Unmanned aerial vehicle sar wounded person detection living body discrimination method based on micro-doppler-distance

Abstract

The application discloses a kind of unmanned aerial vehicle SAR wounded detection living body discrimination method based on micro-doppler-distance. The method first generates SAR image under wide-area search mode, and adopts neural network screening candidate target based on physical prior constraint;Then switch to review mode, obtain review mode SAR echo data and extract slow-time complex echo sequence, construct micro-doppler-distance joint representation graph;SAR image block and micro-doppler-distance joint representation graph of candidate target are simultaneously input into state space model fusion network for joint feature extraction and state space modeling, and the unified semantic representation fusion feature obtained is detected, segmented, living body discrimination and positioning based on it. The application can effectively reduce the false detection rate of complex background false target under all-weather conditions, improve the accuracy of wounded target detection, living body discrimination and positioning, and is suitable for disaster site rapid search and rescue task.

Description

Technical Field

[0001] This invention belongs to the field of UAV-borne synthetic aperture radar imaging and intelligent detection technology, specifically involving a UAV SAR wound detection and liveness detection method based on micro-Doppler-range. Background Technology

[0002] Following disasters and accidents, the scene often presents challenges such as complex terrain, severe obstructions, harsh environments, and urgent rescue time, necessitating the use of unmanned aerial vehicle (UAV) platforms for rapid search and location. Existing emergency rescue UAVs primarily employ visible light and infrared sensors for target detection; however, these methods are susceptible to environmental factors such as nighttime, rain, fog, smoke, and insufficient lighting, making it difficult to meet the needs of search and rescue under all-weather, 24 / 7 conditions.

[0003] Synthetic Aperture Radar (SAR) possesses advantages such as active imaging, minimal susceptibility to lighting conditions, and strong adaptability to complex weather, making it suitable for target detection at disaster sites. Mounting SAR on a lightweight, small unmanned aerial vehicle (UAV) platform enables rapid, mobile reconnaissance of disaster areas, demonstrating promising application prospects.

[0004] In recent years, research has explored a two-stage search and rescue model that utilizes unmanned aerial vehicles (UAVs) for wide-area search, combined with subsequent hovering radar or life detection modules for target verification. In SAR target detection, a cascaded detection framework combining constant false alarm rate (CFAR) initial screening with deep convolutional neural network (DCNN) secondary discrimination has been widely adopted. Regarding vital sign detection, numerous publicly available studies have utilized the micro-Doppler effect to extract vital activity features such as respiration and heartbeat, and employed methods such as blind source separation and independent component analysis (ICA) to separate weak life signals from radar echoes. Furthermore, state-space models such as Mamba have been attempted for SAR image feature extraction and multi-source remote sensing data fusion. However, current technologies still lack a complete and systematic solution for jointly modeling SAR spatial scattering characteristics and micro-motion characteristics of vital activities, and for using a unified state-space model to achieve high-precision detection, liveness detection, and localization of wounded targets in complex rescue environments.

[0005] Specifically, existing UAV SAR still has the following problems when used for casualty detection and location: First, lightweight unmanned aerial vehicle (UAV) platforms are prone to attitude disturbances and track deviations during flight, resulting in range migration and phase errors in SAR echoes, affecting imaging quality and thus reducing the detectability of weakly scattering casualty targets. Existing motion compensation and autofocus methods still struggle to guarantee ideal focusing results when the signal-to-clutter ratio of casualty targets is extremely low.

[0006] Second, wounded targets in SAR images typically appear as weakly scattered, small-scale, and discrete sets of scattering points, easily confused with complex backgrounds such as ruins, trees, and metal structures. Although detection methods combining CFAR and CNN exist, the discrimination ability of single spatial domain features is limited when faced with wounded targets that are varied in shape and partially occluded, resulting in a still relatively high false detection rate.

[0007] Third, most existing vital sign radar detection methods require the target to be at close range or in a known location beforehand, making it difficult to effectively integrate with wide-area SAR searches. Although some studies have attempted to extract the slow-time signal corresponding to the target after SAR imaging for time-frequency analysis, their ability to jointly suppress platform disturbances, clutter, and breathing harmonics is insufficient, making it difficult to stably extract weak heartbeat components, and they lack a mechanism for simultaneously verifying and distinguishing multiple candidate targets.

[0008] Fourth, most existing methods process SAR spatial domain features and life activity temporal domain features separately, lacking a unified fusion network capable of simultaneously modeling long-range spatial contextual dependencies and temporal life activity cycle characteristics. Traditional feature extraction struggles to balance global contextual information and local detailed features in complex scenarios, affecting the accuracy of casualty target detection, segmentation, and localization, and further failing to output life activity scores that can be used for rescue priority ranking.

[0009] Therefore, it is necessary to propose a new UAV SAR method for detecting and locating wounded personnel, so as to achieve high-precision detection, location and auxiliary identification of wounded targets in complex rescue environments. Summary of the Invention

[0010] The purpose of this invention is to provide a liveness detection method for casualties based on micro-Doppler-range SAR on unmanned aerial vehicles (UAVs), in order to solve the problems in the prior art, such as the lack of effective coordination between wide-area SAR search and vital sign verification, the difficulty in steadily extracting weak life signals under platform disturbance and complex clutter backgrounds, and the lack of a deep fusion mechanism between the spatial scattering structure features of SAR images and the time-frequency features of life activities. This will enable high-precision detection, reliable liveness detection, and accurate positioning of casualty targets in complex rescue environments.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: a method for liveness detection of wounded personnel by unmanned aerial vehicle (UAV) SAR based on micro-Doppler-range, comprising: The raw SAR echo data and platform positioning attitude data of the target area are acquired simultaneously using the UAV platform, and a SAR image of the target area is generated in wide-area search mode. Based on the SAR image, a candidate target screening network based on physical prior constraints is used to screen suspected wounded targets, and candidate targets and their coarse localization results are obtained. The control drone platform is switched to verification mode, hovering and staring above the candidate target to acquire verification mode SAR echo data of the candidate target; Based on the aforementioned verification mode SAR echo data, the candidate targets and their coarse positioning results are mapped to the range gates of the verification mode SAR echo data to determine the center range gate and its adjacent range gates corresponding to the candidate targets; within the center range gate and its adjacent range gates, slow-time complex echo sequences are extracted. Time-frequency analysis was performed on the slow-time complex echo sequence, and a time-range-Doppler data cube was formed by stacking along the distance dimension. A micro-Doppler-distance joint characterization map was constructed based on the time-range-Doppler data cube. The SAR image patch and the micro-Doppler-range joint representation map of the candidate target are simultaneously input into the state space model fusion network for joint feature extraction and state space modeling, resulting in a unified semantic representation fusion feature that combines boundary localization capability and liveness detection capability. Based on unified semantic representation fusion features, suspected wounded targets are detected, segmented, liveness-based and located, and output the location, boundary, confidence level, vital activity score and rescue priority information of the wounded targets.

[0012] Further preferably, the candidate target screening network based on physical prior constraints includes a two-dimensional constant false alarm rate (CFAR) initial screening module, a scattering geometry prior screening module, and a convolutional feature secondary screening module; wherein, the two-dimensional CFAR initial screening module is used to extract initial candidate regions from SAR images, the scattering geometry prior screening module is used to perform prior constraint screening on the initial candidate regions based on the statistical characteristics of the initial candidate regions, and the convolutional feature secondary screening module is used to perform rapid secondary discrimination on the candidate regions after prior constraint screening and output the confidence scores of the candidate targets; The candidate target screening network filters out suspected wounded candidates that meet the requirements based on the comprehensive score of the candidate region, and outputs their coarse localization results; the comprehensive score of the candidate region is expressed as: ; in, This represents the overall score of the candidate regions. This represents the score of the two-dimensional constant false alarm rate (CFAR) detection. This represents the prior score of the scattering geometry. This represents the score from the convolutional feature re-screening. , , These are the weighting coefficients for each score item.

[0013] Further optimization involves first performing envelope alignment and phase compensation on the SAR echo data in the verification mode; then constructing a multi-channel observation matrix using the center range gate corresponding to the candidate target and the slow-time complex echo sequences within adjacent range gates; whitening the multi-channel observation matrix to decouple the signals of each channel in a second-order statistical sense; subsequently constructing a higher-order statistical matrix of the whitened signal and performing joint approximate diagonalization to decompose the mixed signal into several independent source components; by performing frequency domain analysis and time domain continuity analysis on the independent source components, selecting the component with stable waveform within the breathing frequency range as the main breathing component; then constructing a feedback notch filter using this main breathing component as a reference signal to perform narrowband suppression on the sequence containing vital activity components, resulting in an enhanced slow-time complex echo sequence.

[0014] Further optimization involves using a three-dimensional constant false alarm rate (CFAR) detection module to perform three-dimensional CFAR detection on the time-distance-Doppler data cube, obtaining significant micro-motion component distribution results, and generating a micro-Doppler-distance joint characterization map based on the significant micro-motion component distribution results.

[0015] Further preferably, the state space model fusion network includes a SAR spatial domain branch, a life activity time-frequency domain branch, and a cross-domain fusion module; The SAR spatial domain branch is used to embed features into SAR image patches of candidate targets to obtain initial features of spatial scattering structure, and to perform state space modeling on the initial features of spatial scattering structure to obtain deep features of spatial scattering structure; the life activity time-frequency domain branch is used to encode features into the micro-Doppler-range joint characterization map of candidate targets to obtain initial features of life activity time-frequency, and to perform state space modeling on the initial features of life activity time-frequency to obtain deep features of life activity time-frequency; the deep features of spatial scattering structure and the deep features of life activity time-frequency are concatenated to obtain fused features; the fused features are processed by the cross-domain fusion module to form a unified semantic representation fused feature that has both boundary localization capability and liveness detection capability.

[0016] Further preferred, the SAR spatial domain branch first embeds features into the SAR image blocks of the candidate target to obtain initial features of the spatial scattering structure. Then, the initial features of the spatial scattering structure are expanded into sequences along four directions: forward horizontal direction, reverse horizontal direction, forward vertical direction, and reverse vertical direction. These sequences are then input into the Mamba state space modeling unit for long-range context modeling to obtain deep features of the spatial scattering structure. The time-frequency domain branch of the life activity first encodes the micro-Doppler-distance joint representation map to obtain the initial time-frequency features of life activity; then, it performs global temporal dependency modeling on the initial time-frequency features of life activity: an outer sequence is constructed along the slow time direction to model the global periodic dependency of respiration and heartbeat throughout the observation period; then, it performs local time-frequency texture modeling: an inner sequence is constructed along the local frequency distance window within each slow time segment to model the coupling relationship between local time-frequency texture changes and adjacent distance gates; then, it performs nested space modeling to obtain the deep time-frequency features of life activity.

[0017] Further optimization involves the region-aware attention module generating spatial attention weights and channel attention weights based on the spatial response intensity and salience of life activities in candidate regions. This enhances the features of candidate targets by weighting them while suppressing background features. The gated recursive context refinement module recursively fuses high-level semantic fusion features with low-level boundary detail features. In each round of recursive updates, the gating unit adjusts the retention weights of the high-level semantic fusion features and low-level boundary detail features, ultimately forming a unified semantic representation fusion feature that combines boundary localization and liveness detection capabilities.

[0018] Further optimization employs a multi-task output network to jointly discriminate the fused features, outputting casualty detection results, casualty segmentation results, vital signs scores, and geolocation results respectively; the following multi-task training loss function is used during training: ; Where L is the total loss, For target detection loss, To segment the loss for the target, Score the loss of vital functions. The consistency constraint loss between the deep features of spatial scattering structures and the deep features of time-frequency biological activities is used. To pinpoint the loss, These are the weighting coefficients for the corresponding loss terms.

[0019] The present invention also provides an electronic device, including a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program to enable the electronic device to perform the UAV SAR wounded person detection and liveness determination method.

[0020] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor performs the aforementioned UAV SAR wounded person detection and liveness determination method.

[0021] The beneficial effects of this invention are as follows: This invention employs a two-stage working mode combining wide-area search and target verification, enabling rapid screening of suspected casualty targets in large-scale disaster scenarios and long-term verification of candidate targets, effectively balancing search efficiency and discrimination accuracy. By performing range compression, motion compensation, and phase stabilization correction on SAR echo data, it improves the imaging quality of the UAV platform under disturbance conditions and enhances the detectability of weak casualty targets in complex backgrounds. By constructing a micro-Doppler-range joint representation, it fully utilizes the micro-motion features generated by vital activities such as breathing and heartbeat, improving the reliability of liveness detection. By constructing a state-space model fusion network, it achieves joint modeling of the spatial scattering structure features of SAR images and the time-frequency features of vital activities, improving the accuracy of casualty target detection, segmentation, and localization in complex scenarios. Therefore, this invention can effectively reduce the false detection rate of pseudo-targets in complex backgrounds and improve the accuracy of casualty target detection, liveness detection, and localization under all-day, all-weather conditions, making it suitable for rapid search and rescue missions at disaster sites. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the micro-Doppler-distance joint characterization construction process provided in the embodiments of the present invention.

[0024] Figure 3 This is a schematic diagram of the state-space model fusion network structure provided in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to embodiments. It should be noted that the following embodiments are only for further illustration of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-essential improvements and adjustments made by those skilled in the art based on the disclosure of the present invention should fall within the scope of protection of the present invention.

[0026] like Figure 1 As shown, this embodiment provides a liveness detection method for unmanned aerial vehicle (UAV) SAR casualty detection based on microDoppler-range. This method employs a two-stage processing approach combining wide-area search and target verification, including the following steps: Step S1: Simultaneously acquire the raw SAR echo data and platform positioning attitude data of the target area using the UAV platform, and generate a SAR image of the target area in wide-area search mode; Step S2: Based on the SAR image, a candidate target screening network based on physical prior constraints is used to screen suspected wounded targets to obtain candidate targets and their coarse localization results; Step S3: Control the UAV platform to switch to verification mode, hover and stare above the candidate target to acquire the verification mode SAR echo data of the candidate target; Step S4: Based on the verification mode SAR echo data, map the candidate target and its coarse positioning result to the range gate of the verification mode SAR echo data, and determine the center range gate and its adjacent range gates corresponding to the candidate target; extract the slow time complex echo sequence within the center range gate and its adjacent range gates; Step S5: Perform time-frequency analysis on the slow-time complex echo sequence, and stack along the distance dimension to form a time-distance-Doppler data cube. Construct a micro-Doppler-distance joint characterization map based on the time-distance-Doppler data cube. Step S6: Simultaneously input the SAR image patch and the micro-Doppler-range joint representation map of the candidate target into the state space model fusion network for joint feature extraction and state space modeling, to obtain a unified semantic representation fusion feature that has both boundary localization capability and liveness detection capability; Step S7: Based on the unified semantic representation fusion features, detect, segment, determine the liveness of suspected wounded targets and locate them, and output the location, boundary, confidence level, vital activity score and rescue priority information of the wounded targets.

[0027] In this embodiment, the UAV platform is a lightweight multi-rotor aircraft, equipped with a SAR radar system, a navigation and positioning system, an attitude measurement unit, a mission control module, and an onboard data storage module. The navigation and positioning system provides the platform's position, velocity, and time information; the attitude measurement unit provides the platform's roll, pitch, and yaw angles; the SAR radar system transmits radar signals to the target area and receives echoes; and the onboard data storage module synchronously records the raw SAR echo data and the platform's attitude data. To ensure subsequent motion compensation and positioning accuracy, a high-precision clock or pulse synchronization signal is used to synchronize the SAR echo acquisition time with the platform's attitude data acquisition time. The synchronization relationship satisfies the following: ; in, Indicates the SAR echo acquisition time. This indicates the time when the platform acquired the positioning attitude data. This indicates the upper limit of the synchronization error allowed by the system. The above synchronization method ensures that the raw SAR echo data and platform pose data are processed under the same time reference, thereby improving the accuracy of subsequent slant range compensation and imaging focusing.

[0028] During the wide-area search phase, the UAV flies over the search area according to a preset trajectory, continuously collecting raw SAR echo data. Let the raw echo signal be... ,in To save time, For slow time, the raw SAR echo data is first preprocessed, including abnormal pulse removal, amplitude-phase consistency correction, echo normalization, and DC removal. After preprocessing, range compression is performed on the SAR echo signal to improve range resolution. Subsequently, motion compensation is performed by combining the navigation and positioning information and attitude information of the UAV platform. Because lightweight UAVs are prone to attitude disturbances and track deviations during flight, platform motion errors can cause changes in target slant range, thus introducing additional phase errors into the SAR echo signal. Let the slant range deviation caused by platform motion errors be... Then its phase compensation term is expressed as: ; in, This represents the phase compensation value. This indicates the radar's operating wavelength. By applying the aforementioned phase compensation to the range-compressed SAR echo signal, the main phase distortion caused by platform motion can be eliminated. After motion compensation, residual phase errors are eliminated through phase stabilization processing to improve the focusing quality of the imaging results.

[0029] In this embodiment, a temporal back-projection imaging method is used to generate a SAR image of the target area. Let the pixels to be imaged be... The platform is in slow time at pixel The slant distance between them is The complex image value at a pixel is represented as: ; in, This represents the complex image value corresponding to a pixel. Indicates the radar carrier frequency. Represents the speed of light. This represents the SAR echo signal after range compression, where j is the imaginary unit. A high-resolution SAR image of the target area is obtained by performing point-by-point backprojection accumulation on all pixels within the scene grid. The SAR image preserves the spatial scattering distribution of the scene target, local bright spot structures, and the scattering contrast between the target and the background, providing input for subsequent candidate target selection.

[0030] After acquiring SAR images, suspected casualty targets are rapidly screened. In this embodiment, the candidate target screening network includes a two-dimensional constant false alarm rate (CFAR) initial screening module, a scattering geometry prior screening module, and a convolutional feature secondary screening module. The two-dimensional CFAR initial screening module first performs sliding window statistical processing on the SAR image, sets up a protection unit and a background reference unit around each target unit, and adaptively generates a detection threshold based on the estimated local clutter power within the background reference unit. The detection threshold is expressed as: ; in, Indicates the detection threshold. This represents the estimated background clutter power. This represents the preset false alarm probability. When the response intensity of a pixel or candidate region to be detected is greater than the detection threshold, it is determined as an initial candidate region. The purpose of using a two-dimensional constant false alarm rate initial screening module is to quickly extract strong response or locally significant response regions from the entire SAR image, narrowing down the subsequent fine screening range and improving the efficiency of large-scene search.

[0031] The initial candidate regions obtained from the two-dimensional constant false alarm rate (CFAR) screening are further filtered using prior constraints based on scattering geometry. Specifically, statistical features such as the number of scattering points, the degree of scattering point aggregation, area, aspect ratio, compactness, and contrast between the initial candidate region and the surrounding background are extracted for each initial candidate region. Since wounded targets in SAR images typically appear as weak scattering, small scale, discrete bright spot sets, or locally irregular scattering clusters, while metal components, trees, rocks, and other environmental pseudo-targets in ruins differ from human targets in scattering morphology and structural distribution, these scattering geometric features can be used to perform prior constraint screening of the initial candidate regions. Initial candidate regions that do not meet the requirements for area range, aspect ratio range, or local contrast are eliminated; initial candidate regions with high geometric prior consistency are retained and proceed to the lightweight convolutional feature re-screening stage.

[0032] Specifically, let the pixel area of ​​the initial candidate region be... The longer side of the circumscribed rectangle is The shorter side is The regional average intensity is The average intensity of the neighborhood background is The standard deviation of the neighborhood background intensity is The aspect ratio is... and local contrast They are represented as follows: ; ; in, To prevent extremely small constants with a denominator of zero, the area range, aspect ratio range, and local contrast requirements are expressed as follows: ; ; ; Preferably, when the spatial resolution of the SAR image is 0.1m to 0.5m, To determine the minimum pixel area, a value of 3 to 8 pixels is used. The maximum pixel area is set at 80-300 pixels; the minimum aspect ratio is... Take 1.0 as the maximum aspect ratio. Set the contrast ratio to 6.0, the minimum contrast ratio. Take a value between 1.2 and 3.0. If the area of ​​the initial candidate region is less than... If the area is larger than a certain value, it is considered more likely to be an isolated noise point; if the area is larger than a certain value, it is considered more likely to be an isolated noise point. If the aspect ratio exceeds the set range, it is considered more likely to be a large-area feature, a linear structure, or a non-human false target; if the local contrast is lower than If the candidate region fails to meet the above constraints, it is considered that the candidate region is not sufficiently distinguishable from the background. Initial candidate regions that meet the above constraints are then entered into the convolutional feature re-screening module.

[0033] The convolutional feature re-screening module receives candidate region image blocks after geometric prior screening and performs further secondary discrimination on them. Considering that wounded targets in SAR images have weak texture, low contrast, and sparse local scattering points, this embodiment uses a multi-scale convolution method to extract candidate region features in the convolutional feature re-screening module. Specifically, features are simultaneously introduced into the same feature extraction layer. , and Three types of convolution kernels, among which Convolutional kernels are used to adjust channel information and suppress redundant features. Convolution kernels are used to extract local scattering textures and short-side structures. Convolutional kernels are used to extract the contour distribution and adjacent scattering structure relationships within a larger receptive field. To prevent weak scattering features in shallow layers from being weakened during multi-layer convolutional operations, cross-layer feature reuse connections are set between convolutional layers. This allows effective weak features preserved in the previous layer to be directly passed to subsequent feature extraction layers, thereby enhancing the ability to preserve weak scattering structures of the human body. Finally, the comprehensive score of the candidate region is expressed as: ; in, This represents the overall score of the candidate regions. This represents the score of the two-dimensional constant false alarm rate (CFAR) detection. This represents the prior selection score for scattering geometry. This represents the score from the convolutional feature re-screening. , , , where is the weighting coefficient. Based on the comprehensive score of the candidate regions, suspected wounded targets that meet the requirements are selected, and their coarse localization results are output. This processing method can both preserve weak targets in complex backgrounds and reduce the false target detection rate in large-area scenes.

[0034] In this embodiment, the weighting coefficient , , All are non-negative numbers and satisfy the following conditions: Preferably, Take a value of 0.20 to 0.40. Take a value of 0.20 to 0.35. Take a value of 0.30 to 0.55.

[0035] in, This is used to adjust the impact of the two-dimensional constant false alarm rate (CFAR) detection score on the overall score of the candidate region. Used to adjust the influence of scattering geometry prior screening scores Used to adjust the impact of convolutional feature screening scores. When the application scenario requires improving the recall rate of suspected injured targets, it improves... The value of ; when there are many false targets such as ruins, trees, rocks, or metal components in the scene, improve . The value of ; when the training samples of the convolutional feature screening module are sufficient and the discrimination ability is strong, the improvement The value of can be determined based on the candidate target recall rate, false target false detection rate, and comprehensive evaluation index on the validation set.

[0036] After candidate target screening is completed, the UAV platform is switched to verification mode. Verification mode is a hovering staring mode. After the UAV flies above the candidate target, the flight control system maintains stable attitude and continuously illuminates the candidate target with the SAR radar beam to acquire SAR echo data with a long dwell time. This mode is used because the displacement of the chest cavity caused by vital activities such as breathing and heartbeat is extremely small. Only under long-term, slow-time observation sequences can stable, weak vital activity components be extracted through time-frequency analysis and statistical separation.

[0037] In the verification mode, the SAR echo signal is first envelope aligned to compensate for distance migration caused by minor platform displacements during UAV hovering. Let two adjacent slow-time echo sequences be... and Then its cross-correlation function is expressed as: ; in, For two adjacent slow-time echo sequences, the time shift is... The cross-correlation function values ​​under the following conditions This is the previous slow-time echo sequence. For the next slow-time echo sequence, Echo sampling time, The relative time delay or time shift between two echo sequences. express The result of the complex conjugate operation.

[0038] By calculating the time shift corresponding to the peak value of the cross-correlation function, the distance offset between adjacent slow-time echo sequences can be obtained: ; in, This represents the optimal time offset between two adjacent slow-time echo sequences; Represents all candidate time shifts In the brackets, the time shift that causes the function within the brackets to reach its maximum value; Represents the cross-correlation function The amplitude of the cross-correlation function. Since a larger cross-correlation function amplitude indicates a higher degree of envelope matching between the two slow-time echo sequences at that time shift, therefore, when... When the maximum value is obtained, the corresponding This refers to the range offset of the subsequent slow-time echo sequence relative to the preceding slow-time echo sequence. The subsequent slow-time echo sequence is then shifted and compensated according to this range offset to ensure that the target scattering envelopes in adjacent echoes remain consistent in the range direction, thereby achieving envelope alignment between adjacent slow-time echo sequences.

[0039] The above process is repeated for the entire verification mode sequence to complete the envelope alignment of the entire sequence. After envelope alignment, the average phase of several aligned echoes is used as a reference to perform phase compensation on the current echo, in order to eliminate the overall phase change introduced by the slow attitude drift of the platform. Through envelope alignment and phase compensation, the same candidate target can be stably constrained within a similar range gate range throughout the entire slow-time observation process, reducing energy diffusion and inter-gate drift during subsequent life signal extraction.

[0040] After envelope alignment and phase compensation, platform perturbation cancellation and background clutter suppression are performed on the observed signals within the range gate range corresponding to the candidate target. Based on the position of the candidate target and its coarse localization result in the wide-area imaging map, the candidate target is mapped to the range gate in the verification mode SAR echo data. This mapping is based on the coarse localization result of the candidate target. Let the first... The distance gate index function corresponding to each candidate target is: Then its slow-time complex echo sequence Represented as: ; in, Indicates the first Slow-time complex echo sequences of candidate targets.

[0041] Considering that the scattering center of the human thoracic cavity may be distributed across multiple adjacent range gates in the range direction, this embodiment uses the range gate with the highest energy as the central range gate, and simultaneously extracts the slow-time complex echo sequences of the two adjacent range gates above and below it, forming a total of five observation signals. After removing baseline drift from the five observation signals, a multi-channel observation matrix is ​​constructed, as shown below: ; in, Represents a multi-channel observation matrix. Represents a mixture matrix. The source signal matrix is ​​represented by a matrix comprising the target respiratory component, the target heartbeat component, the plateau disturbance component, and the background clutter component. First, the multi-channel observation matrix is ​​whitened to decouple the signals in a second-order statistical sense. Then, a higher-order statistical matrix of the whitened signal is constructed, and a joint approximate diagonalization operation is performed to decompose the mixed signal into several independent source components. Through frequency domain analysis and time domain continuity analysis of the independent source components, the component with stable waveform within the respiratory frequency range is selected as the principal respiratory component. Then, a feedback notch filter is constructed using this principal respiratory component as a reference signal to perform narrowband suppression on the sequence containing vital activity components, progressively weakening the respiratory fundamental frequency and its higher harmonic components, and recovering the weaker heartbeat component.

[0042] In this embodiment, the feedback notch filter is a second-order IIR notch filter. Let the respiratory dominant frequency estimated from the respiratory principal component be... The slow time sampling frequency is Then and its higher harmonics As the center frequency of the notch filter, , The preferred values ​​are 2 to 4. For the first... Notch center frequency The transfer function of a second-order IIR notch filter is expressed as: ; in, The pole radius is used to control the notch bandwidth. Preferably, Use a value of 0.90–0.99; when the dominant respiratory rate is relatively stable, A value of 0.96–0.99 is used to obtain a narrower notch bandwidth; when there is a drift in the dominant breathing frequency or strong plateau disturbance, Use a value of 0.90 to 0.96 to expand the notch bandwidth.

[0043] Furthermore, the feedback notch filter employs a sliding time window method for adaptive updating. The respiratory frequency is re-estimated within each time window. The notch center frequency for the current time window is updated according to the following formula: ; in, This indicates the primary respiratory frequency after the current time window update. This indicates the dominant respiratory frequency in the previous time window. For smoothing, a value of 0.6 to 0.9 is preferred. Through this method, the feedback notch filter can adaptively adjust the notch center frequency and notch bandwidth according to changes in the respiratory fundamental frequency, thereby suppressing the respiratory fundamental frequency and its harmonic components while preserving weaker vital activity components such as heartbeat.

[0044] By combining the above-mentioned multi-channel observation, blind source separation and feedback notch filtering, the target's life activity components can be separated from platform disturbances and background clutter, resulting in an enhanced slow-time complex echo sequence.

[0045] like Figure 2 As shown, after obtaining the enhanced slow-time complex echo sequence, time-frequency analysis of the slow-time complex echo sequence can be performed, which can be expressed by the short-time Fourier transform as follows: ; in, For the first The slow-time complex echo sequence of the candidate targets in time ,frequency The short-time Fourier transform results at the specified location are used to characterize the micro-Doppler time-frequency distribution of the candidate target within the local time window; Indicates the first Slow-time complex echo sequences of candidate targets; Indicates slow time; This represents the sliding window function. By performing a short-time Fourier transform on the slow-time complex echo sequence, the micro-Doppler time-frequency characteristic map of the candidate target is obtained.

[0046] When the candidate target has a slight velocity along the radar line of sight At that time, its micro-Doppler frequency satisfies: ; in, Indicates that the candidate target is in slow time The micro-Doppler frequency caused by the slight motion velocity; Indicates that the candidate target is in slow time The minute velocity along the radar line of sight; This indicates the radar's operating wavelength. Micro-displacements in the chest cavity caused by breathing and heartbeat will appear as low-frequency, weak-amplitude, periodic textures in the time-frequency diagram. Since inanimate background scatterers typically lack stable periodic structures, this difference can be used to improve the ability to detect living organisms.

[0047] To reduce the overlapping effects of different micro-motion components in a single time-frequency plane, this embodiment not only performs a short-time Fourier transform on the center distance gate, but also performs time-frequency analysis on multiple adjacent distance gates above and below the center distance gate, stacking them along the distance dimension to form a time-distance-Doppler three-dimensional data cube. This three-dimensional data cube simultaneously characterizes the response distribution of candidate targets in the time, distance, and Doppler dimensions. Based on this, a three-dimensional constant false alarm rate (CFAR) detection module is used to perform three-dimensional CFAR detection on the time-distance-Doppler data cube. Specifically, for each voxel, the average background energy outside its neighborhood protection unit is calculated, and a detection threshold is set according to the CFAR principle. When the voxel energy is higher than the threshold, the voxel is retained; otherwise, it is set to zero. After three-dimensional CFAR detection processing, significant micro-motion component distribution results are obtained, and a micro-Doppler-distance joint characterization map is generated based on this. This micro-Doppler-distance joint characterization map reflects both the periodic variation characteristics of human life activities in the frequency domain and preserves the differences in scattering distribution at different distance gates, thereby enhancing the distinguishability between living targets and false targets.

[0048] The time-distance-Doppler data cube constructed in this method introduces the distance dimension as an independent dimension into micromotion analysis. Chest micromotions caused by breathing and heartbeat in real humans exhibit a continuous and correlated energy distribution between the central distance gate and its adjacent distance gates; while the micromotions of background pseudo-targets (such as vegetation swaying in the wind, metal components, etc.) often concentrate energy at a single distance gate or exhibit discontinuous jumps along the distance dimension. By extracting significant micromotion components through 3D CFAR, the distance dimension information can be effectively used to distinguish between real life micromotions and background pseudo-micromotions, reducing misjudgments caused by overlapping micromotion components in a single time-frequency plane.

[0049] After constructing the micro-Doppler-range joint representation map, the SAR image patches of the candidate targets and the micro-Doppler-range joint representation map are simultaneously input into the state-space model fusion network, such as... Figure 3 As shown, the state-space model fusion network includes a SAR spatial domain branch, a life activity time-frequency domain branch, and a cross-domain fusion module. Specifically, the SAR spatial domain branch first embeds features from SAR image patches of candidate targets to obtain initial spatial scattering structure features. Then, these initial spatial scattering structure features are expanded into sequences along four directions: forward horizontal direction, reverse horizontal direction, forward vertical direction, and reverse vertical direction. These sequences are then input into the Mamba state-space modeling unit (SSM) for long-range context modeling to obtain deep spatial scattering structure features. .

[0050] The time-frequency domain branch for life activities first encodes the micro-Doppler-distance joint representation map to obtain initial time-frequency features of life activities. Then, it performs global temporal dependency modeling on these initial features: an outer sequence is constructed along the slow time direction to model the global periodic dependency of respiration and heartbeat throughout the observation period. Next, it performs local time-frequency texture modeling: an inner sequence is constructed within each slow time segment along a local frequency distance window to model local time-frequency texture changes and the coupling relationship between adjacent distance gates. Finally, nested spatial modeling (Nested SSM) is performed to obtain deep time-frequency features of life activities. The state-space modeling unit corresponding to the outer sequence in the time-frequency domain branch for life activities is called the outer Mamba module, used to model the global periodic dependency of respiration and heartbeat in the observation time dimension; the state-space modeling unit corresponding to the inner sequence is called the inner Mamba module, used to model the coupling relationship between weak spectral peaks, local textures, and adjacent distance gates within the local frequency distance window. The outer and inner Mamba modules together constitute a nested Mamba structure.

[0051] The state-space modeling process satisfies: ; ; in, Indicates input features, This indicates a hidden state. Indicates output features, , , , This is a matrix of learnable parameters related to the input.

[0052] Deep features of spatial scattering structure output by the two branches after state-space modeling and deep characteristics of life activities The fusion features are obtained by splicing channels: ; in, This represents the deep spatial scattering structure features extracted by the SAR spatial domain branch. This represents the deep time-frequency features of life activities extracted from the time-frequency domain branch of life activities.

[0053] The fused features are input into a cross-domain fusion module, which includes a region-aware attention module and a gated recursive context refinement module. The region-aware attention module generates spatial attention weights and channel attention weights based on the spatial response intensity and salience of vital activities in candidate regions, performing feature weighting enhancement on candidate target features and suppressing background region features. The gated recursive context refinement module recursively fuses high-level semantic fusion features with low-level boundary detail features. In each round of recursive updates, the retention weights of high-level semantic fusion features and low-level boundary detail features are adjusted through a gating unit, thereby highlighting areas with salience of vital activities while preserving local scattering details near the boundary of the wounded target, ultimately forming a unified semantic representation fusion feature that combines boundary localization and liveness detection capabilities.

[0054] Based on unified semantic representation fusion features, a multi-task output network is constructed to perform target detection, target segmentation, liveness detection, and geolocation. The detection branch outputs the bounding boxes and confidence scores of candidate targets; the segmentation branch outputs the segmentation mask for the injured target region; the liveness detection branch outputs continuous score values ​​representing the intensity of the target's liveness; and the geolocation branch combines the center of the detection box, the geometric center of the segmented region, and the imaging geometry to output the target's position parameters in the geographic coordinate system. To ensure the network uniformly learns the capabilities of injured target detection, boundary delineation, liveness detection, and geolocation during training, the following multi-task training loss function is adopted: ; Where L is the total loss, For target detection loss, To segment the loss for the target, Score the loss of vital functions. The consistency constraint loss between the deep features of spatial scattering structures and the deep features of time-frequency biological activities is used. To pinpoint the loss, These are the weighting coefficients for the corresponding loss terms. In this embodiment, the weighting coefficients... ~ All are non-negative numbers, used to balance the contributions of each loss term in joint training. Preferably, Take a value of 0.8 to 1.5. Take a value of 0.5 to 1.2. Take a value of 0.8 to 1.5. Take a value of 0.1 to 0.6. Take a value of 0.3 to 1.0.

[0055] The weighting coefficients are determined based on the magnitude of each loss term, the importance of the task, and the evaluation results of the validation set. Specifically, when the focus is on the accuracy of casualty detection, the weighting coefficients are increased. The value of ; when focusing on the delineation of the wounded's boundaries, improve The value of ; when focusing on the ability to detect liveness, improve The value of ; when focusing on cross-domain feature consistency constraints, improve The value of ; when focusing on geographic positioning accuracy, improve The value of .

[0056] The object detection loss consists of object classification loss and bounding box regression loss. The object classification loss enhances the constraint on positive and hard-to-classify samples, while the bounding box regression loss constrains the positional consistency between the predicted and ground truth boxes. The object segmentation loss uses the Dice loss, whose expression is: ; in, Indicates the predicted segmentation map at the pixel level. The output value at that location, This indicates the corresponding real label. It is a very small constant; In this embodiment, the vital activity scoring label The life activity scoring label is used to characterize the strength of the life activity signals of candidate targets, and its value ranges from 0 to 1. The physiological signals synchronously acquired by the contact-type physiological sensor are used as the primary calibration basis, supplemented by radar micro-motion feature normalization results for auxiliary verification. The contact-type physiological sensor includes a breathing belt and a heart rate sensor. The breathing belt is used to acquire chest rise and fall signals of the test target and obtain respiratory stability information, while the heart rate sensor is used to acquire heartbeat signals of the test target and obtain heartbeat stability information. The contact-type physiological sensor is only used in the tagging and calibration phase of training samples; it is not required for the target to wear the contact sensor during actual search and rescue operations.

[0057] During the training data acquisition phase, the chest cavity fluctuation signals acquired by the breathing belt, the heartbeat signals acquired by the heart rate sensor, and the SAR echo data from the verification mode were synchronized in time. For the first... Based on training samples, and according to respiratory stability, heart rate stability, and duration of effective vital signs within the corresponding time window, a real vital activity score is generated. When both respiratory and heartbeat signals are stable, Take the higher value; when only one of the respiratory or heartbeat signals is detected, Take the median value; when no stable vital signs are detected, Take the lower value or 0.

[0058] Preferably, real life activity score Determine as follows: ; in, Indicates the first The respiratory stability score of each training sample within the corresponding time window. Indicates heart rate stability score. The score indicates the duration of effective vital signs. This means limiting the calculation result to the range of 0 to 1. , , This represents the weight of the scoring item. Preferably, , , All are non-negative numbers and satisfy the following conditions: .

[0059] Furthermore, based on the SAR echo data from the verification mode, the energy of the respiratory principal component, the energy of the heartbeat candidate component, the periodic stability of vital activity components, and the signal-to-noise ratio are extracted. These radar micro-motion characteristics are then normalized to obtain a radar micro-motion auxiliary score. This radar micro-motion auxiliary score is used to verify the consistency of vital activity score labels obtained from contact physiological sensor calibration, or as a supplementary calibration basis when some training samples lack contact physiological sensor records.

[0060] Therefore, each training sample can obtain a true life activity score for calculating the life activity score loss. .

[0061] Based on real life activity scores The loss for vital activity scoring uses the mean squared error loss, which is expressed as follows: ; in, Indicates the first Predicted vital activity scores for each sample Indicates the first The true vital activity score of each sample, where N is the total number of samples; the localization loss is calculated based on the relationship between location vectors. Distance constraint, its expression is: ; in, Indicates the first The predicted geographic location vector of each sample. Indicates the first The true geographical location vector of each sample.

[0062] During training, each input sample carries a target detection bounding box label, a target segmentation mask label, a vital activity score label, and a geographic location label. During forward propagation, the detection branch determines the presence of a wounded target in a candidate region based on fused features and regresses its bounding box; the segmentation branch outputs the target region mask; the vital activity score branch outputs a vital activity score based on the combined response of deep spatial scattering structure features, deep temporal features of vital activity, and fused features of unified semantic representation; and the localization branch regresses the target's geographic coordinates based on the detection results, segmented regions, and imaging geometry. All loss terms are weighted and summed before backpropagation, enabling the network to learn the capabilities of wounded target detection, boundary delineation, vital activity discrimination, and geographic localization while sharing feature representations.

[0063] In practical applications, the UAV platform performs wide-area search missions over disaster areas, generating SAR images of the disaster area and completing preliminary candidate target screening. Subsequently, hovering verification is performed on the selected candidate targets, acquiring long-stay SAR echo data in verification mode. After joint discrimination by a state-space model fusion network, the location, boundary, confidence level, and vital signs score of the injured target are output. Based on the confidence level, vital signs score, and target distance, rescue priority information is generated. In this embodiment, for the first... Each suspected injured person is assigned a priority score for rescue. Represented as: ;

[0064] in, Indicates the first The confidence level of detection for a suspected wounded target. Represents the distance priority factor. , , Let be the weight coefficient, and satisfy... Distance priority factor The distance is determined based on the distance between the suspected injured person and the rescue starting point or drone platform; the closer the distance, the better. The larger the value, the better.

[0065] Score based on rescue priority Classify suspected wounded targets. When At that time, it was marked as a first-priority rescue target; when At that time, it was marked as a secondary priority rescue target; when At that time, the targets are marked as pending review or low priority targets. Ground control stations or mobile terminals display alarms and prioritize suspected casualty targets based on the rescue priority score and rescue priority level, and provide location guidance and rescue order reference for the rescue team. Another embodiment of the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it implements the aforementioned micro-Doppler-range-based UAV SAR casualty detection and liveness detection method.

[0066] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the aforementioned method for detecting and determining liveness of casualties using UAV SAR based on microDoppler-range.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements and adjustments made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for liveness detection of wounded personnel using UAV SAR based on microDoppler-range, characterized in that, include: The raw SAR echo data and platform positioning attitude data of the target area are acquired simultaneously using the UAV platform, and a SAR image of the target area is generated in wide-area search mode. Based on the SAR image, a candidate target screening network based on physical prior constraints is used to screen suspected wounded targets, obtaining candidate targets and their coarse localization results. The candidate target screening network based on physical prior constraints includes a two-dimensional constant false alarm rate (CFAR) initial screening module, a scattering geometry prior screening module, and a convolutional feature secondary screening module. The two-dimensional CFAR initial screening module extracts initial candidate regions from the SAR image; the scattering geometry prior screening module performs prior constraint screening on the initial candidate regions based on their statistical characteristics; and the convolutional feature secondary screening module performs rapid secondary discrimination on the candidate regions after prior constraint screening and outputs the candidate target confidence score. The candidate target screening network selects suspected wounded candidate targets that meet the requirements based on the comprehensive score of the candidate regions and outputs their coarse localization results. The comprehensive score of the candidate regions is expressed as: ； in, This represents the overall score of the candidate regions. This represents the score of the two-dimensional constant false alarm rate (CFAR) detection. This represents the prior score of the scattering geometry. This represents the score from the convolutional feature re-screening. , , These are the weighting coefficients for each score item; The control drone platform is switched to verification mode, hovering and staring above the candidate target to acquire verification mode SAR echo data of the candidate target; Based on the aforementioned verification mode SAR echo data, the candidate targets and their coarse positioning results are mapped to the range gates of the verification mode SAR echo data to determine the center range gate and its adjacent range gates corresponding to the candidate targets; within the center range gate and its adjacent range gates, slow-time complex echo sequences are extracted. Time-frequency analysis was performed on the slow-time complex echo sequence, and a time-range-Doppler data cube was formed by stacking along the distance dimension. A micro-Doppler-distance joint characterization map was constructed based on the time-range-Doppler data cube. The SAR image patch and the micro-Doppler-range joint representation map of the candidate target are simultaneously input into the state space model fusion network for joint feature extraction and state space modeling, resulting in a unified semantic representation fusion feature that combines boundary localization capability and liveness detection capability. Based on unified semantic representation fusion features, suspected wounded targets are detected, segmented, liveness-based and located, and output the location, boundary, confidence level, vital activity score and rescue priority information of the wounded targets.

2. The UAV SAR wounded person detection and liveness determination method according to claim 1, characterized in that, First, envelope alignment and phase compensation are performed on the SAR echo data in the verification mode. Then, a multi-channel observation matrix is ​​constructed using the center range gate corresponding to the candidate target and the slow-time complex echo sequences within the adjacent range gates. The multi-channel observation matrix is ​​then whitened to decouple the signals of each channel in a second-order statistical sense. Subsequently, a higher-order statistical matrix of the whitened signal is constructed, and joint approximate diagonalization is performed to decompose the mixed signal into several independent source components. By performing frequency domain analysis and time domain continuity analysis on the independent source components, the component with stable waveform within the breathing frequency range is selected as the main breathing component. Then, a feedback notch filter is constructed using this main breathing component as a reference signal to perform narrowband suppression on the sequence containing vital activity components, resulting in an enhanced slow-time complex echo sequence.

3. The UAV SAR wounded person detection and liveness determination method according to claim 2, characterized in that, A three-dimensional constant false alarm rate (CFAR) detection module was used to perform three-dimensional CFAR detection on the time-distance-Doppler data cube, and significant micro-motion component distribution results were obtained. Based on the significant micro-motion component distribution results, a micro-Doppler-distance joint characterization map was generated.

4. The UAV SAR wounded person detection and liveness determination method according to claim 1, characterized in that, The state space model fusion network includes a SAR spatial domain branch, a life activity time-frequency domain branch, and a cross-domain fusion module. The SAR spatial domain branch is used to embed features into SAR image patches of candidate targets to obtain initial features of spatial scattering structure, and to perform state-space modeling on the initial features of spatial scattering structure to obtain deep features of spatial scattering structure; the life activity time-frequency domain branch is used to encode features into the micro-Doppler-range joint characterization map of candidate targets to obtain initial features of life activity time-frequency, and to perform state-space modeling on the initial features of life activity time-frequency to obtain deep features of life activity time-frequency; the deep features of spatial scattering structure and the deep features of life activity time-frequency are concatenated to obtain fused features; The fusion features are processed by the cross-domain fusion module to form a unified semantic representation fusion feature that combines boundary localization capability and liveness detection capability.

5. The UAV SAR wounded person detection and liveness determination method according to claim 4, characterized in that, The SAR spatial domain branch first embeds features into the SAR image blocks of the candidate targets to obtain initial features of the spatial scattering structure. Then, the initial features of the spatial scattering structure are expanded into sequences along four directions: forward horizontal direction, reverse horizontal direction, forward vertical direction, and reverse vertical direction. These sequences are then input into the Mamba state space modeling unit for long-range context modeling to obtain deep features of the spatial scattering structure. The time-frequency domain branch of the life activity first encodes the micro-Doppler-distance joint representation map to obtain the initial time-frequency features of life activity; then, it performs global temporal dependency modeling on the initial time-frequency features of life activity: an outer sequence is constructed along the slow time direction to model the global periodic dependency of respiration and heartbeat throughout the observation period; then, it performs local time-frequency texture modeling: an inner sequence is constructed along the local frequency distance window within each slow time segment to model the coupling relationship between local time-frequency texture changes and adjacent distance gates; then, it performs nested space modeling to obtain the deep time-frequency features of life activity.

6. The UAV SAR wounded person detection and liveness determination method according to claim 5, characterized in that, The region-aware attention module generates spatial attention weights and channel attention weights based on the spatial response intensity and salience of life activities in the candidate region, thereby enhancing the features of the candidate target by weighting and suppressing the features of the background region. The gated recursive context refinement module recursively fuses high-level semantic fusion features with low-level boundary detail features. In each round of recursive update, the gated unit adjusts the retention weights of high-level semantic fusion features and low-level boundary detail features, ultimately forming a unified semantic representation fusion feature that combines boundary localization and liveness detection capabilities.

7. The UAV SAR wounded person detection and liveness determination method according to claim 1, characterized in that, A multi-task output network is used to jointly discriminate the fused features, outputting the casualty detection result, casualty segmentation result, vital signs score, and geolocation result respectively; the following multi-task training loss function is used during training: ； Where L is the total loss, For target detection loss, To segment the loss for the target, Score the loss of vital functions. The consistency loss is constrained between the deep features of spatial scattering structures and the deep features of time-frequency biological activities. To pinpoint the loss, These are the weighting coefficients for the corresponding loss terms.

8. An electronic device comprising a processor, a memory, and a computer program; wherein, The processor is connected to the memory, and the computer program is stored in the memory. The electronic device is characterized in that, when the electronic device is running, the processor executes the computer program to cause the electronic device to perform the UAV SAR wounded person detection and liveness determination method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the processor performs the UAV SAR wounded person detection and liveness determination method as described in any one of claims 1 to 7.

Images