Multi-target sensing and tracking method for detecting unmanned aerial vehicle swarm through single-photon laser radar

Abstract

The invention discloses a multi-target sensing and tracking method for detecting an unmanned aerial vehicle swarm through a single-photon laser radar, and belongs to the field of photoelectric detection and signal processing. The problem that a traditional image reconstruction algorithm and a tracking algorithm are difficult to adapt to ultra-low signal-to-noise ratio, high dynamic and multi-target complex sensing requirements is solved. The method comprises the following steps: carrying out systematic prepeak noise self-adaptive suppression on global photon arrival time data acquired by a single-photon laser radar; performing photon event physical statistical modeling and background noise suppression based on the suppressed data to obtain an enhanced residual signal; performing space-time consistency enhancement processing on the enhanced residual signal to realize extraction and reconstruction of a multi-target area; based on the extracted photon events in the multi-target area, performing trajectory fitting and prediction by adopting a Bayesian weighted RANSAC (Random Sample Consensus) fitting method; and designing a target tracking algorithm based on scale adaptive updating and template matching of an optical flow method and pyramid search assistance, and carrying out robust tracking on multiple targets. The method is used in target tracking field.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of photoelectric detection and signal processing, and particularly relates to a multi-target perception and tracking method for single-photon laser radar detection of a UAV swarm. BACKGROUND

[0002] With the rapid development and wide application of UAV technology, however, a single UAV is difficult to independently complete a large-scale, highly coordinated complex task due to its load capacity and task execution complexity. In recent years, the progress of distributed control and autonomous cooperation technology enables UAVs to form an efficient cluster system through local communication and cooperation strategies, i.e., a UAV swarm. This swarm combat mode improves task efficiency while also posing unprecedented challenges to existing air defense and countermeasure technologies.

[0003] Currently, radar technology is the mainstream means of air defense and surveillance. Among them, multiple-input multiple-output (MIMO) radar and multi-beam phased array radar have made significant breakthroughs in resolution relying on their spatial degrees of freedom and virtual array technology. Some systems using coprime arrays have achieved early warning of UAV swarms at a distance of kilometers. However, radar technology has inherent limitations in dealing with low-altitude, slow-speed, and small targets: low-altitude complex environments easily cause distortion and distortion of radar echoes, while the high mobility of UAV swarms causes serious range migration and Doppler shift phenomena, exacerbating the aliasing between multi-target echoes, thereby greatly limiting its accurate detection and stable tracking performance in swarm scenarios. Radio frequency monitoring technology is of concern due to its low cost and long action distance, but its performance is heavily dependent on prior knowledge of target signals, and is easily disturbed by radio frequency noise and multipath effects in complex electromagnetic environments, making it difficult to achieve stable and continuous tracking of swarm targets. As an effective auxiliary identification means, photoelectric detection systems, among which infrared detection relies on the thermal radiation characteristics of targets, and visible light detection can achieve target identification and tracking in combination with deep learning algorithms. However, as a passive perception method, photoelectric detection can only provide two-dimensional information of the target, and its performance is easily affected by environmental lighting conditions and severe weather such as fog, haze, rain, and snow, greatly restricting its action distance and reliability in actual environments.

[0004] With the development of semiconductor devices and high-speed electronic timing technology, single-photon lidar (SPL) has become an important means for long-range and low-reflectivity target detection due to its single-photon-level detection sensitivity and picosecond-level time resolution. Studies have shown that SPL systems can successfully achieve three-dimensional imaging of a mountain 201.5 kilometers away under the extreme condition of an average of only 0.44 photoelectrons per pixel, fully demonstrating its great potential in the field of ultra-long-range perception. At the same time, the scale of avalanche photodiode (APD) arrays is constantly expanding, enabling SPL systems to obtain large field-of-view three-dimensional point cloud data without mechanical scanning, thereby significantly improving the spatiotemporal resolution and system response capability of dynamic scenes. Domestic and foreign research institutions have made preliminary progress in this field. For example, the China Electronics Technology Group has achieved real-time tracking and measurement of a single unmanned aerial vehicle (UAV) 0.3 kilometers away based on a 64x64 pixel Gm-APD SPL and a high-precision encoder; the Swedish Defense Research Institute has achieved profile measurement of a stationary UAV 1.3 kilometers away based on SPL under the condition of less than 1 photoelectron per pixel; and Harbin Institute of Technology has achieved real-time tracking of a dynamic UAV 3 kilometers away based on a Gm-APD SPL.

[0005] Although the above studies have proven the feasibility of SPL in detecting single static or slow dynamic UAVs, the technical difficulty increases exponentially for the detection and tracking of high-dynamic and high-density UAV swarms. In the high-dynamic scenario of swarm formation flight, complex situations such as mutual occlusion, trajectory intersection, and image projection overlap frequently occur between UAVs, which can easily cause target loss, false association, or trajectory drift in tracking algorithms. The detection essence of SPL systems relies on the statistics of limited photon events, which are used to complete target reconstruction by constructing a histogram of photon arrival time frequency distribution. The high-dynamic characteristics of the swarm target make the statistical amount of effective photon events in the reflected echo extremely sparse, making the target extraction process face the severe challenge of extremely low signal-to-noise ratio (SNR<-10 dB). The effective photon events returned by the target are extremely sparse and random, and are easily overwhelmed by strong background noise, resulting in highly incomplete and discontinuous inter-frame target observation information. Traditional image reconstruction algorithms and tracking algorithms usually rely on stable target area reconstruction and complete observation frame data for modeling and matching, making it difficult to adapt to the complex perception requirements of ultra-low SNR, high-dynamic, and multi-target. SUMMARY

[0006] Therefore, the present application aims to provide a multi-target perception and tracking method for single-photon lidar detection of UAV swarms to solve the problem that traditional image reconstruction algorithms and tracking algorithms usually rely on stable target area reconstruction and complete observation frame data for modeling and matching, making it difficult to adapt to the complex perception requirements of ultra-low SNR, high-dynamic, and multi-target.

[0007] To achieve the above object, the present application adopts the following technical solutions: A single-photon laser radar detects a multi-target perception and tracking method of a UAV swarm, the method comprising: step S1: statistics is performed on global photon arrival time data collected by a single-photon laser radar, a time frequency distribution histogram is constructed, and wavelet denoising processing is performed on the time frequency distribution histogram; step S2: a front-end noise tail decay trend in the time frequency distribution histogram is identified and fitted, a first derivative of the decay trend is calculated to determine a demarcation time point of the front peak noise and the stable background area, and non-physical noise data before the demarcation time point is removed; step S3: background noise exponential decay modeling is performed on the time frequency distribution histogram from which the non-physical noise data is removed, and background parameters are fitted by a linear least squares method; a residual of the global trigger frequency and the background fitting frequency is calculated; step S4: wavelet transform, soft threshold compression, wavelet inverse transform reconstruction and median filtering are performed on the residual to obtain an enhanced smooth residual signal; step S5: candidate pulse peaks are detected from the smooth residual signal, and a high confidence pulse peak set is obtained based on a half-width constraint and a structural symmetry constraint; step S6: photon events in a time window corresponding to each high confidence pulse peak are mapped to a pixel space, a normalized statistical deviation of each pixel is calculated, and a candidate target area is extracted based on spatial connectivity analysis; step S7: based on photon arrival time data of each candidate target area, a trajectory fitting and prediction is performed by using a Bayesian weighted RANSAC algorithm to obtain a motion model and a predicted arrival time of the target, and a target range image is reconstructed accordingly; step S8: for each candidate target area segmented, a local optical flow field of the target in the field of view is calculated based on a Farneback dense optical flow method, a target area center position is updated by weighted average, a scale adaptive update is introduced, and a tracking window size is dynamically adjusted according to an optical flow amplitude change.

[0008] Based on the same inventive concept, the present application further provides a computer device comprising a memory and a processor, the memory storing a computer program, and when the processor runs the computer program stored in the memory, the processor executes the multi-target perception and tracking method of the UAV swarm detected by the single-photon laser radar according to any one of the above.

[0009] Based on the same inventive concept, the present application further provides a computer readable storage medium, the computer readable storage medium storing a computer program, and when the computer program is run by a processor, the steps of the multi-target perception and tracking method of the UAV swarm detected by the single-photon laser radar according to any one of the above are executed.

[0010] Compared with the prior art, the present application has the following beneficial effects: Different from traditional methods which directly image or simply threshold the raw photon counts, the present invention designs a two-stage preprocessing mechanism of front-peak noise adaptive suppression and residual enhancement. First, the front-peak noise caused by avalanche breakdown overflow, post-pulse and other non-physical mechanisms is identified and modeled, and based on its derivative characteristics, the noise area is accurately located and removed, thus purifying the data environment from the source and restoring the Poisson statistical consistency of photon events. Second, the frequency residual is enhanced by combining log-domain background fitting, wavelet multi-scale denoising and median filtering, effectively suppressing statistical fluctuations and pseudo-pulses, while at the same time preserving the symmetry structure and morphological integrity of real target pulses to the greatest extent, which is not involved in traditional methods.

[0011] Traditional methods rely only on time domain or spatial domain information for target judgment under low signal-to-noise ratio, which is extremely vulnerable. The present invention proposes a separation principle of coupling third-order pulse peak detection and spatial clustering: first-order time domain preliminary screening to extract candidate pulse peaks in the enhanced residual signal; second-order structure screening: introducing half-width constraint and structural symmetry constraint to physically exclude noise spikes or diffuse background that do not conform to the characteristics of laser echoes, obtaining high-confidence time pulses; third-order spatial verification: mapping each time pulse to the spatial domain, calculating the spatial standardized statistical deviation of photon events within its corresponding time window, and extracting candidate target regions that show significant aggregation in both time and space dimensions through connected component analysis. This closed-loop mechanism of time domain detection and spatial verification ensures that only signals consistent in both time and space are determined as real targets, fundamentally solving the contradiction between false alarm and missed detection under extremely low signal-to-noise ratio.

[0012] In the trajectory fitting stage, facing the challenge of sparse target photon events mixed with a large number of background noise outliers, the present invention discards traditional methods such as least squares which are sensitive to outliers, and introduces a Bayesian weighted RANSAC algorithm. Through the random sampling and consistency verification mechanism of RANSAC, the inlier set of the target motion model can be robustly estimated, effectively resisting the interference of noise and outliers. At the same time, the introduction of Bayesian prior can guide the sampling process, improving the fitting efficiency and stability. This makes it possible to obtain accurate target motion parameters and prediction values even in a strong noise background.

[0013] In the target tracking stage, the application breaks through the limitation of the traditional fixed scale tracker, and designs a double-mode adaptive tracking mechanism dominated by Farneback dense optical flow and based on BANRAC distance guided template matching. By using the dense displacement information provided by the optical flow field, the target position is updated by Gaussian weighted average, and the target imaging scale is adaptively estimated and updated according to the optical flow amplitude change, which effectively compensates the scale distortion caused by longitudinal motion and maintains the geometric consistency of tracking. When the target is seriously occluded, moves quickly or the pixel is too sparse to cause the optical flow estimation to fail, the backup mechanism is automatically triggered. The mechanism uses the predicted distance information provided by BANRAC fitting to constrain the search range of template matching, and combines with the pyramid layered search strategy, which greatly improves the success rate and efficiency of recovering tracking in extreme cases. Experimental data show that the method of the application pushes the lower limit of the single-photon laser radar target detection and segmentation signal-to-noise ratio of the unmanned aerial vehicle to below-14dB. Under the same conditions, the traditional matching filter-11.34dB, kernel density estimation-10.77dB have all failed. The application successfully solves the target aliasing problem caused by occlusion, trajectory intersection and projection overlap in the bee swarm scene. The embodiment proves that the algorithm can clearly distinguish and separate multiple adjacent targets with a longitudinal distance difference of ≥4.95m, corresponding to a time interval of 33ns, and can effectively handle double-target and even triple-target intersection scenes in the middle and far field. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The schematic diagram of the multi-target dynamic perception architecture of the ultra-low signal-to-noise ratio photon event modeling and structured segmentation described in the application; Figure 2 The flowchart of the scale adaptive target tracking algorithm described in the application; Figure 3 The schematic diagram of the multi-frame intensity image reconstructed based on MF described in the application; Figure 4 The schematic diagram of the multi-frame intensity image reconstructed based on PKF described in the application; Figure 5 The schematic diagram of the multi-frame intensity image reconstructed based on MLE described in the application; Figure 6 The schematic diagram of the multi-frame intensity image reconstructed based on KDE described in the application; Figure 7 The schematic diagram of the multi-frame intensity image reconstructed based on the method described in the application; Figure 8 The schematic diagram of the frequency distribution histogram statistical fitting and residual distribution described in the application, wherein Photon arrival time represents the photon arrival time corresponding to the triggered photon event, and Trigger frequency represents the trigger frequency; Figure 9 Residual pulse peak distribution determined by two-order fitting according to the application; Figure 10 Frequency peak distribution after three-order spatial domain screening according to the application; Figure 11 Target candidate connected region calibration diagram according to the application; Figure 12 Photon arrival time fitting and distance reconstruction of target candidate region 1 according to the application; Figure 13 Photon arrival time fitting and distance reconstruction of target candidate region 2 according to the application; Figure 14 Two target corresponding frequency distribution marking and intensity image reconstruction result according to the application; Figure 15 Reconstruction process of the method according to the application when SNR=-14.0826dB; Figure 16 Connected domain detection result of the method according to the application when SNR=-14.0826dB; Figure 17 Reconstruction result after target 1 segmentation according to the application when SNR=-14.0826dB; Figure 18 Reconstruction result after target 2 segmentation according to the application when SNR=-14.0826dB; Figure 19 Far field reconstruction result based on matching filtering according to the application; Figure 20 Residual and frequency distribution histogram position distribution of two target corresponding time domain in far field according to the application; Figure 21 Reconstruction result of segmented target 1 according to the application; Figure 22 Reconstruction result of segmented target 2 according to the application; Figure 23 Photon arrival time fitting result distribution of two targets according to the application; Figure 24 Reconstruction result based on matching filtering algorithm according to the application; Figure 25 Residual and frequency distribution histogram position distribution of two target corresponding time domain in medium distance according to the application; Figure 26 Reconstruction result of two targets in medium distance according to the application; Figure 27A schematic diagram of a photon arrival time fitting distribution of two targets at a medium distance according to the present application; Figure 28 A schematic diagram of a reconstruction result of multiple unmanned aerial vehicle targets in a near field according to the present application based on matched filtering; Figure 29 A schematic diagram of a residual and frequency distribution histogram position distribution of two targets in a near field according to the present application corresponding to a time domain; Figure 30 A schematic diagram of a reconstruction result of two targets in a near field according to the present application; Figure 31 A schematic diagram of a photon arrival time fitting distribution of two targets in a near field according to the present application; Figure 32 A schematic diagram of a residual resolution of three unmanned aerial vehicles according to the present application; Figure 33 A schematic diagram of a segmentation result of three unmanned aerial vehicle targets according to the present application; Figure 34 A schematic diagram of a photon event fitting of three unmanned aerial vehicle targets according to the present application corresponding to a time domain; Figure 35 A schematic diagram of a residual position distribution of three unmanned aerial vehicle targets according to the present application; Figure 36 A schematic diagram of a segmentation result of three unmanned aerial vehicle targets according to the present application; Figure 37 A schematic diagram of a tracking result of a far field target 1 according to the present application based on a Farneback dense optical flow algorithm; Figure 38 A schematic diagram of a tracking result of a far field target 2 according to the present application based on a Farneback dense optical flow algorithm; Figure 39 A schematic diagram of a tracking trajectory and error distribution curve of a target 1 according to the present application; Figure 40 A schematic diagram of a tracking trajectory and error distribution curve of a target 2 according to the present application. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict, and the described embodiments are only some of the embodiments of the present application, not all of the embodiments.

[0016] Embodiment one, see Figures 1-2 This embodiment is described. The method for detecting unmanned aerial vehicle swarms by a single-photon lidar and perceiving and tracking multiple targets according to the present application comprises: Step S1: statistics is performed on global photon arrival time data collected by the single-photon lidar, a time frequency distribution histogram is constructed, and wavelet denoising is performed on the time frequency distribution histogram; Step S2: the front-end noise tail decay trend in the time frequency distribution histogram is identified and fitted, the first derivative of the decay trend is calculated to determine the demarcation time point of the front peak noise and the stable background area, and the non-physical noise data before the demarcation time point is removed; Step S3: background noise exponential decay modeling is performed on the time frequency distribution histogram after removing the non-physical noise data, and the background parameters are fitted by the linear least squares method; the residual of the global trigger frequency and the background fitting frequency is calculated; Step S4: wavelet transform, soft threshold compression, wavelet inverse transform reconstruction and median filtering are performed on the residual to obtain an enhanced smooth residual signal; Step S5: candidate pulse peaks are detected from the smooth residual signal, and high-confidence pulse peak sets are obtained based on half-height width constraints and structural symmetry constraints; Step S6: the photon events in the time window corresponding to each high-confidence pulse peak are mapped to the pixel space, the normalized statistical deviation of each pixel is calculated, and the candidate target region is extracted based on spatial connectivity analysis; Step S7: based on the photon arrival time data of each candidate target region, the Bayesian weighted RANSAC algorithm is used for trajectory fitting and prediction to obtain the motion model and predicted arrival time of the target, and the target range image is reconstructed accordingly; Step S8: for each candidate target region segmented, the local optical flow field of the target in the field of view is calculated based on the Farneback dense optical flow method, the target region center position is updated by weighted average, and scale adaptive update is introduced, and the tracking window size is dynamically adjusted according to the change of optical flow amplitude.

[0017] Specifically, the embodiment proposes a multi-target perception and tracking method for detecting a drone swarm by a single-photon lidar, which includes two parts: multi-target dynamic perception based on ultra-low signal-to-noise ratio photon event modeling and structured segmentation, and target tracking based on scale adaptive update. In the detection of a drone swarm at a long distance, the scattering cross-section of a drone is extremely small, and the number of echo photons is extremely limited. At the same time, there is strong competition between background photons, complex scene reflection, system noise and target echo, which leads to a significant decrease in the effective trigger probability of a single-photon lidar (SPL), and severely limits the detection performance and dynamic response. Under the conditions of an extremely low signal-to-noise ratio (SNR < -10 dB) and limited reconstruction frame number, the inter-frame reconstruction stability of a single target is impaired, and the problem is more prominent in a multi-target scene. Traditional threshold methods, deep learning detection or conventional image enhancement strategies are difficult to achieve high-confidence target separation and trajectory recovery. In addition, the occlusion, trajectory intersection and projection overlap among multiple drones further exacerbate the difficulty of segmentation and dynamic modeling. To solve the above problems, the embodiment proposes a multi-target dynamic perception architecture based on ultra-low signal-to-noise ratio photon event modeling and structured segmentation, as shown in the schematic diagram of the architecture Figure 1 . Specifically, it includes: Step 1: Systematic pre-peak noise adaptive suppression SPL target detection relies on the high-frequency trigger statistics and frequency modeling of the gated-in photon events. However, in the first few time channels after the gate is opened, obvious abnormal peak values of non-physical noise are observed, which seriously interferes with the extraction of the true signal and limits the effective imaging and dynamic range of the system. The causes of this pre-peak noise are complex, mainly including the following aspects: avalanche breakdown overshoot and post-pulse effect, system gate delay response, APD response "memory effect", charge retention effect. These factors together cause the pre-peak noise to be abnormally prominent, and the statistical characteristics do not conform to the standard Poisson process, showing asymmetry, tail delay, and seriously damaging the time-intensity distribution consistency of the echo signal. Therefore, we propose a pre-peak detection and truncation method based on frequency modeling and derivative analysis to identify and truncate the tail region of the pre-peak, thereby suppressing systematic noise interference and restoring the distribution consistency of the true photon events. It contains the following three contents: The global photon arrival time data collected by the single-photon lidar is statistically analyzed, and the time frequency distribution histogram is constructed, and the time frequency distribution histogram is processed by wavelet denoising; The photon events returned by each pixel of SPL are composed of photon arrival time, frame index, count number, etc. Set the echo data as a three-dimensional data tensor . Among them, represents the pixel coordinates; is the imaging resolution; is the arrival time of the returned photons; is the set gate width.

[0018] The global photon events of all pixels triggered are unfolded in the time dimension, denoted as the global photon event set: , where N is the total number of pixel triggered photon events. The histogram statistics of with a reconstruction frame number of T=400 are obtained to obtain the global time frequency distribution vector: , where is the Dirac impulse function. The wavelet denoising method is introduced to smooth the frequency distribution histogram: .

[0019] Identify and fit the decay trend of the tail of the pre-peak noise in the time frequency distribution histogram; The overall trend of the pre-peak noise can be approximately modeled by an exponential decay function, although its source is not a single noise mechanism. The global maximum point in can be regarded as the index of the main peak of the in-gate noise . In this embodiment, from , the right tail sequence is extracted for exponential fitting: to fit the main peak height of the photon event trigger frequency; to fit the exponential decay constant; to represent the noise stable baseline, to fit the main peak index of the in-gate noise, x is the time bin index corresponding to the triggered photon event to fit the initial offset.

[0020] The demarcation time point of the front peak noise and the stable background area is determined by calculating the first derivative of the decay trend, and the non-physical noise data before the demarcation time point is removed; Since the front peak noise trigger frequency presents the characteristics of steep rise to peak and rapid decay; Once it enters the interval mainly dominated by background noise and target echo, its change trend tends to be flat. Therefore, the first derivative is calculated from the fitting curve to describe the rate of frequency change: , wherein, is the fitting value corresponding to the next triggered photon event time index, is the fitting value corresponding to the current triggered photon event time index, is the time bin width, is the time index corresponding to the next triggered photon event, is the time index corresponding to the current triggered photon event.

[0021] And according to this, the position coordinates of the first derivative entering the slow change area are found, which are regarded as the demarcation of the transition from the systematic non-steady state response to the stable background, denoted as . And all time channels will be marked as non-physical front peaks and removed in subsequent target detection and reconstruction: .

[0022] Step 2: Photon event modeling and background noise suppression Due to the extreme sparsity of target photon events, even after front-end noise suppression, the target echo signal is still submerged in a large amount of background noise, making target detection under the condition of ultra-low signal-to-noise ratio (SNR<-10 dB) extremely challenging. To achieve the timing separation and dynamic prediction of multiple target photon events, the present embodiment proposes a multi-target dynamic photon event detection and prediction method driven by structure enhancement-time domain statistical significance. This method comprehensively utilizes physical modeling, statistical enhancement and structure consistency constraints to achieve multi-target detection and continuous trajectory fitting under sparse extreme conditions. Specifically, it includes the following contents: Step 21: Physical statistical modeling of photon events After adaptive gating, the frequency distribution histogram , under ideal conditions, will present multiple isolated peaks, corresponding to the echo responses of multiple targets. Its statistical behavior conforms to the non-homogeneous Poisson process: , is the photon arrival rate per unit time, and the background noise term is the target echo term is composed of: .

[0023] Step 22: Background noise exponential decay modeling is performed on the time-frequency distribution histogram of the non-physical noise data removed, and the background parameters are fitted by linear least squares. After suppressing abnormal front-end noise, the background count frequency still presents a statistically consistent tail decay behavior on the time axis. Therefore, the following exponential model is used in this embodiment to model the frequency of background noise trigger events: .

[0024] Taking the natural logarithm on the frequency distribution histogram , we have: , where is the random noise.

[0025] This embodiment uses linear least squares to fit the background parameters in the background dominant interval, and determines the background noise trigger photon event frequency distribution curve: .

[0026] Step 23: Calculate the residual of the global trigger frequency and the background fitted frequency; perform wavelet transform, soft threshold compression, inverse wavelet transform reconstruction, and median filtering on the residual to obtain an enhanced smooth residual signal; define the residual of the trigger frequency as the difference between the global trigger photon event frequency and the background noise trigger photon event frequency: .

[0027] Under low signal-to-noise ratio conditions, the residual signal contains a large number of isolated spikes, pulse glitches, or short-time mutations. The sources of these pseudo-pulses include: statistical fluctuations in the background fitting residual (Poisson fluctuations are more intense at low count rates); system electronic noise or time jitter of photon detection; accidental dark count responses of the detector in non-target time windows.

[0028] Pseudo-pulses appear as sharp, random, asymmetric, and non-structural mutations. If peak detection is performed directly, it will seriously affect subsequent image reconstruction. Therefore, this embodiment designs an enhanced process of wavelet transform and median filtering to suppress glitches while preserving the symmetric structure of real target pulses, including: performing three-layer Daubechies-4 discrete wavelet transform (DWT) on the residual signal to obtain high-frequency detail coefficients: , where is the 3rd layer low-frequency approximation coefficient, to are the corresponding detail high-frequency coefficients.

[0029] The Daubechies-4 wavelet basis is selected because of its good localization of signal discontinuities and its ability to distinguish between pulses and noise.

[0030] For each layer of high-frequency coefficients The soft threshold function is applied:

[0031] where the threshold , is the median absolute deviation estimate (MAD) of . The processed and to are inverse wavelet transformed to to , which are the three processed detail coefficients; the denoised residual is reconstructed as .

[0032] This process preserves the target pulse shape while suppressing multi-scale noise interference. The two-stage median filtering (Median Filtering) performs a sliding window median filtering on to remove residual spikes and obtain an enhanced smoothed residual signal , which is: where is the denoised reconstructed residual of the adjacent previous time index, and is the denoised reconstructed residual of the adjacent next time index.

[0033] Step 3: Detect candidate pulse peaks from the smoothed residual signal and filter based on half-width constraint and structural symmetry constraint to obtain a high-confidence pulse peak set; map the photon events in the time window corresponding to each high-confidence pulse peak to the pixel space, calculate the standardized statistical deviation of each pixel, and extract the candidate target region based on spatial connectivity analysis; identify possible target pulses from the enhanced smoothed residual signal , which requires two-stage screening combining peak value characteristics in the time domain and signal structural features: based on the initial screening mechanism of the time domain peak value, extract candidate pulses with local features; combine the structural features to judge the geometric consistency and eliminate noise false peaks.

[0034] Residual peak point is defined as the local maximum value within a certain symmetric window, which satisfies the following conditions: .

[0035] For each candidate peak, the peak position , the peak height and the half-width are calculated. Based on this, a preliminary set of maxima is extracted from the residual signal as: .

[0036] To remove noise-triggered false pulses from the preliminary set of maxima , the following constraint criteria are defined: half-width constraint: based on the emitted laser pulse width, exclude too narrow (noise spikes) or too wide (background diffusion or additional broadening of extended targets) pulses; structural symmetry constraint: enhance the peak shape integrity judgment, including pulse peak intensity and pulse peak structural symmetry constraints. Keep all the above constraints that meet the high-confidence pulse set . It only reflects the clustering behavior in the local time domain. However, under the condition of ultra-low signal-to-noise ratio, the trigger of background noise concentration may also lead to pulse peaks that meet the time domain distribution characteristics, and form false target points based on the corresponding photon event reconstruction. Such false detection points often show discrete distribution in the spatial domain, rather than the spatial aggregation of real targets. Therefore, it is crucial to further enhance the clustering significance in the spatial domain. We design a spatial event enhancement strategy based on spatio-temporal joint saliency to extract the credible target region.

[0037] In the current frame , each high-confidence pulse corresponds to a local time window : .

[0038] Let denote the photon event count at time label within the gating window at pixel . Define the local time window photon event density function for each pixel as: .

[0039] That is, the cumulative event count within the corresponding time window is counted on each pixel, and this statistic reflects whether there is a photon dense phenomenon in this region within this time period, and its expectation depends on the background light flow distribution and target signal strength. Under natural light and no target condition, the background photon statistics conform to the spatially weakly correlated and temporally stationary Poisson distribution assumption. Therefore, the photon count of each pixel can be regarded as a Poisson random variable at spatial position . According to the central limit theorem, it can be approximately modeled as a normal distribution within the local window: , where and are the expectation and standard deviation of the background, respectively. To avoid the interference of real target pixels on the background estimation, the spatial sliding window method is used to estimate the background in the background dominant area and :

[0040] and the normalized statistical deviation (Z-score) is defined as: .

[0041] For any position , the pixels in the space that satisfy are considered as the abnormal photon aggregation points in the time window, usually taking corresponding to the 5σ significance level, corresponding to the false alarm probability . According to this, a binary response mask is constructed to indicate the possible candidate target area: , wherein, represents any spatial position; , represents the normalized statistical deviation, wherein, is the estimated mean, is the estimated standard deviation, is the photon count of each pixel; The response mask triggered by the photon event characterizes the spatial response area of the potential target in the time window . Since individual high value pixels may be caused by incidental noise, the embodiment further introduces spatial consistency analysis to extract the real candidate target area, the specific steps including: Connected component analysis: based on 8-neighborhood, connected region labeling is performed on , and all pixel clusters are obtained; Area filtering: only the cluster area (such as ) is retained to remove isolated noise points or false detection points; Target region center definition: the centroid position of each retained connected region is calculated as the tracking target center: .

[0042] Temporal correlation assignment: for each region , its triggered time window center is associated to form a structured candidate area .

[0043] Finally, after spatial statistical enhancement and cluster rejection, the set of all effective candidate target regions in the current frame is obtained . wherein, is the number of targets to be extracted. The pixel points in each region will be input as a unified target to participate in the subsequent multi-target dynamic trajectory matching and three-dimensional tracking.

[0044] For each candidate region mask , the count value of the corresponding pixel photon event directly reflects the energy distribution of the target, is the mask exponential function, and the intensity of the candidate region is: .

[0045] Step 4: Based on the photon arrival time data of each candidate target region, a Bayesian weighted RANSAC algorithm is used for trajectory fitting and prediction to obtain the motion model and predicted arrival time of the target, and the target range image is reconstructed accordingly; for each candidate target region , the photon arrival time and the trigger frame index corresponding to the triggered photon event of the constituent pixel are extracted, and a two-dimensional time sequence point set is constructed: , wherein, is the pixel space corresponding to the candidate target region; is a robust fitting target motion trajectory under the condition of background noise and outlier interference, and a Bayesian prior weighted RANSAC (Bayesian-RANSAC) method is used to perform linear model fitting on : , wherein, and are the motion slope and bias term of the th target region, respectively, and the longitudinal flight speed of the target can be further calculated. The optimal model parameters are obtained by iterative sampling, and the corresponding inlier set corresponds to the photon events of the target echo: , wherein, represents the maximum allowed residual of the target peak (determined by the corresponding main peak width ), is the candidate photon time set.

[0046] At the same time, this model can be further extended to a second-order dynamic model to capture acceleration or maneuvering characteristics. Finally, based on the fitted motion model of each target i, the target arrival time at the end frame is predicted: .

[0047] Assign this value to the entire mask region, and reconstruct the range image of the target i region based on the time-of-flight method: .

[0048] In dynamic unmanned aerial vehicle (UAV) imaging process, the relative motion between the target and the detector will cause the imaging scale to change dynamically with the longitudinal distance. If not corrected, the traditional fixed scale tracking algorithm often appears positioning deviation or even target loss, which seriously affects the stability of multi-target tracking. To solve this problem, the embodiment proposes a scale adaptive updating tracking method based on Farneback dense optical flow, aiming to improve the robust tracking performance of single-photon lidar for dynamic multi-UAV targets. The method first calculates the optical flow vector field in the local area of the target, and realizes the dynamic updating of the target position through statistical analysis; then combines the BANSAC distance guided template matching bottom mechanism and the pyramid multi-layer search strategy to enhance the robustness of tracking and effectively avoid target loss; finally, the scale adaptive updating module is introduced to correct the imaging size change caused by longitudinal motion, so as to maintain the geometric consistency of ROI in time sequence. The overall algorithm process is as shown in Figure 2 .

[0049] Initial target region determination step: in the candidate target region obtained in the segmentation stage , a rectangular region of interest (ROI) is selected as the tracking window, denoted as: , wherein is the upper left corner coordinate of ROI; is the width and height of the target region; i is the number of targets. The region serves as the search basis for subsequent optical flow estimation and template matching.

[0050] Target local area optical flow field estimation step: the Farneback method is used to calculate the dense optical flow between the intensity image sequences and of any time and . The Farneback method is based on the fact that the pixel intensity in the local area of the target in the field of view satisfies the following quadratic approximation form: . Then, through multi-scale pyramid modeling and minimizing the overall reconstruction error, the optical flow vector field of any pixel in each target region is obtained.

[0051] Target region average optical flow calculation and ROI updating step: in each frame t, the optical flow is weighted and averaged in the ROI to obtain the position offset:

[0052] wherein the weight function takes Gaussian distribution to further weaken the edge noise and highlight the target region: ​

[0053] where, is the ROI center.

[0054] The position update is: .

[0055] Template matching bottom-up mechanism step: when the optical flow vector in ROI is insufficient or abnormally distributed (for example, the proportion of valid pixels <10%), the bottom-up mechanism is triggered. Based on BANSAC fitting, the distance range of the target region and the neighborhood determined near the ROI are calculated to calculate the normalized cross-correlation:

[0056] where T is the target template, and the maximum response position is taken as the new position of the ROI. When local failure occurs, the pyramid layer-by-layer scaling search image and template method is used for recovery, avoiding tracking loss caused by large displacement or rapid motion.

[0057] Tracking scale adaptive update step: to adapt to the imaging scale change of the unmanned aerial vehicle, the mean value of the regional optical flow amplitude is introduced as the scale estimation index. In consecutive frames, the local optical flow amplitude change of the target can approximately reflect the projection scale transformation trend of the target in the image. The local average amplitude of the target region is defined as: .

[0058] The amplitude ratio of the current frame to the previous frame is taken as the relative scale factor: where, is the local average amplitude of the target region in the previous frame, is a small constant to prevent division by zero; the tracking window size is updated to:

[0059] where, is the width of the tracking window in the current frame; is the length of the tracking window in the current frame; is the width of the tracking window in the previous frame; is the length of the tracking window in the previous frame; , , and are the coordinate scales of the window size position. To suppress the jitter of the size update, an exponential moving average is introduced to perform low-pass filtering on : where, is the smoothed estimation value of the local average amplitude of the target region, is the smoothed estimation value of the local average amplitude of the target region in the previous frame. The jump range of is also constrained.​ .

[0060] If , it represents the maximum allowed 2% scale change per frame. Considering that the target size fluctuation between consecutive frames is not significant, the algorithm updates the scale with a fixed frame interval to improve the real-time performance of adaptive processing.

[0061] Multi-target simultaneous tracking step: the above updating process can be used for multiple targets , to achieve multi-target tracking.

[0062] Embodiment two, the computer device of the embodiment, comprising a memory and a processor, the memory has stored a computer program, when the processor runs the computer program stored in the memory, the processor executes the multi-target perception and tracking method of the single-photon laser radar detection of the swarm of unmanned aerial vehicles in embodiment one.

[0063] Embodiment three, a computer readable storage medium, the computer readable storage medium has stored a computer program, the computer program is run by the processor to execute the steps of the multi-target perception and tracking method of the single-photon laser radar detection of the swarm of unmanned aerial vehicles in embodiment one.

[0064] Embodiment four, see Figures 3-40 This embodiment integrates the technical solutions of the above-mentioned multiple embodiments, and further verifies and explains the technical effects of the present application through specific examples in combination with the actual application scene and the use process of the computer program product for implementing the method of the present application.

[0065] In order to evaluate the detection and segmentation performance of the proposed algorithm, the experiment compares it with various reconstruction methods at different distances, including matched filter (MF), maximum likelihood estimation (MLE), peak method (PKF) and kernel density estimation (KDE). The reconstruction results of the above methods under multi-frame conditions are shown in Figures 3-7 The experimental results show that: when the signal-to-noise ratio is lower than -4.6040dB, the peak method fails to reconstruct; when the signal-to-noise ratio is lower than -7.0666dB, the maximum likelihood estimation fails; when the signal-to-noise ratio is lower than -10.7697dB, the kernel density estimation method cannot effectively reconstruct the target; when the signal-to-noise ratio is lower than -11.3427dB, the matched filter method cannot effectively reconstruct the target.

[0066] Taking frame 3412 as an example, the reconstruction and target segmentation process of this invention is as follows: For the DJI Mavic Pro near the imaging center position (7, 20), the average number of photoelectrons in the target region is Ns = 0.1719; the average number of photoelectrons in the noise region is Nn = 3.8391, and the regional signal-to-noise ratio is as low as approximately -13.4908 dB. After adaptively gating and suppressing systematic front-peak noise, a frequency histogram of photon arrival times is plotted, as shown below. Figure 8 As shown in the left-hand diagram; Figure 8 As shown in the right-hand figure, a frequency residual signal (red line) is constructed by fitting an exponential decay model (blue dashed line) to highlight potential echo pulses. First-order peak detection is used to initially identify significant peaks, such as... Figure 9 The left-hand figure shows the location of potential target echo photon events. Second-order structural screening further identifies high-confidence pulse peaks; the shaded area represents the full width at half maximum (FWHM). Only nine pulse peaks matching the target characteristics are retained, such as... Figure 9 As shown in the right figure, after completing the first two stages of screening, each high-confidence peak has a clear pulse width and peak position in the time domain. However, under ultra-low signal-to-noise ratio conditions, relying solely on the time domain peak may result in isolated noise triggering, failing to accurately reflect the spatial distribution of target photon events. To further eliminate isolated noise and extract spatially continuous target regions, the photon arrival time intervals corresponding to each peak pulse width are statistically analyzed. Photon events are mapped to a pixel array, and through 8-neighborhood connected component analysis, only spatially continuous effective pixel clusters are retained. Third-order pulse peak judgment is performed to identify two pulse peaks that meet the conditions, and the trigger counts of the corresponding connected regions are counted to achieve intensity image reconstruction. The results are shown below. Figure 10 and Figure 11 As shown.

[0067] The BANSAC algorithm is used to fit the photon time series of each candidate region, outliers are removed, and the arrival time of target photons at the end of the frame is predicted. This, combined with the time-of-flight method, completes the range profile reconstruction. The results are as follows: Figures 12-14 As shown in the figure, the position of the trigger frequency of segmenting the target region in the global frequency distribution histogram is as follows. Figure 16 As shown. Further experiments show that the experimental results are available in [reference needed]. Figures 15-18 Even when the signal-to-noise ratio is as low as -14.08 dB (Ns=0.1744, Nn=4.4636), this invention can still accurately reconstruct and segment multiple targets, reaching the critical point for algorithm application. During reconstruction, the connected regions retain relatively clear spatial distribution characteristics. Compared to previous methods, it can be determined that the target corresponding to the first peak is a DJI Phantom 4, and the target corresponding to the second peak is a DJI Mavic 3, providing a more reliable basis for subsequent target tracking, type recognition, and target size calculation.

[0068] To solve the problem of trajectory intersection and projection overlap in the dynamic flight of UAV swarm, the experiment selects typical frames in the far field, the middle distance and the near field for analysis to verify the adaptability of the algorithm in different distance conditions. In the far field (Ns=0.6013, Nn=4.1082, SNR≈-8.3453dB), based on the matching filter reconstruction result as shown in Figure 19 , the number of UAVs existing in the field of view cannot be effectively distinguished. The present application accurately judges that there are two targets by three-order pulse peak detection, as shown in Figure 20 , and marks the corresponding pulse peak position. Then the two target intensity images are reconstructed based on the connected domain and time interval trigger frequency. Based on the BANSAC model prediction and the flight time method, the flight distances of Phantom 4 and Mavic Pro are 312.363 m and 264.000 m respectively. The reconstruction results of the two targets and the photon arrival time fitting distribution are shown in Figures 21-23 . In the middle distance (Ns=1.8596, Nn=4.0673, SNR=-3.3990dB), the matching filter method only generates a fuzzy intensity distribution, as shown in Figure 24 , which makes it difficult to realize target differentiation. The present application successfully separates the two UAVs under occlusion, and the flight distances are 170.537m (Phantom 4) and 94.350m (Mavic Pro) respectively. The corresponding time domain residual and frequency distribution histogram are shown in Figure 25 , the reconstruction results of the two targets and the photon arrival time fitting distribution are shown in Figures 26-27 . In the near field (Ns=1.5002, Nn=4.4213, SNR≈-4.6939dB), the matching filter reconstruction result of the near field multi-UAV target is shown in Figure 28 , the corresponding time domain residual and frequency distribution histogram are shown in Figure 29 , the reconstruction results of the two targets and the photon arrival time fitting distribution are shown in Figures 30-31 . As can be seen from the figure, the present application successfully distinguishes the two UAV targets and can determine the target type from the imaging results, and the calculated target distance is: 87.153m for Phantom 4 and 69.185m for Mavic Pro.

[0069] To further evaluate the separation performance of the algorithm in the multi-target scene, the present embodiment analyzes the different longitudinal distance between UAVs, as shown in Figures 32-34 . The experiment shows that when the pulse time interval of the two targets is not less than 33ns (corresponding to a longitudinal distance difference of ≥4.95m), the present application can effectively distinguish and reconstruct multiple target echoes; when the pulse time interval is less than 33ns, as shown in Figure 35 , the multiple pulse peaks overlap significantly in the time domain, and the three-order peak detection cannot accurately separate them, as shown in Figure 36 .As shown, leading to the appearance of artifacts in the reconstructed intensity image. Accordingly, the multi-target minimum resolvable range quantification of the present application is 4.95 m.

[0070] The present application also compares tracking with two representative baseline methods in the far-field multi-target unmanned airfield scenario, including: (1) Particle Filter (PF), using a particle size of 150 to ensure convergence stability and real-time performance; (2) MeanShift + Kalman fusion (Mean-Kalm), representing an engineering commonly used fusion framework based on appearance color features and linear motion prior.

[0071] All methods use uniform detection results and initial boxes to eliminate initialization bias and maintain consistent sampling frequency and reconstruction frame number. Tracking accuracy is measured by the two-dimensional Euclidean distance between the estimated target center and the manually labeled reference position frame by frame, and further statistics of mean error (Mean), standard deviation (STD) and root mean square error (RMSE) are obtained. Wherein, the target tracking center position of the t-th frame is denoted as, and the center position of the manually labeled target is denoted as , then: .

[0072] In the far-field double-airfield scenario (SNR = -13.83 dB, Ns = 0.1618, Nn = 3.9055), the differences between different methods are significant. The results are shown in Table 1. Figures 37-40 For the far-field double-airfield scenario, the mean error distance and the root mean square error distance of target 1 are 0.3801 and 0.4548, respectively, which are significantly better than Particle Filter (1.4849 and 1.7814) and better than Mean-Kalm (0.6868 and 0.7695); the mean error distance and the root mean square error distance of target 2 are 0.6147 and 0.8775, respectively, which are also significantly better than PF (18.9899 and 23.1695) and slightly better than Mean-Kalm (0.7001 and 0.9433). PF almost completely loses tracking on target 2, with a high level of error for a long time, while OpticalFlow can fully utilize local intensity information for reliable estimation by calculating Farneback dense optical flow in the candidate region and updating the ROI center by Gaussian weighted average, even if the pixels are sparse. Further combined with normalized cross-correlation matching, pyramid layer-by-layer search and ROI adaptive adjustment, the present application can effectively deal with fast motion or partial occlusion, ensuring that the target is not lost between consecutive frames, while avoiding tracking drift and mismatch.

[0073] Table 1 Error comparison table of far-field double-airfield tracking algorithm Table 1 Error comparison table of far-field double-airfield tracking algorithm

[0074] In summary, the scale adaptive update tracking method based on Farneback dense optical flow (OpticalFlow) exhibits excellent tracking performance in various scenarios: in the near-field three-aircraft scene, the Mean Distance is reduced by an average of about 22.1% and the RMSE Distance is reduced by an average of about 16.2% compared to the traditional method of fusing appearance color features and linear motion priors; in the far-field two-aircraft sparse photon condition, the Mean Distance and RMSE Distance are reduced by an average of about 28.5% and 23.9%, respectively. These quantitative results fully demonstrate that the method has significant robustness and accuracy advantages in complex dynamic unmanned aircraft scenarios.

[0075] The above examples are only used to illustrate the technical solutions of the present disclosure and not to limit the protection scope thereof. Although the present disclosure has been described in detail with reference to the above examples, those skilled in the art can make various changes, modifications or equivalent replacements to the specific implementation manners of the present disclosure after reading the present disclosure. However, these changes, modifications or equivalent replacements are all within the protection scope of the disclosed claims.

Claims

1. A method for multi-target perception and tracking of UAV swarms using single-photon lidar, characterized in that, The method includes: Step S1: Statistically analyze the global photon arrival time data collected by the single-photon lidar, construct a time frequency distribution histogram, and perform wavelet denoising processing on the time frequency distribution histogram; Step S2: Identify and fit the decay trend of the front noise tail in the time frequency distribution histogram, determine the boundary time point between the front peak noise and the stable background area by calculating the first derivative of the decay trend, and remove non-physical noise data before the boundary time point. Step S3: Model the background noise exponential decay of the time frequency distribution histogram of the data excluding non-physical noise, and fit the background parameters using the linear least squares method; calculate the residual between the global trigger frequency and the background fitted frequency; Step S4: Perform wavelet transform, soft threshold compression, inverse wavelet transform reconstruction, and median filtering on the residual to obtain an enhanced smooth residual signal; Step S5: Detect candidate pulse peaks from the smoothed residual signal, and filter them based on the full width at half maximum (FWHM) constraint and structural symmetry constraint to obtain a set of high-confidence pulse peaks; Step S6: Map the photon events within the time window corresponding to each high-confidence pulse peak to the pixel space, calculate the standardized statistical bias of each pixel, and extract candidate target regions based on spatial connectivity analysis; Step S7: Based on the photon arrival time data of each candidate target region, the Bayesian weighted RANSAC algorithm is used to perform trajectory fitting and prediction to obtain the target's motion model and predicted arrival time, and the target range profile is reconstructed accordingly. Step S8: For each segmented candidate target region, calculate the local optical flow field of the target within the field of view based on the Farneback dense optical flow method, update the center position of the target region by weighted averaging, and introduce scale adaptive update to dynamically adjust the tracking window size according to the change of optical flow amplitude.

2. The method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 1, characterized in that, The attenuation trend of the front-end noise tail in the fitted time frequency distribution histogram in step S2 includes: Starting from an offset after the global maximum point of the time frequency distribution histogram, an exponential function model is used for fitting. The exponential function model is as follows: in, To fit the peak height of the photon event triggering frequency; It is the exponential decay constant; Indicates the noise-stabilized baseline. This is the index for the main noise peak within the gate. x The bin index of the time corresponding to the triggering photon event This is the initial offset for fitting.

3. The method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 2, characterized in that, Determining the boundary time point between the leading noise peak and the stable background region in step S2 includes: Find the time point at which the first derivative of the fitted curve first enters the interval [-1, 0]: ， in, This serves as the boundary for the transition from a systemic unsteady response to a stable background. The first derivative of the curve is expressed as: ， in, This is the fitted value corresponding to the time index of the next triggering photon event. This is the fitted value corresponding to the time index of the currently triggered photon event. For time bin width, This is the time index corresponding to the next triggering photon event. This is the time index corresponding to the current photon event.

4. The method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 2, characterized in that, Step S4 specifically involves: Perform a three-level Daubechies-4 discrete wavelet transform on the residual signal to obtain high-frequency detail coefficients: ， in, For the third layer low-frequency approximation coefficient, to These are the corresponding high-frequency coefficients for details; The obtained high-frequency detail coefficients Applying the soft threshold function: ， Among them, threshold , for The median absolute deviation estimate, The length of the trigger photon event is input. These are the detail coefficients for each level of the discrete wavelet transform; use Perform an inverse wavelet transform on the high-frequency coefficients of the details after soft thresholding to reconstruct the denoised residual: ， in, to These correspond to the three processed detail coefficients; right By performing sliding window median filtering, an enhanced smooth residual signal is obtained while preserving the pulse edge structure. for: ， in, The denoised reconstruction residuals of the adjacent previous time index. This is the denoised reconstruction residual of the adjacent next time index.

5. A method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 1, characterized in that, The extraction of candidate target regions in step S6 includes: Construct a binary response mask based on standardized statistical bias: ， in, Represents any spatial location; , representing standardized statistical bias, where, To estimate the mean, To estimate the standard deviation, Count the number of photons for each pixel; Perform connectivity analysis on the binary response mask, filter out regions with areas smaller than a set minimum value, and use the centroids of the remaining connected regions as candidate target centers: ， in, Indicates pixel clustering; For each region Associate it with the center of the time window it triggers , forming structured candidate regions .

6. The method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 5, characterized in that, Step S7 includes: For each candidate target region Extract the photon arrival time corresponding to the trigger photon event of the constituent pixels. With trigger frame index Construct a two-dimensional time series point set: ， in, The pixel space corresponding to the candidate target region; The Bayesian prior weighted RANSAC method is used to... Perform linear model fitting: ， in, For the first The slope of motion in each target region For the first The bias term for each target region To fit the time index corresponding to the photon event, For fitting constants; Optimal model parameters are obtained through iterative sampling. The corresponding set of interior points corresponds to the photon events of the target echo as follows: ， in, For candidate photon time sets, This represents the maximum permissible residual for the target peak. Extending the optimal model to a second-order dynamic model, based on the fitted motion model of each target i, predicts the target arrival time in the final frame: ， in, The time index corresponding to the termination frame; Will Assign the entire mask region and reconstruct the range image of target region i based on the time-of-flight method: 。 7. The method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 1, characterized in that, The scale-adaptive update in step S8 includes: The ratio of the amplitude of the current frame to that of the previous frame is used as the relative scale factor: ， in, The local average amplitude of the target region. This represents the local average amplitude of the target region in the previous frame. To prevent division by zero of small constants; Update the tracking window size based on the relative scale factor: ， in, The width of the tracking window for the current frame; The length of the tracking window for the current frame; This is the width of the tracking window from the previous frame; The length of the tracking window in the previous frame; , , and The coordinate scale represents the window size and position.

8. A method for multi-target perception and tracking of UAV swarms using single-photon lidar according to claim 7, characterized in that, Low-pass filtering is performed on the local average amplitude of the target region using exponential sliding motion. ， in, This is the estimated value of the local average amplitude of the target region after smoothing. This is the estimated value of the local average amplitude of the target region in the previous frame after smoothing.

9. 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 multi-target perception and tracking method for detecting UAV swarms using a single-photon lidar, as described in any one of claims 1-8.

10. 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 multi-target perception and tracking method for detecting UAV swarms using a single-photon lidar as described in any one of claims 1-8.

Images