A three-dimensional tracking method applied to single-photon laser radar for detecting a dynamic unmanned aerial vehicle point target in a far field
By extending Kalman filtering and adaptive process noise adjustment mechanisms, a six-dimensional state vector is constructed. Combined with air resistance and turning acceleration models, a three-dimensional tracking of point targets on a long-range dynamic UAV by single-photon lidar is realized, solving the problem of insufficient accuracy in traditional methods and achieving a tracking accuracy of 0.4 meters.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-09-29
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies fail when the size of dynamic target imaging is less than one pixel at long distances, rendering detection and tracking algorithms based on shape, texture, or edge features ineffective. Furthermore, traditional lidar systems struggle to achieve high-precision 3D tracking in complex outdoor environments.
An extended Kalman filter framework is adopted, combined with an adaptive process noise adjustment mechanism, to construct a six-dimensional state vector, including the target's three-dimensional position and velocity. Air resistance and turning acceleration models are introduced to perform state prediction and updating, thereby realizing three-dimensional tracking of a single-photon lidar system.
In a complex field environment, continuous and accurate three-dimensional spatial tracking of a dynamic UAV point target 1.4 km away was achieved, with a three-dimensional tracking accuracy of 0.4 meters. This significantly improves the accuracy of traditional methods, simplifies the system structure, and reduces hardware costs.
Smart Images
Figure CN121115030B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of lidar detection and target tracking technology, and in particular relates to a three-dimensional tracking method for detecting point targets of far-field dynamic UAVs using single-photon lidar. Background Technology
[0002] The detection and tracking of moving targets in the airspace is one of the key technologies for maintaining airspace security and strategic early warning systems. With the widespread deployment of new aerial platforms such as drones and low-altitude, slow-moving, and small aircraft, the number of targets in low-altitude airspace has surged and their maneuverability has increased, posing a severe challenge to traditional air situation surveillance systems. Against this backdrop, achieving early detection, accurate identification, and continuous tracking of potentially high-threat targets has become a core task for enhancing airspace protection capabilities.
[0003] Under long-range observation conditions, such targets often appear as "point targets" with extremely small imaging dimensions, making it difficult for traditional imaging or tracking sensors to provide sufficient spatial resolution and detection capabilities. LiDAR, with its extremely high spatial resolution and all-weather operation, demonstrates superior performance in tasks such as target detection and obstacle avoidance. However, as the detection distance increases, laser energy decays rapidly in free space, severely limiting the effective range and detection capabilities of traditional LiDAR.
[0004] The rise of Single-Photon Detection (SPD) technology has provided a breakthrough path for target detection in long-range, low signal-to-noise ratio environments. Combined with Time-Correlated Single-Photon Counting (TCSPC), SPD systems can respond horizontally to a single photon, thereby reconstructing complete 3D information of the target within the field of view through point / line scanning or flash imaging, achieving low-power, long-range detection. Currently, SPD has achieved significant results in long-range detection in static scenes. NASA reported the first photon-counting lidar used in airborne and airborne applications, which achieved a horizontal resolution of 5 m and a vertical accuracy of 10 m at an altitude of 300 km. Li et al. used a coaxial scanning single-photon lidar system to achieve 3D imaging of mountains at 201.5 km using only 0.44 signal photons per pixel.
[0005] With the continuous evolution of technology, the detection and tracking of long-range dynamic targets is gradually becoming a research hotspot. Tachella et al. used a single-photon array detector to detect the contour changes of a dynamic pedestrian at a distance of 320m; Du et al. used a photon counting laser ranging system to simulate the tracking of non-cooperative objects moving beyond line of sight, but the scenario involved only radial motion and has not yet achieved true lateral and three-dimensional tracking; Zhang et al., based on the Gm-APD single-photon array system, achieved real-time two-dimensional planar tracking of UAVs within a distance of 300m through position loop design. However, SPD is still in its infancy in the field of dynamic target detection, with most research limited to the hundred-meter scale, two-dimensional imaging and radial ranging, and lacking the ability to continuously track long-range dynamic "point targets" in three dimensions.
[0006] Furthermore, existing SPD systems mostly rely on passive imaging devices such as CCD and CMOS for collaborative operation, resulting in system complexity and difficulties in optical axis adjustment. Moreover, at long distances, the target imaging size is smaller than a single pixel, rendering detection and tracking algorithms based on shape, texture, or edge features ineffective. Currently, there are no literature or research reports on high-precision two-dimensional or three-dimensional tracking of long-distance, dynamic, point targets in outdoor environments using only single-photon lidar systems. Summary of the Invention
[0007] In view of this, the present invention aims to propose a three-dimensional tracking method for detecting point targets of far-field dynamic UAVs using single-photon lidar, in order to solve the problem that existing detection and tracking algorithms based on shape, texture or edge features fail when the target imaging size is less than one pixel under long-range conditions.
[0008] To achieve the above objectives, the present invention adopts the following technical solution: A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar, the method comprising: Construct a six-dimensional state vector containing the target's three-dimensional position and three-dimensional velocity; The state vector is predicted based on the state transition function that includes air resistance and cornering acceleration. An extended Kalman filter framework is used to update the predicted state and obtain the target state estimate. The extended Kalman filter framework includes an adaptive process noise adjustment mechanism based on innovation feedback, which is used to adjust the process noise covariance matrix online.
[0009] Furthermore, a preferred embodiment is proposed, wherein the six-dimensional state vector includes:
[0010] in, This indicates the three-dimensional spatial position of the target within the field of view of the single-photon lidar array. Distance to the target , The row and column coordinates within the Gm-APD field of view , This represents the velocity component in the corresponding direction.
[0011] Furthermore, a preferred method is proposed, wherein the air resistance is modeled as follows: the air resistance experienced by the target is proportional to the square of the velocity, and the direction is opposite to the velocity vector; the air resistance term in each direction in three-dimensional space is:
[0012] in, Let x be the acceleration due to air resistance in the x-direction. Let be the acceleration due to air resistance in the y-direction. Let be the acceleration due to air resistance in the z-direction. To measure noise, For discrete time step index, For time increments, Let k be the velocity update in the x-direction at time k. Let k be the velocity update in the y-direction at time k. Let k be the velocity update in the z-direction at time k.
[0013] Furthermore, a preferred method is proposed, wherein the turning acceleration is modeled as follows: , , , in, Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let be the velocity component of the target in the xoz plane; This refers to the magnitude of the turning acceleration; Let be the velocity components of the target within the yoz plane; Let be the target's turning acceleration component in the y-direction of the yoz plane; Let be the target's turning acceleration component in the z-direction of the yoz plane; Let x be the component of the target's turning acceleration in the x-direction of the xoy plane; Let be the target's turning acceleration component in the y-direction of the xoy plane; Let be the target velocity component in the xoy plane.
[0014] Furthermore, a preferred embodiment is proposed, wherein the state transition function is: .
[0015] Furthermore, a preferred method is proposed, wherein updating the predicted state includes: The information is calculated by introducing actual parameters based on the observation model:
[0016] in, For observation models, This is the observation matrix of the algorithm; The target state vector predicted by the algorithm; Calculate the observation error covariance matrix:
[0017] in, Let be the covariance matrix of the algorithm's prediction error; Here is the observation noise covariance matrix of the algorithm; update the filter gain based on the error covariance matrix:
[0018] The final state filtering and prediction error covariance update are as follows:
[0019] in, This is the target state vector updated by the algorithm; The new information obtained is the difference between the observation and the prediction. This is the updated error covariance; It is an identity matrix.
[0020] Furthermore, a preferred embodiment is proposed, wherein the adaptive process noise adjustment mechanism includes: Collect velocity component information data within a fixed number of steps in a sliding window:
[0021] in, To adjust the sliding window size, This refers to the information data at time k. For the innovation data at time k-N+1; extract the innovation components of the velocity in three directions:
[0022] in, The information data collected in the x-direction within the sliding window; This refers to the information data collected in the y-direction within the sliding window; The z-direction innovation data collected within the sliding window; the actual covariance of the velocity innovation and the predicted innovation covariance of the theoretical corresponding component are calculated. Based on the ratio of the actual covariance matrix to the predicted innovation covariance of the corresponding theoretical components, the adjustment coefficients in each velocity direction are calculated; and the proportional threshold is designed. Based on the relationship between the adjustment coefficient and the proportional threshold, the base values of the corresponding dimensions in the process noise covariance matrix are dynamically amplified or reduced, including: when This indicates that the actual innovation variance exceeds the theoretical prediction, and the estimation process is insufficiently noisy.
[0023] when This indicates that the actual information is too small and the estimation process is too noisy.
[0024] in, This represents the process noise covariance matrix, used for adaptive updates. This represents the ratio of actual to theoretical innovation variance, used to determine whether process noise needs adjustment. The maximum adjustment factor, As the minimum adjustment factor, and As a shrinkage-sensitive factor, An adaptive threshold is used to determine whether the actual innovation variance exceeds the theoretical prediction, thereby triggering the amplification or reduction of process noise.
[0025] Furthermore, a preferred embodiment is proposed, wherein the adjustment coefficient is:
[0026] in, To calculate the actual information covariance; To calculate the theoretical information covariance.
[0027] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to any one of the above claims.
[0028] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a three-dimensional tracking method for detecting far-field dynamic UAV point targets using a single-photon lidar as described above.
[0029] Compared with the prior art, the beneficial effects of the present invention are: The proposed three-dimensional tracking method, for the first time in a complex outdoor environment, successfully achieved continuous and accurate three-dimensional spatial tracking of a dynamic UAV point target 1.4 kilometers away, relying solely on a single-photon lidar sensor. Its three-dimensional tracking accuracy reached the 0.4-meter level, with average errors of 0.2676 meters and 0.3780 meters. Compared to the traditional Kalman filter method with an average error of approximately 2.75 meters, this represents an order of magnitude improvement in accuracy, solving the fundamental problem of existing technologies being limited to tracking on a scale of hundreds of meters, in a two-dimensional plane, or in a radial one-dimensional manner.
[0030] The proposed 3D tracking method addresses the core challenges of complex and variable background noise in far-field detection, including drastic target maneuvers and intense background noise. It employs an adaptive process noise adjustment mechanism that can perceive system modeling errors and external environmental disturbances online in real time and dynamically adjust filter parameters. This allows the method to maintain stable tracking even when the target performs complex maneuvers such as cross-shaped maneuvers and deep-range flight, effectively avoiding the filtering divergence or sharp decline in accuracy caused by model mismatch in traditional algorithms.
[0031] The three-dimensional tracking method proposed in this invention not only tracks the position of the target, but also achieves accurate and smooth estimation of the target's three-dimensional velocity vector by introducing a refined dynamic model of air resistance and three-plane turning acceleration.
[0032] The three-dimensional tracking method proposed in this invention does not rely on passive imaging devices such as CCD / CMOS for assistance; the entire process of detecting, identifying, and tracking point targets can be completed solely by the single-photon lidar system itself. This simplifies the system structure, avoids the difficulties of multi-axis calibration, and reduces hardware costs and system integration complexity. Attached Figure Description
[0033] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar, as described in this invention. Figure 2 This is a schematic diagram of the intensity imaging results of a distant point target detected by the lidar according to the present invention; Figure 3 This is a schematic diagram of the target tracking trajectory during circular cruise flight as described in this invention; Figure 4 This is a schematic diagram of the position estimation trajectory in three dimensions during circular cruise flight as described in this invention; Figure 5 This is a schematic diagram of the velocity estimation trajectory in three dimensions during circular cruise flight as described in this invention; Figure 6 This is a schematic diagram of the three-dimensional velocity error distribution during circular cruise flight as described in this invention; Figure 7 This is a schematic diagram of the three-dimensional tracking error distribution during circular cruise flight as described in this invention; Figure 8 This is a schematic diagram of the target tracking trajectory during the cross-shaped hovering and deep flight described in this invention; Figure 9 This is a schematic diagram of the target's position estimation trajectory in three dimensions during the cross-shaped hovering and deep flight described in this invention; Figure 10 This is a schematic diagram of the velocity estimation trajectory of a target in three dimensions during cross-shaped maneuvers and deep flight as described in this invention; Figure 11 This is a schematic diagram of the velocity error distribution of the 3D-AEKF algorithm during cross-shaped maneuvers and deep-flight as described in this invention; Figure 12 This is a schematic diagram of the three-dimensional tracking error distribution of the 3D-AEKF algorithm during cross-shaped hovering and deep flight as described in this invention; In the figure, Row represents the row direction of the Gm-APD detector plane; Distance represents the target distance of the measurement result; and Column represents the column direction of the Gm-APD detector plane. Detailed Implementation
[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.
[0035] Implementation Method 1, see Figure 1 This embodiment describes a three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar. The method includes: Construct a six-dimensional state vector containing the target's three-dimensional position and three-dimensional velocity; The state vector is predicted based on the state transition function that includes air resistance and cornering acceleration. An extended Kalman filter framework is used to update the predicted state and obtain the target state estimate. The extended Kalman filter framework includes an adaptive process noise adjustment mechanism based on innovation feedback, which is used to adjust the process noise covariance matrix online.
[0036] This implementation proposes a 3D tracking method for detecting far-field dynamic UAV point targets using single-photon lidar, addressing the technical challenge of high-precision 3D tracking of dynamic point targets (such as UAVs) by single-photon lidar in long-range, low signal-to-noise ratio environments. The technical process includes: estimating and updating the target's 3D position and 3D velocity vectors; introducing the target's turning acceleration and air resistance into the nonlinear state transition function for modeling; and achieving fine estimation of the vectors during recursive filtering. To address the insufficient modeling capabilities of traditional EKF in handling drastic motion state changes and background noise disturbances, this implementation designs an adaptive process noise covariance adjustment mechanism based on innovation feedback. This mechanism analyzes the observation residuals during state updates and dynamically adjusts the process noise covariance matrix Q, achieving online response to system modeling uncertainties and environmental background disturbances, thereby effectively improving tracking accuracy and robustness. The overall process of the 3D tracking method is as follows: Figure 1 As shown.
[0037] Step 1: System Model and State Definition The target's motion state is defined as a six-dimensional state vector x, which contains the target's position and velocity information in three-dimensional space: (1) in, This indicates the three-dimensional spatial position of the target within the field of view of the single-photon lidar array. Distance to the target , The row and column coordinates within the Gm-APD field of view , This represents the velocity component in the corresponding direction.
[0038] The dynamic model of the system is described by nonlinear state transition equations: (2) in, This is the state transition function. State noise, For covariance.
[0039] The observation model is a linear identity mapping, defined as: (3) in, To measure noise, the covariance matrix is: .
[0040] Step 2: State Prediction and Nonlinear Dynamic Modeling In the prediction phase, this implementation not only considers uniform motion but also introduces air resistance and turning acceleration to more realistically simulate the target's dynamic behavior. Specifically, the three-dimensional position update (Euler integral) and the three-dimensional velocity vector update are as follows: (4) in, The composite acceleration consists of two parts: the air resistance component and the cornering acceleration component.
[0041] The air resistance model is as follows: Assume that the air resistance experienced by the target is proportional to the square of its velocity and in the opposite direction to the velocity vector. Influenced by the overall velocity magnitude, its acceleration components on each coordinate axis are: (5) Therefore, the air resistance term in each direction in three-dimensional space is: (6) in, Let x be the acceleration due to air resistance in the x-direction. Let be the acceleration due to air resistance in the y-direction. Let be the acceleration due to air resistance in the z-direction. To measure noise, For discrete time step index, For time increments, Let k be the velocity update in the x-direction at time k. Let k be the velocity update in the y-direction at time k. Let k be the velocity update in the z-direction at time k.
[0042] The turning acceleration in the three-dimensional plane is modeled as follows: To describe the target's maneuvering (such as turning), this embodiment describes it in three planes of space ( xoy , xoz , yoz The acceleration is corrected within the range. For example, xoz In the plane, the velocity components of the target are The direction of acceleration is perpendicular to the current direction of motion. Taking the negative perpendicular vector of its unit vector is: (7) in, This represents the magnitude of the turning acceleration.
[0043] Similarly, it can be obtained in yoz and xoy Cornering acceleration in a plane: (8) (9) in, Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let be the velocity component of the target in the xoz plane; This refers to the magnitude of the turning acceleration; Let be the velocity components of the target within the yoz plane; Let be the target's turning acceleration component in the y-direction of the yoz plane; Let be the target's turning acceleration component in the z-direction of the yoz plane; Let x be the component of the target's turning acceleration in the x-direction of the xoy plane; Let be the target's turning acceleration component in the y-direction of the xoy plane; Let be the target velocity component in the xoy plane.
[0044] Adding the air resistance and turning acceleration components together, we obtain the following formulas for updating the target's velocity in each direction: (10) Therefore, the state transition function This can be represented as a combination of target 3D position update and 3D velocity vector update, with forward integration prediction performed in the prediction step, and can be expressed as: (11) Step 3: Linearize the state transition function, i.e., calculate the Jacobian matrix: Because a nonlinear state transition function is used, linearization is required after each prediction step to apply the Kalman filter framework. This process is achieved by calculating its Jacobian matrix. The Jacobian matrix is constructed to approximate the local linearity of the state near the current estimation point, including: Performing a first-order Taylor expansion on the state transition function, its Jacobian matrix is: (12) The Jacobian matrix consists of two main parts: the partial derivative of position with respect to velocity, and the partial derivative of velocity with respect to state update. According to formula (4), the update of the target position can be obtained and its partial derivative can be calculated: (13) In the other part, the velocity update is calculated based on the chain rule and the principle of fractional differentiation to ensure first-order approximation accuracy. Using formula (10)... Taking the update as an example, let's look at the partial derivatives for the air resistance term and the cornering acceleration term. Let: (14) Calculate using the chain rule We can obtain: (15) Similarly, for right and The partial derivative is: (16) for and The air resistance component is similar and will not be elaborated here.
[0045] And the turning acceleration term, here it is also... For example, the update includes two contributions: exist xoz In a plane, find the ratio of the turning acceleration components. The partial derivative of . Let . but (17) right The partial derivative is: (18) Multiply by The corresponding item is obtained.
[0046] In the xoy plane, find the pair of turning acceleration components. The partial derivative of . Similarly, let . but (19) therefore, The update formula is (20) Taking the partial derivatives with respect to each component yields (twenty one) Similarly, calculate separately and Each partial derivative. Finally, the complete Jacobian matrix is constructed and used to update the prediction error covariance matrix.
[0047] Step 4: Status Update This implementation compares the actual observations and predicted values of the single-photon lidar system to obtain the observation residual (innovation). Then, the innovation covariance matrix is calculated and the Kalman filter gain is updated. Finally, the predicted state and its error covariance are optimally corrected using the gain and innovation to obtain the optimal state estimate and updated error covariance at the current time.
[0048] Among them, the information is calculated by introducing actual parameters based on the observation model (3): (twenty two) This is the observation matrix of the algorithm; The target state vector predicted by the algorithm; The observation error covariance matrix is calculated as follows: (twenty three) in, Let be the covariance matrix of the algorithm's prediction error; The observation noise covariance matrix of the algorithm; Update the filter gain based on the error covariance matrix: (twenty four) The state filtering and prediction error covariance are updated as follows: (25) in, This is the target state vector updated by the algorithm; The new information obtained is the difference between the observation and the prediction. This is the updated error covariance; It is an identity matrix.
[0049] Step 5: Adaptive process noise conditioning: The performance of traditional extended Kalman filters (EKFs) is highly dependent on a preset process noise covariance matrix Q. A fixed value for Q is insufficient to handle sudden changes in target maneuvering or environmental noise. To overcome the performance degradation caused by a fixed process noise covariance, this implementation proposes an adaptive adjustment mechanism for the process noise covariance based on innovation feedback: Collect velocity component information data within a fixed number of steps in a sliding window: (26) in, Sliding window size (only reserved) (New information at this moment) This refers to the information data at time k. For the innovation data at time k-N+1; extract the innovation components of the velocity in three directions: (27) in, The information data collected in the x-direction within the sliding window; This refers to the information data collected in the y-direction within the sliding window; The innovation data in the z-direction collected within the sliding window; the actual statistical covariance of the innovation data is calculated, yielding: (28) Calculate the predicted information covariance of the theoretically corresponding components: (29) in, and Observation matrix Covariance matrix of observation noise The submatrix corresponding to the velocity dimension.
[0050] Adjustment coefficient calculation: Calculate the ratio of actual to theoretical variance for each of the three velocity components, and design a proportional threshold. Define the adjustment coefficient as: (30) in, To calculate the actual information covariance; To calculate the theoretical innovation covariance. When This indicates that the actual innovation variance exceeds the theoretical prediction, and the estimation process is insufficiently noisy. (31) when This indicates that the actual information is too small and the estimation process is too noisy. (32) in, This represents the process noise covariance matrix, used for adaptive updates. This represents the ratio of actual to theoretical innovation variance, used to determine whether process noise needs adjustment. The maximum adjustment factor, As the minimum adjustment factor, and As a shrinkage-sensitive factor, An adaptive threshold is used to determine whether the actual innovation variance exceeds the theoretical prediction, thereby triggering the amplification or reduction of process noise.
[0051] The adjustment coefficient is calculated based on the ratio of the two values. If the actual innovation variance is greater than the theoretical value, it indicates an increase in model prediction uncertainty (such as a target undergoing violent maneuvers). In this case, the value of the corresponding dimension in the process noise covariance matrix Q is increased proportionally to make the filter respond to the observed data more quickly. Conversely, the Q value is decreased accordingly to smooth the estimation results. The adaptive adjustment method based on innovation feedback can dynamically capture changes in noise in the actual system, enabling the filter to maintain robust and accurate state estimation even when the background noise of the single-photon lidar or the target's motion state changes rapidly.
[0052] Existing technologies typically employ simple uniform velocity or uniform acceleration models, which cannot accurately describe the air resistance and turning maneuverability experienced by UAVs in real three-dimensional space. This implementation constructs a state prediction model that more closely resembles physical reality. This model, for the first time in the field of single-photon tracking, introduces an air resistance term proportional to the square of the velocity to simulate the aerodynamic effects experienced by the aircraft. Simultaneously, it models turning acceleration in three spatial planes (xoy, xoz, yoz), with the direction always perpendicular to the current direction of motion. This composite dynamics model fundamentally improves the accuracy of predicting the motion state of targets, especially UAVs performing high-speed, high-maneuverability flight, in the short term.
[0053] Traditional Extended Kalman Filters (EKFs) employ a fixed process noise covariance matrix Q, which is insufficient to handle sudden target maneuvers or abrupt changes in environmental noise. This implementation proposes an adaptive adjustment mechanism based on innovation (observation residual) feedback. This mechanism analyzes the actual statistical characteristics (actual covariance) of the velocity innovation components within a sliding window in real time and compares them with theoretical predictions to determine whether the current process noise setting is "insufficient" or "excessive." Subsequently, according to a preset adjustment law, the values of the corresponding dimensions in the process noise covariance matrix are dynamically amplified or reduced. This approach transforms the static filter into a dynamic filter with online learning capabilities, automatically adapting to unknown changes in the target's motion state and the external environment.
[0054] Existing solutions largely rely on imaging sensors to provide target information, or are only validated at extremely short distances under ideal conditions. This implementation constructs a complete technology chain from single-photon level signal detection and point cloud reconstruction to final 3D tracking, all based on echo data from single-photon lidar. Addressing the textureless and shapeless nature of "point targets," this implementation abandons all traditional image feature-based methods, instead relying entirely on the target point's 3D spatial coordinates and motion trajectory sequence, achieving stable tracking through the aforementioned state estimation and filtering algorithms.
[0055] Implementation Method 2: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar, as described in Implementation Method 1.
[0056] Implementation Method 3: A computer-readable storage medium as described in this embodiment stores a computer program, which, when executed by a processor, performs the steps of a three-dimensional tracking method for detecting far-field dynamic UAV point targets using a single-photon lidar, as described above.
[0057] Implementation Method Four, see below Figures 2 to 12 This embodiment describes a specific experiment based on the three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar described in Embodiment 1. This further verifies the effectiveness of 3D-AEKF applied to single-photon lidar for detecting far-field, dynamic, point targets in real, complex environments.
[0058] In this embodiment, the Gm-APD single-photon lidar detection system was placed in Yanqing District, Beijing for field experiments. The experiment was conducted at 14:00, under clear skies with a visibility of 31 km and a background illuminance of 113,500 Lux. The system output wavelength was 1064 nm, the repetition frequency was 2.5 kHz, the output power was 6.80 W, the field of view was 1°, and the array pixel count was 64×64.
[0059] The target intensity image is obtained by reconstructing 500 frames based on the peak method, as shown below. Figure 2 As shown, the point with the highest intensity represents the location of the UAV. A distance-gating technique is used to filter out the mountain background, and intensity and distance masking is used to reconstruct the 3D point cloud. 3D-AEKF and KF methods are used for tracking, and the 3D tracking trajectories for the two flight states are shown below. Figures 3 to 5 , Figures 8 to 10 As shown.
[0060] Flight State 1: During circular cruise flight, using Euclidean distance as the tracking error, the distribution of the algorithm's tracking error and velocity estimation error is as follows: Figure 6 and Figure 7 As shown; Flight state 2: During cross-shaped maneuvering and deep flight, using Euclidean distance as the tracking error, the distribution of the algorithm's tracking error and velocity estimation error is obtained as follows. Figure 11 and Figure 12 As shown in Table 1, under both flight conditions, 3D-AEKF can stably estimate the target's three-dimensional trajectory, with average tracking errors of 0.2676 m and 0.3780 m, respectively, significantly outperforming the traditional KF method (2.7505 m and 2.7494 m, respectively). The root mean square (RMSE) of the tracking error in the three-dimensional directions is also significantly lower than the corresponding distribution of the KF method. Especially in the complex flight of state 2, 3D-AEKF maintains a low position error, demonstrating good dynamic adaptability and strong robustness. In summary, under conditions of strong background light and complex terrain interference, 3D-AEKF can still achieve long-range tracking with an accuracy of 0.4 m at 1.4 km, significantly outperforming the divergence trend exhibited by the traditional KF method in such scenarios.
[0061] Table 1. Algorithm position error distribution under two flight conditions
[0062] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0063] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the published pending claims.
Claims
1. A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar, characterized in that, The method includes: Construct a six-dimensional state vector containing the target's three-dimensional position and three-dimensional velocity; The state vector is predicted based on the state transition function that includes air resistance and cornering acceleration. The predicted state is updated using an extended Kalman filter framework to obtain the target state estimate. The extended Kalman filter framework includes an adaptive process noise adjustment mechanism based on innovation feedback, which is used to adjust the process noise covariance matrix online. The state transition function is: ； in, Let k be the velocity update in the x-direction at time k. Let k be the velocity update in the y-direction at time k. Let k be the velocity update in the z-direction at time k. For discrete time step index, For time increments, Let be the target's turning acceleration component in the y-direction of the yoz plane; Let be the target's turning acceleration component in the z-direction of the yoz plane; Let x be the component of the target's turning acceleration in the x-direction of the xoy plane; Let be the target's turning acceleration component in the y-direction of the xoy plane; Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let x be the target's turning acceleration component in the x-direction of the xoz plane; For measuring noise.
2. The three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to claim 1, characterized in that, The six-dimensional state vector includes: in, This indicates the three-dimensional spatial position of the target within the field of view of the single-photon lidar array. Distance to the target , The row and column coordinates within the Gm-APD field of view , This represents the velocity component in the corresponding direction.
3. A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to claim 2, characterized in that, The air resistance model is as follows: the air resistance experienced by the target is proportional to the square of the velocity, and its direction is opposite to the velocity vector; the air resistance term in each direction in three-dimensional space is: in, Let x be the acceleration due to air resistance in the x-direction. Let be the acceleration due to air resistance in the y-direction. Let be the acceleration due to air resistance in the z-direction. To measure noise, For discrete time step index, For time increments, Let k be the velocity update in the x-direction at time k. Let k be the velocity update in the y-direction at time k. Let k be the velocity update in the z-direction at time k.
4. A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to claim 3, characterized in that, The turning acceleration is modeled as follows: ， ， ， in, Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let x be the target's turning acceleration component in the x-direction of the xoz plane; Let be the velocity component of the target in the xoz plane; This refers to the magnitude of the turning acceleration; Let be the velocity components of the target within the yoz plane; Let be the target's turning acceleration component in the y-direction of the yoz plane; Let be the target's turning acceleration component in the z-direction of the yoz plane; Let x be the component of the target's turning acceleration in the x-direction of the xoy plane; Let be the target's turning acceleration component in the y-direction of the xoy plane; Let be the target velocity component in the xoy plane.
5. A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to claim 1, characterized in that, The process of updating the predicted state includes: The information is calculated by introducing actual parameters based on the observation model: in, For observation models, This is the observation matrix of the algorithm; The target state vector predicted by the algorithm; The observation error covariance matrix is calculated as follows: in, Let be the covariance matrix of the algorithm's prediction error; The observation noise covariance matrix of the algorithm; Update the filter gain based on the error covariance matrix: The final state filtering and prediction error covariance update are as follows: in, This is the target state vector updated by the algorithm; For the calculated new information; This is the updated error covariance; It is an identity matrix.
6. A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to claim 1, characterized in that, The adaptive process noise adjustment mechanism includes: Collect velocity component information data within a fixed number of steps in a sliding window: in, To adjust the sliding window size, This refers to the information data at time k. This refers to the new information at time k-N+1; Extraction speed in three directions of information components: in, The information data collected in the x-direction within the sliding window; This refers to the information data collected in the y-direction within the sliding window; The information data collected in the z-direction within the sliding window; The actual covariance of the computational speed innovation and the predicted innovation covariance of the theoretical corresponding component; Based on the ratio of the actual covariance matrix to the predicted innovation covariance of the corresponding theoretical components, the adjustment coefficients in each velocity direction are calculated; and the proportional threshold is designed. Based on the relationship between the adjustment coefficient and the proportional threshold, the base values of the corresponding dimensions in the process noise covariance matrix are dynamically amplified or reduced, including: when This indicates that the actual innovation variance exceeds the theoretical prediction, and the estimation process is insufficiently noisy. when This indicates that the actual information is too small and the estimation process is too noisy. in, Represents the process noise covariance matrix. This represents the ratio of actual to theoretical innovation variance. The maximum adjustment factor, As the minimum adjustment factor, and As a shrinkage-sensitive factor, This is an adaptive threshold.
7. A three-dimensional tracking method for detecting far-field dynamic UAV point targets using single-photon lidar according to claim 6, characterized in that, The adjustment coefficient is: in, To calculate the actual information covariance; To calculate the theoretical information covariance.
8. A computer device, characterized in that: The device includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a three-dimensional tracking method for detecting far-field dynamic UAV point targets using a single-photon lidar, as described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a three-dimensional tracking method for detecting far-field dynamic UAV point targets using a single-photon lidar, as described in any one of claims 1-7.