0016] This invention provides a ground-based observation method for locating aircraft positions based on the age of contrails, which has the following advantages: This invention achieves high-precision, all-weather measurement of the three-dimensional spatial position of contrails by combining binocular vision; and by combining the three-dimensional geographic information of the contrails with the daily humidity conditions, it quantitatively analyzes the "state age" of the contrails to extract time dimension information from static images, providing key input for retrieving motion trajectories; and by directly inferring the heading using the physical evolution (age gradient) of the contrails themselves, and by extrapolating the real-time position using three-dimensional geometric and kinematic models, it achieves purely passive positioning without relying on target cooperation, significantly improving the continuity, accuracy, and real-time performance of positioning; this method deeply integrates the morphological age characteristics of the contrail with its precise three-dimensional geographic spatial coordinates, and by deploying multiple observation stations for data fusion, it significantly improves the accuracy of locating and tracking aircraft. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of a ground observation method for locating an aircraft position based on the age of contrails, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the standard YOLOv8 network structure in an embodiment of the present invention; Figure 3 This is a schematic diagram of the age recognition results of the contrail cloud in an embodiment of the present invention. Detailed Implementation

[0018] 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 only some embodiments of the present invention, and not all embodiments.

[0019] Continuous and precise location and tracking of high-altitude, high-speed aircraft with radar stealth capabilities is a major challenge facing existing air surveillance systems. The main technical approaches currently employed and their shortcomings include: (1) Active radar detection: relies on the reflection of electromagnetic waves by the target. For stealth aircraft that use radar-absorbing materials and have special shape designs, their radar cross section (RCS) is significantly reduced, resulting in a sharp reduction in detection range, unstable tracking, or even failure. In addition, actively transmitted signals are susceptible to electronic interference and have geographical blind spots.

[0020] (2) Direct optical / infrared tracking: attempts to directly capture images of the aircraft itself. However, at long distances, the target occupies only a few pixels in the image. Under complex sky backgrounds, cloud cover, or strong light conditions, the signal-to-noise ratio is extremely low, making target extraction and continuous tracking extremely difficult.

[0021] (3) Satellite remote sensing and reconnaissance: Although it can monitor a wide area, the revisit cycle is long, the cost is high, and the real-time performance is poor, making it difficult to meet the tactical level continuous monitoring requirements.

[0022] (4) Surveillance of cooperative targets (such as ADS-B): cannot be applied to non-cooperative targets that are not equipped with the system.

[0023] In summary, the fundamental reason for the shortcomings of existing technologies lies in the fact that their detection signals directly affect the aircraft itself, making them easily evaded, either actively or passively. In contrast, persistent contrails generated by aircraft under specific atmospheric conditions are a natural physical phenomenon strongly correlated with their flight paths and difficult to actively eliminate. Contrails are formed by the condensation of volatile and non-volatile particles from aircraft engine emissions in the ice supersaturated region (ISSR). They can remain at high altitudes for several hours and disperse, providing targets with a difficult-to-hide indirect signal source.

[0024] Based on this, the present invention proposes a ground observation method for locating aircraft positions by analyzing the age of contrails. This method analyzes the morphology, three-dimensional position and temporal evolution characteristics of aircraft contrails (condensed contrails) to invert, locate and track high-altitude aircraft in a purely passive manner. It can be applied to scenarios such as airspace management and aviation stealth feature analysis.

[0025] like Figure 1 As shown, this method is applied to at least two ground observation stations deployed in different geographical locations, and specifically includes the following steps: S1. Synchronous Image Acquisition: Select 5 to 10 open locations in the shooting area to deploy Raspberry Pi 4B devices, each equipped with two HQ Camera Modules (12 megapixels), a gimbal, and a GPS module for fixed shooting of airspace images.

[0026] A synchronous controller is used to ensure that the timestamps of the binocular image acquisition are strictly aligned; it runs at a frequency of 1 frame per second for at least 30 days, retaining about 2,000 sets of JPEG images with consistent timestamps (5-10 images per set), and simultaneously recording the timestamp, GPS positioning information, and azimuth and pitch angles of each image.

[0027] S2. Image Annotation: Acquired images are uploaded to the CVAT platform. Each contrail segment is divided into 5 blocks. Annotators manually and meticulously annotate the average age of the contrail in each block, combining time and image features, and construct a contrail evolution status label system, categorized as follows: Category 0: Contrail - Newly generated ~ 0 minutes ago (the lines are thin, clear, and consistent with the flight path).

[0028] Category 1~300: Wheal cloud generation assigns a label every 2 seconds until 10 minutes later.

[0029] Category 301: Non-contrail clouds (natural clouds, blue sky background, other cirrus clouds, etc.).

[0030] Each image supports multi-target annotation, using the YOLO standard (class x_center y_centerwidth height), with all coordinates normalized to the default image resolution (1920×1080). The total number of annotated samples is controlled to be no less than 10,000 images, ensuring that there are no less than 2,000 images for each class of labels.

[0031] S3. Contrail Recognition: Load a multi-task convolutional neural network for temporal contrail recognition. This network takes a pre-processed monocular (such as the left camera) image sequence frame as input and outputs a contrail segmentation mask to distinguish the contrail from the sky background. It will also be used to determine whether the contrail extends outside the image and whether the contrail changes and lengthens over time, thereby making a preliminary judgment on whether the aircraft is still within the field of view.

[0032] The YOLOv8-tiny model is used as the detection backbone structure. The main network structure includes: Backbone: CSPDarknet-Tiny, which has residual connections and C2f structure; Neck: PAN-FPN structure, used for multi-scale feature fusion; Detection Head: three output branches, which are responsible for small, medium and large scale object detection respectively; The subsequent steps will be different depending on whether the aircraft is in the image. If the aircraft is in the image, the steps will be S4→S5→S8. If the aircraft is not in the image, the steps will be S4→S6→S7→S8.

[0033] S4. Contrail 3D Height Calculation: Receive synchronized detection images and use step S3 to detect the contrail position. Then, process the binocular images as follows to calculate the contrail height H. cloud : Distortion and epipolar corrections are performed on the left and right images.

[0034] Identify the wake cloud region as the region of interest in the left image.

[0035] Dense stereo matching is performed within this area to calculate the disparity of each pixel.

[0036] Based on the binocular vision geometric model Z=(f×B) / d, where Z is depth, f is focal length, B is baseline distance, and d is parallax; combined with camera intrinsic and extrinsic parameters, the parallax map can be converted into three-dimensional point cloud coordinates of each point on the contrail, and then the average altitude H of the contrail relative to the observation station can be calculated. cloud .

[0037] S5. Temporal Image Localization: When the aircraft is in the image but its main body is difficult to observe, a high-frequency continuous acquisition of binocular image sequences is performed. First, a two-dimensional mask of the contrail region in each frame is extracted using step S3. Then, based on the results of the three-dimensional altitude calculation module, the binocular images are quickly matched and reconstructed using this mask to generate a temporal three-dimensional point cloud of the contrail head (i.e., the most recently generated part) in real time. By tracking the displacement of the centroid or feature points of this head point cloud in three-dimensional space between consecutive frames, the spatial vector representing the growth direction and rate of the contrail is directly calculated. This vector is the high-precision inversion of the aircraft's real-time motion vector. Finally, this motion vector is combined with the three-dimensional coordinates of the contrail head in the latest frame. Compensation is performed using a preset geometric offset model between the aircraft and the contrail generation point to directly calculate the aircraft's precise three-dimensional geographical location, velocity vector, and heading at the current moment. This achieves direct localization without identifying the aircraft itself, relying solely on the temporal morphological changes of the contrail.

[0038] S6. Contrail Age Analysis: When the aircraft is not in the image and only the visible condensation contrail is retained, the theoretical diffusion image of the contrail over a 5-minute period is simulated using the contrail's latitude, longitude, altitude, and air humidity. This theoretical image and a sequence of synchronized image frames detected by contrail position detection are input into a YOLOv8 neural network. The contrail within the mask is segmented along its principal axis, and the probability distribution of the "apparent age" of each segment is output. Based on the morphological characteristics of the contrail (such as diffusion width, edge sharpness, and optical thickness) and the "state age" mapped from the thermodynamic model, the age gradient vector ▽Age of the contrail along its extension direction is fitted according to the segmented position sequence and its corresponding "state age." The direction of this vector indicates the age of the contrail and points in the opposite direction to the aircraft's flight direction.

[0039] S7. Aircraft Position Deduction: When the aircraft is not in the image and only a visible contrails remain, the core of deducing the aircraft's position through contrail age includes the following steps: Heading determination: Based on the ▽Age direction and with true north as the reference, determine the approximate heading angle θ of the aircraft.

[0040] Historical trajectory point inversion: Two feature points P1 and P2 with ages t1 and t2 (t2>t1) on the contrail cloud are selected. The observation times T1 and T2 corresponding to these two points are known, as well as their three-dimensional spatial positions (longitude λ1, latitude φ1, altitude H1) and (λ2, φ2, H2) calculated through binocular vision (where H1≈H2≈H2). cloud Assuming the aircraft flies at a constant speed in a straight line during the time interval Δt = t2 - t1, its historical average velocity vector can be calculated.

[0041] Real-time position extrapolation: Using the historical velocity vector V and heading θ obtained from inversion, the motion is extrapolated along the heading starting from the three-dimensional position of the latest observed contrail cloud point; extrapolation time t extrapolate The position is obtained by subtracting the observation time corresponding to the latest contrail cloud point from the current system time. This allows for the estimation of the aircraft's predicted position (λ) at the current moment. current ,φ current H current ).

[0042] S8. Multi-station data fusion and visualization output: The central server receives the predicted aircraft position, altitude, heading, speed, and confidence level reported by each observation station. The nearest neighbor method is used to determine whether targets reported by different stations belong to the same aircraft. For multiple successfully correlated observation data, a Kalman filter algorithm is used for fusion, where the weights are inversely proportional to the geometric configuration from the observation station to the target and the estimated confidence level. Finally, the fused real-time 3D position, speed, heading, and tracking identifier of the aircraft are output.

[0043] The control interface is built using Flask and Vue, supporting image preview, recognition result display, contrail cloud status statistics, latitude and longitude display, altitude display, parameter adjustment, and 3D coordinate display of aircraft position prediction.

[0044] This invention creatively shifts the observation target from the aircraft itself to the contrail it generates, which is difficult to conceal. This effectively circumvents stealth designs targeting the aircraft itself and provides a novel technical approach to solving the problem of detecting high-altitude, stealthy targets. Specifically, it offers the following advantages in terms of technological innovation and practical application: (1) Binocular vision and neural network cloud age analysis: Binocular vision was combined to achieve high-precision, all-weather measurement of the three-dimensional spatial position of the contrail cloud; and combined with the three-dimensional geographic information of the contrail cloud and the humidity conditions of the day, the “state age” of the contrail cloud was quantitatively analyzed by a pre-trained neural network, realizing the extraction of time dimension information from static images, which provided key input for inverting the motion trajectory.

[0045] (2) Position estimation algorithm based on age gradient: The heading is directly inferred by the physical evolution (age gradient) of the wake cloud itself, and the real-time position is extrapolated by combining the three-dimensional geometry and kinematic model. A complete and closed technical chain is formed from "passive observation of phenomena" to "active estimation of dynamic targets", realizing pure passive positioning without relying on target cooperation.

[0046] (3) Multi-station data fusion: By deploying multiple observation stations and fusing data, the problems of directional ambiguity and large error in single-station observation are overcome, and the positioning accuracy, system redundancy and resistance to local interference (such as cloud cover of single station) are significantly improved.

[0047] Based on the same inventive concept, this invention proposes a ground-based observation system for locating aircraft positions using contrail age. This system consists of at least two geographically dispersed ground observation stations (nodes) and a central data processing center. Each observation station includes the following core modules: Synchronous Image Acquisition Module: Deploy Raspberry Pi 4B devices at 5 to 10 open locations within the shooting area. Each device is equipped with two HQ Camera Modules (12 megapixels), a gimbal, and a GPS module for fixed aerial image capture. The module has a built-in synchronization controller to ensure strict alignment of timestamps during binocular image acquisition. Run the module at a frame rate of 1 second per day for at least 30 days, retaining approximately 2000 sets of JPEG images (5-10 images per set) with consistent timestamps. Simultaneously record the timestamp, GPS positioning information, and gimbal azimuth and pitch angles for each image.

[0048] Image annotation module: Collected images are uploaded to the CVAT platform, and each contrail cloud segment is divided into 5 blocks. Annotators manually and meticulously annotate the average age of the contrail cloud in each block by combining time and image features, and construct a contrail cloud evolution status label system.

[0049] Contrail Recognition Module: This module loads a multi-task convolutional neural network for temporal contrail recognition. The network takes pre-processed monocular (e.g., left camera) image sequences as input and outputs a contrail segmentation mask to distinguish the contrail from the sky background. This module also determines whether the contrail extends beyond the image and whether it lengthens over time, thus providing a preliminary assessment of whether the aircraft is still within the field of view. This module uses the YOLOv8-tiny model as the detection skeleton structure.

[0050] Contrail 3D Height Calculation Module: Receives synchronized detection images and uses the contrail recognition module to detect the contrail position. Then, it processes the binocular images as follows to calculate the contrail height H. cloud .

[0051] Temporal Image Localization Module: When the aircraft is in the image but its main body is difficult to observe, the system continuously acquires binocular image sequences at high frequency. First, it uses the contrail recognition module to extract a two-dimensional mask of the contrail region in each frame. Then, based on the results of the three-dimensional height calculation module, it performs rapid stereo matching and reconstruction of the binocular images using this mask, generating a temporal three-dimensional point cloud of the contrail head (i.e., the most recently generated part) in real time. By tracking the displacement of the centroid or feature points of this head point cloud in three-dimensional space between consecutive frames, the system directly calculates the spatial vector representing the growth direction and rate of the contrail. This vector is the high-precision inversion of the aircraft's real-time motion vector. Finally, the system combines this motion vector with the three-dimensional coordinates of the contrail head in the latest frame and compensates for it using a preset geometric offset model between the aircraft and the contrail generation point. This directly calculates the aircraft's precise three-dimensional geographical location, velocity vector, and heading at the current moment, thus achieving direct localization without recognizing the aircraft itself, relying only on the temporal morphological changes of the contrail.

[0052] Contrail Age Analysis Module: When the aircraft is not in the image and only the visible condensation contrail is retained, the theoretical diffusion image of the contrail over a 5-minute period is simulated using the contrail's latitude, longitude, altitude, and air humidity. This theoretical image and a sequence of synchronized image frames detected by contrail position detection are input into a YOLOv8 neural network. The contrail within the mask is segmented along its principal axis, and the probability distribution of the "apparent age" of each segment is output. This model determines the "apparent age" based on the "state age" mapped from the contrail's morphological characteristics (such as diffusion width, edge sharpness, and optical thickness) and a thermodynamic model. The system fits the age gradient vector ▽Age of the contrail along its extension direction based on the segmented position sequence and its corresponding "apparent age." The direction of this vector indicates the contrail's age, while the opposite direction points towards the aircraft's flight direction.

[0053] Aircraft position estimation module: This module is the core of aircraft position estimation when the aircraft is not in the image but only the visible contrails are retained. It includes heading determination, historical trajectory point inversion and real-time position extrapolation.

[0054] Multi-station data fusion and visualization output module: The central server receives the predicted aircraft position, altitude, heading, speed, and confidence level reported by each observation station. It uses the nearest neighbor method to determine whether targets reported by different stations belong to the same aircraft. For multiple successfully correlated observation data, a Kalman filter algorithm is used for fusion, where the weights are inversely proportional to the geometric configuration from the observation station to the target and the estimated confidence level. The final output is the fused real-time 3D position, speed, heading, and tracking identifier of the aircraft. A control interface is built using Flask+Vue, supporting image preview, recognition result display, contrail status statistics, latitude and longitude display, altitude display, parameter adjustment, and display of the predicted 3D coordinates of the aircraft position.

[0055] Each observation node runs the following process independently: S1. Synchronous Trigger Acquisition: The system commands the image acquisition module to synchronously capture binocular images according to a sampling rate of once per second, and packages the image data, timestamp, and attitude angle data to form a raw data packet.

[0056] S2. Simultaneous Processing: The raw data packet is simultaneously fed into two processing pipelines. Pipeline A (height calculation module) performs stereo matching and 3D reconstruction, outputting the 3D point cloud of the contrail and its average height H. Pipeline B (cloud age analysis module) feeds the left-eye image into the neural network, outputting the contrail segmentation mask and segmentation information.

[0057] Internal logic of the wake cloud recognition and age analysis module: The neural network structure for contrail recognition adopts the YOLOv8 network structure. The network outputs three heads: a segmentation head (Softmax outputs the probability that each pixel belongs to the contrail / background), a preliminary age regression head (sampling and dividing the segmented contrail region into blocks at equal intervals along its skeleton line, and outputting a regression value representing its apparent age for each block), and an aircraft determination head (determining whether the aircraft is still within the observation range by whether the contrail extends). The backbone network introduces a CBAM attention mechanism module, which guides the neural network to automatically focus on the contrail region during the feature extraction stage through joint modeling of channel attention and spatial attention. The temporal category information is converted into a discrete vector through one-hot encoding. This temporal feature is concatenated with the image spatial feature tensor along the channel dimension to form a joint feature map, enabling the model to not only have spatial perception capabilities but also to identify the state differences of the contrail in the temporal evolution dimension.

[0058] Training Method: A large dataset of annotated contrail images was used for training. Label generation was completed by importing image data into the CVAT annotation platform, with manual annotation performed by annotators based on image content and capture time information. By analyzing the morphological evolution of contrails after generation, annotators categorized image targets into 300 contrail evolution states: newly generated and annotated every 2 seconds from 0 to 10 minutes. A non-contrail category was also added as a control. The bounding boxes and category information of targets in each image were uniformly annotated according to YOLO format, and the coordinates were normalized to suit the training requirements of the neural network. The number of annotated samples was controlled to be over 10,000 to ensure a balanced distribution of labels for each class, facilitating the construction of a high-quality, time-series-defined training dataset.

[0059] The height of the identified contrail is calculated using a binocular vision method, and the latitude and longitude of the contrail are calculated using the shooting location, azimuth angle, and pitch angle to obtain the three-dimensional information of the contrail. It also provides the ability to query the weather conditions (wind speed, humidity, etc.) of the day to determine whether a contrail has formed at the location, and provides the theoretical diffusion image within 20 minutes.

[0060] To address the problem of temporal image inversion when an aircraft is in a photograph, a direct localization method based on real-time contrail cloud growth vectors is used, for each frame at time t. k The system first uses the contrail recognition module to obtain the two-dimensional region of the contrail in the current frame, and then drives the three-dimensional height calculation module to perform fast stereo matching and reconstruction only on this region, generating a cluster of three-dimensional point clouds P corresponding to the latest and smallest part of the contrail head. head(tk) By comparing the point cloud P between consecutive frames head(tk) With P head(tk-1) The system calculates the 3D displacement of the centroid of the point cloud or specific feature points, and defines the growth vector G of the wake cloud in 3D space at the current moment. (tk) The direction of this vector indicates the aircraft's heading, and its magnitude divided by the frame interval gives the direct observation of the aircraft's velocity. Then, the system calculates the growth vector for multiple consecutive frames. Perform smoothing filtering (such as Kalman filtering) to obtain a stable three-dimensional motion direction vector in the current geographic coordinate system (northeast to sky). The horizontal component of this vector directly determines the aircraft's instantaneous heading angle θ, while its time series magnitude, after differentiation, can output the aircraft's real-time ground speed. and acceleration information; then the system will output the current frame's wake cloud head point cloud P head(tk) The point set furthest from the previous frame's point cloud is identified as the "near end" closest to the aircraft nozzle. The centroid X-axis of this "near end" point cloud is then determined. near(tk) It is known that the contrail is continuously generated by the aircraft's exhaust nozzles; therefore, the real-time position X of the aircraft itself is... ac(tk) Not directly equal to X near(tk) Instead, it requires compensating for a tiny spatial offset in the opposite direction of the motion. This offset is estimated by a pre-set model of aircraft type, atmospheric conditions, and engine parameters. The aircraft is at time t k The real-time three-dimensional position is calculated by the following formula: ; The system continuously outputs Heading θ, speed And its timestamps, forming a continuous flight path.

[0061] Contrail Age Analysis Network Structure: To address the problem of accurate age inversion of contrails when the aircraft is not in the image, a YOLOv8 network structure is adopted. The theoretical diffusion image and the binocular vision image are stitched together and then input into the network. The network outputs three heads: a segmentation head (outputs the position and 3D position of each segmented contrail in the image), an age regression head (samples and divides the segmented contrail region into blocks at equal intervals along its skeleton line, and outputs a regression value representing its apparent age for each block), and an age uncertainty head (gives the uncertainty of the apparent age). The training method of this network is roughly the same as that of the recognition model, but the labels are re-labeled using the CVAT annotation platform based on the binocular recognition results.

[0062] Age gradient calculation: The neural network output trail cloud is divided into 5 segments, with the center pixel coordinate sequence being [C1, C2, C3, C4, C5], corresponding to age estimates of [A1, A2, A3, A4, A5]. The system calculates the age gradient for (C1, C2, C3, C4, C5). i , , A i The sequence is linearly fitted, and the direction of the fitted line is the direction of the age gradient ▽Age, and the slope reflects the rate of age change.

[0063] S3. Data Association and Estimation: The system spatially associates the 3D point cloud output from pipeline A with the segmented mask output from pipeline B, assigning real 3D coordinates to each segment and further calculating the age label for each segment. The position estimation module executes heading determination, historical trajectory inversion, and real-time position extrapolation algorithms based on the age gradient, associated 3D coordinates, and timestamps.

[0064] a. Internal logic of the aircraft position estimation module (single station): Input: Age gradient direction ▽Age (image coordinate system); a series of age-3D coordinate pairs (A i ,λ i ,φ i H i ); the corresponding timestamp T i .

[0065] Heading angle calculation: The ▽Age direction is transformed from the image coordinate system to the northeast-sky coordinate system with the observation station as the origin, using the camera model and gimbal attitude angle. The direction opposite to the projection of the transformed vector onto the horizontal plane is the preliminary estimate of the aircraft's horizontal heading (Heading_image).

[0066] b. Velocity and position inversion and extrapolation: Select two pairs of pairs with a sufficient time interval Δt=A k -A j The age points j,k(A) k >A j).

[0067] Calculate the horizontal displacement: Δd horizontal =great_circle_distance((φ j ,λ j ),(φ k ,λ k )), where great_circle_distance represents the function that calculates the distance of spherical movement using latitude and longitude coordinates.

[0068] Calculate the historical average speed magnitude: V h =Δd horizontal / Δt.

[0069] Calculate the magnitude of the historical average vertical velocity: V H =ΔH / Δt, where ΔH=H k -H j .

[0070] Determine the speed direction: Combine Heading_image and the direction of the line connecting the two points to determine the final heading Heading_final.

[0071] Let the newest point (youngest age) be P. new The three-dimensional coordinates are (λ new ,φ new H new The observation time is T. new .

[0072] The current system time is T. now Extrapolation time t ext =T now -T new .

[0073] The current aircraft position (latitude and longitude) is predicted based on the geodesic model using the following formula: φ current =arcsin(sin(φ new )*cos(s)+cos(φ new )*sin(s)*cos(Heading_final)); Δλ=arctan2(sin(Heading_final)*sin(s)*cos(φ new ),cos(s)-sin(φ new )*sin(φ current )); λ current =λ new +Δλ; H current=H new +V H *t ext ; Where R is the Earth's radius, and s = V h *t ext / R is the spherical distance, while (λ) current ,φ current H current () represents the coordinates of the current aircraft position, where the predicted altitude H current Assumptions and H new Linear extrapolation can be performed based on the same or a small number of historical high points.

[0074] S4. Data Center Fusion and Output: Observation nodes encapsulate their estimated aircraft position, speed, heading, altitude, and their own coordinates, along with the estimated confidence level, into a message and send it to the central data processing center via an encrypted communication link. The central data center receives the messages from each node, executes target association and data fusion algorithms, generates unified, higher-precision global aircraft trajectory information, and distributes it to display and control terminals or higher-level command systems.

[0075] The system triggers the synchronous image acquisition module, controlling the binocular cameras to expose synchronously and packaging the image data with sensor data. The altitude calculation module receives the image pairs, executes a stereo matching algorithm to generate a disparity map, and calculates the 3D point cloud coordinates based on calibration parameters. Simultaneously, the cloud age analysis module inputs the left-eye image into the loaded neural network model; after forward propagation, the network outputs a wake cloud segmentation map and segmented age feature maps. The position estimation module reads the point cloud data from the altitude module and the age feature map from the cloud age module, correlates age and position information in 3D space, fits the age gradient vector, and calls the heading calculation and position extrapolation algorithm to generate the target's 3D position state (λ). current ,φ current H current The message is estimated. Finally, the message is sent to the communication module, which transmits it to the central data processing center to Kalman merge the data from different sites.

[0076] The system boasts core technological highlights such as low-cost deployment, high temporal response sensitivity, and excellent environmental adaptability. It is not only suitable for contrail cloud identification under complex weather conditions, but can also be further extended to multiple practical application areas such as aviation stealth performance evaluation, airspace dynamic monitoring and early warning, and meteorological observation data calibration.

[0077] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for aircraft positioning based on contrail age, deployed at at least two ground observation stations in different geographical locations, characterized in that... Includes the following steps: Each ground observation station simultaneously acquires a sequence of binocular images containing the contrails of the target aircraft; Based on binocular image sequences, the contrail regions in the images are identified; and based on the morphological characteristics of the contrail regions and their physical evolution over time, the apparent ages of different segments along the contrail extension direction are analyzed. And calculate the three-dimensional position information of each feature point on the wake cloud region based on the binocular vision geometric model; Based on the three-dimensional location information and apparent age, an age gradient vector reflecting the direction of change in the age of the wake cloud is obtained by fitting. The flight course of the aircraft is inferred from the direction of the age gradient vector, and the average speed of the aircraft in the corresponding time period is retrieved by combining the three-dimensional position information of points with different age characteristics on the contrail region and their corresponding timestamps. Based on the latest observed contrail feature points, extrapolate the motion along the flight path and based on the average motion speed to predict the three-dimensional position of the aircraft at the current moment. The three-dimensional position, flight heading, and average speed predicted from different ground observation stations are correlated and fused to generate continuous three-dimensional flight path information for the aircraft.

2. The aircraft positioning method based on contrail age according to claim 1, characterized in that, The process of fitting an age gradient vector reflecting the direction of change in the wake cloud based on three-dimensional location information and apparent age includes the following steps: The identified wake cloud region is divided into multiple continuous segments along its main axis. Based on the morphological characteristics of each segment of the contrail cloud, and combined with a thermodynamic model, a state age is assigned to each segment; the morphological characteristics include diffusion width, edge sharpness, and optical thickness; Linear fitting is performed based on the spatial location sequence of each segment and its corresponding state age to obtain the age gradient vector of the wake cloud along its extension direction.

3. The aircraft positioning method based on contrail age according to claim 1, characterized in that, The process of identifying the contrail cloud region in a binocular image sequence specifically includes the following steps: The acquired binocular image sequence is input into the YOLOv8-tiny-based wake cloud recognition model. Through the backbone network, multi-scale feature extraction is performed on the binocular images to obtain the initial feature map. By using a feature enhancement layer incorporating the CBAM attention mechanism, channel and spatial attention weighting is applied to the initial feature map to obtain an enhanced feature map focused on the wake cloud region. Multi-scale feature fusion is performed on the enhanced feature map using a feature pyramid network; Based on the fused feature map, the probability of each pixel belonging to the trail cloud category is calculated and output through the segmentation head; Based on the probability, the whet cloud region in the image is identified.

4. The aircraft positioning method based on contrail age according to claim 1, characterized in that, The calculation of the three-dimensional position information of each feature point on the contrail cloud region based on the binocular vision geometric model specifically includes: the following: Distortion and epipolar correction are performed on binocular images; Identify the wake cloud region as the region of interest in the corrected binocular image; Dense stereo matching is performed within the region of interest to generate a disparity map for each pixel. Based on the binocular vision geometric model, the three-dimensional coordinates of each point on the contrail are calculated by combining the camera's intrinsic and extrinsic parameters, and then the average altitude of the contrail relative to the observation station is calculated; wherein, the binocular vision geometric model is expressed as Z=(f×B) / d, where Z is the depth, f is the focal length, B is the baseline distance, and d is the parallax.

5. The aircraft positioning method based on contrail age according to claim 1, characterized in that, The method of predicting the aircraft's three-dimensional position at the current moment by extrapolating motion along the flight path and based on the average motion speed, using the latest observed contrail feature points as a reference, also includes determining whether the aircraft body has left the current image field of view. Specifically, when it is determined that the aircraft body is within the current image field of view, the direct growth vector method is used to locate the aircraft's three-dimensional position at the current moment; when it is determined that the aircraft body has left the current image field of view, the age gradient extrapolation method is used to locate the aircraft's three-dimensional position at the current moment.

6. The aircraft positioning method based on contrail age according to claim 5, characterized in that, When it is determined that the aircraft body is within the current image field of view, the direct vector growth method is used to locate the three-dimensional position of the aircraft at the current moment, specifically including the following steps: Temporal 3D point cloud of the head of the wake cloud is extracted from a series of continuously acquired binocular images; By tracking the displacement of the centroid of the head point cloud in three-dimensional space between consecutive frames, the spatial vector representing the growth direction and rate of the wake cloud can be directly calculated. The direction of the spatial vector representing the growth direction and rate of the contrail cloud is inverted to the real-time heading of the aircraft, and its magnitude is divided by the time interval to invert the real-time speed of the aircraft. The real-time heading of the aircraft is combined with the three-dimensional coordinates of the contrail head, and the real-time three-dimensional position of the aircraft is directly calculated by compensating for a preset geometric offset between the aircraft and the contrail generation point in the opposite direction to the spatial vector; wherein the geometric offset is estimated by a preset model of aircraft type, atmospheric conditions and engine parameters.

7. The aircraft positioning method based on contrail age according to claim 5, characterized in that, When it is determined that the aircraft body has left the current image view, the age gradient extrapolation method is used to locate the three-dimensional position of the aircraft at the current moment, which specifically includes the following steps: The age gradient vector is transformed from the image coordinate system to the geographic coordinate system with the observation station as the origin, and the horizontal heading of the aircraft is initially estimated based on the opposite direction of the projection of the transformed age gradient vector onto the horizontal plane. Select two time intervals Δt=A k -A j The age points j and k; Based on the coordinates and time difference of age points j and k, calculate the historical average speed and vertical speed of the aircraft within the time period Δt; Using the youngest feature point on the contrail as the starting point, the displacement from the observation time corresponding to the starting point to the current time is extrapolated along the initially estimated horizontal heading of the aircraft and based on its historical average speed and vertical speed, to calculate the predicted three-dimensional position of the aircraft at the current time.

8. The aircraft positioning method based on contrail age according to claim 7, characterized in that, The predicted three-dimensional position of the aircraft at the current moment is calculated using the geodesic model. The calculation process is as follows: spherical distance s=V h *t ext / R; latitude current =arcsin(sin(φ new )*cos(s)+cos(φ new )*sin(s)*cos(Heading_final)); Longitude difference Δλ = arctan2(sin(Heading_final)*sin(s)*cos(φ new ), cos(s) - sin(φ new )*sin(φ current )); Longitude λ current = λ new + Δλ; Height H current =H new +V H *t ext ; Where R is the Earth's radius; V h The magnitude of the historical average speed; V H The historical average vertical velocity magnitude; the feature point P with the youngest age on the contrail. new The three-dimensional coordinates are (λ new ,φ new H new The observation time is T. new Heading_final is the final heading; t ext =T now -T new For extrapolation time, T now λ represents the current moment. new For P new longitude, φ new For P new latitude, H new For P new The height.

9. The aircraft positioning method based on contrail age according to claim 1, characterized in that, The process of correlating and fusing predicted three-dimensional position, flight heading, and velocity calculated from different ground observation stations to generate continuous three-dimensional flight path information for the aircraft specifically includes the following steps: Receive target status data from multiple ground observation stations; the target status data includes at least the predicted position, heading, velocity, and the station's estimated confidence level. Based on the similarity of target location and motion state, determine whether the data reported by different observation stations belong to the same aircraft target; For multiple sets of data identified as the same target, a Kalman filter algorithm is used to fuse them, and the fused three-dimensional position, velocity, heading and unique tracking identifier of the aircraft are output. The fusion weight of the data from each observation station is related to the estimated confidence level and the geometric configuration between the station and the target.

10. An aircraft positioning system based on contrail age, characterized in that, include: The acquisition module is used to simultaneously acquire binocular image sequences containing the contrails of the target aircraft at each ground observation station; The contrail age analysis module is used to identify contrail regions in images based on binocular image sequences; and to analyze the apparent age of different segments along the contrail extension direction based on the morphological characteristics of the contrail regions and their physical evolution over time. And calculate the three-dimensional position information of each feature point on the wake cloud region based on the binocular vision geometric model; Based on the three-dimensional location information and apparent age, an age gradient vector reflecting the direction of change in the age of the wake cloud is obtained by fitting. The aircraft position estimation module is used to infer the aircraft's flight heading based on the direction of the age gradient vector, and combine the three-dimensional position information of points with different age characteristics on the contrail cloud region and their corresponding timestamps to infer the aircraft's average speed during the corresponding time period. Based on the latest observed contrail feature points, extrapolate the motion along the flight path and based on the average motion speed to predict the three-dimensional position of the aircraft at the current moment. The generation module is used to correlate and fuse the three-dimensional position, flight heading, and average speed predicted from different ground observation stations to generate continuous three-dimensional flight path information of the aircraft.

Images