Vital sign perception method and system based on time sequence infrared image

Abstract

The invention relates to the technical field of vital sign monitoring, and provides a vital sign sensing method and system based on a time sequence infrared image, and the method comprises the steps: obtaining a disaster rescue scene time sequence infrared image, and carrying out the brightness contrast standardization; short-time and long-time features are separated and extracted through a dual-channel convolutional network, and a weight coefficient is calculated according to ruin temperature variance for adaptive fusion; performing multi-domain noise detection suppression and physiological signal frequency filtering reconstruction on the time sequence characteristics; the image material type is detected, and temperature attenuation is compensated based on the heat conduction parameter database; a three-dimensional space heat distribution model is constructed in combination with multi-frame information, and life body candidate areas are screened; and correlating and matching the space coordinates of the candidate positions with the reconstructed time sequence characteristic physiological signals, and outputting a sensing result. According to the method, the heat conduction attenuation effect is accurately compensated according to the ruin material distribution characteristics, and the accuracy of vital sign perception of trapped persons in a disaster rescue scene is improved.

Description

Technical Field

[0001] This invention relates to the field of vital sign monitoring technology, and in particular to a method and system for sensing vital signs based on time-series infrared images. Background Technology

[0002] Infrared thermal imaging technology is a non-contact method for temperature measurement and vital sign monitoring based on the thermal radiation characteristics of objects. In disaster relief scenarios, infrared thermal imaging technology can penetrate harsh environmental conditions such as smoke and dust to detect weak thermal radiation signals from people trapped deep in rubble. However, the complex environmental conditions at disaster sites, the thermal conduction interference from various materials, and the weak physiological signals of trapped people pose many challenges to the perception of vital signs in infrared images. The key lies in how to effectively eliminate the thermal conduction interference from rubble materials, accurately extract weak physiological signal features, and achieve precise three-dimensional positioning of trapped people.

[0003] In existing technologies, vital sign detection based on infrared images mainly employs traditional image processing and basic signal analysis methods to achieve basic body temperature detection and coarse vital sign monitoring functions. However, existing methods do not adequately consider the inherent physical correlation mechanism between complex ruin structures and the propagation of infrared thermal radiation signals in disaster relief scenarios. They struggle to organically integrate objective physical constraints of heat conduction with the actual distribution characteristics of ruin materials, resulting in the inability to achieve high-precision and high-reliability automatic perception and accurate location of the vital signs of trapped personnel. Summary of the Invention

[0004] In view of this, the present invention proposes a vital signs perception method and system based on time-series infrared images, which solves the problem that existing methods do not adequately consider the inherent physical correlation mechanism between complex ruin structures and infrared thermal radiation signal propagation in disaster relief scenarios, and are unable to organically integrate objective physical constraints of heat conduction with actual ruin material distribution characteristics, resulting in the inability to achieve high-precision and high-reliability automatic perception and accurate positioning of the vital signs of trapped personnel.

[0005] The technical solution of this invention is implemented as follows: On one hand, this invention provides a method for sensing vital signs based on time-series infrared images, comprising the following steps: A time-series infrared image sequence of a disaster relief scene is acquired, and the brightness and contrast of the time-series infrared image sequence are normalized using an image processor to obtain a normalized time-series infrared image. The standardized temporal infrared image is separated and processed by a dual-channel convolutional network model to obtain short-time features and long-time features. The dual-channel weight coefficients are calculated based on the variance of the ruin temperature detected by the temperature sensor, and the short-time features and long-time features are weighted and fused to obtain temporal features. Multi-domain noise detection and suppression are performed on the time-series features to identify the frequency components of impulse noise. Filtering is then performed in conjunction with the frequency range of physiological signals to obtain the reconstructed time-series features. The material types in the standardized time-series infrared images are detected and classified. The thermal conduction attenuation coefficient is calculated based on a preset material thermal conduction parameter database. The image temperature values ​​are then compensated and corrected to obtain a compensated time-series infrared image. Spatial structure analysis is performed on the compensated temporal infrared image, and a three-dimensional spatial heat distribution model is constructed by combining multi-frame temporal information. Candidate regions of living organisms are screened according to the heat source intensity threshold to obtain candidate locations. The spatial coordinate information of the candidate location is correlated and matched with the physiological signal information of the reconstructed temporal features to output the vital sign perception result.

[0006] Based on the above technical solutions, preferably, the separation processing of the standardized temporal infrared image using a dual-channel convolutional network model includes: The dual-channel convolutional network model includes a fast-channel convolutional module, a slow-channel convolutional module, and a cascaded attenuation compensation module; The standardized temporal infrared image is input into the fast channel convolution module, and the time window of the fast channel is set to 1 to 10 seconds. Short-period physiological signal features in the image are extracted to obtain short-time features. The standardized temporal infrared image is input into the slow channel convolution module, and the time window of the slow channel is set to 1 to 6 hours. Long-period background signal features in the image are extracted to obtain long-time features. A cascaded attenuation compensation module is used to perform inverse signal attenuation compensation for both short-time and long-time characteristics.

[0007] Based on the above technical solutions, preferably, the calculation of the dual-channel weight coefficients and the weighted fusion of short-term and long-term features include: Temperature sensors are used to detect the temperature values ​​at multiple monitoring points in the ruins environment in real time, and the variance of the temperature values ​​is calculated as the temperature variance of the ruins. The environmental stability level is determined based on the magnitude of the temperature variance of the ruins. When the temperature variance of the ruins exceeds the preset threshold, the weight coefficient of the fast channel is set to a high weight value and the weight coefficient of the slow channel is set to a low weight value. The short-term features are multiplied by the fast channel weight coefficient, and the long-term features are multiplied by the slow channel weight coefficient. The weighted features are then numerically summed to obtain the time-series features.

[0008] Based on the above technical solutions, preferably, the detection and classification of material types in the standardized time-series infrared image and the compensation and correction of image temperature values ​​include: The materials of the ruins in the standardized time-series infrared images are classified at the pixel level to identify the material types, which include concrete, steel bars, bricks, wood and soil, and a material distribution map is obtained. Based on the spatial distribution of each material type in the material distribution map, the corresponding thermal conductivity coefficient, thermal diffusivity and heat capacity parameters are extracted from the material thermal conductivity parameter database. The thermal bridge modeling algorithm is used to calculate the thermal conduction coupling effect of multi-material interfaces, and the temperature attenuation degree of different material regions is quantitatively analyzed to generate a thermal conduction attenuation coefficient matrix. The thermal conduction attenuation coefficient matrix is ​​multiplied pixel-by-pixel with the standardized time-series infrared image to obtain the compensated time-series infrared image.

[0009] Based on the above technical solutions, preferably, the calculation of the multi-material interface thermal conduction coupling effect using the interface thermal bridge modeling algorithm includes: Identify the boundary pixel positions between different materials in the material distribution map and establish the topological connection relationship of the material interface; Based on the variation law of thermophysical parameters of various materials under disaster environment recorded in the specific material library of ruins, the effective thermal conductivity at the interface is calculated; The heat transfer path and attenuation law at the material interface are analyzed by using a multilayer coupled heat transfer model. The influence of material thickness, density distribution and porosity on heat conduction is considered to obtain the heat conduction attenuation coefficient matrix.

[0010] Based on the above technical solutions, preferably, the step of performing spatial structure analysis on the compensated temporal infrared image and constructing a three-dimensional spatial heat distribution model by combining multi-frame temporal information includes: The compensated temporal infrared image is subjected to multi-scale spatial feature extraction using a spatiotemporally coupled convolutional processor, and spatiotemporal correlation analysis is performed by combining the temperature change gradient information of adjacent time frames. The thermal tomography algorithm is used to vertically stitch together the heat conduction information of different depth layers to construct a four-dimensional spatiotemporal data structure containing spatial coordinates, temperature values ​​and time dimensions. The four-dimensional spatiotemporal data structure is converted into a three-dimensional spatial heat distribution model using a three-dimensional reconstruction algorithm. A threshold range for heat source intensity is set, and spatial clustering analysis is performed on heat source regions that exceed the threshold to screen out candidate regions that conform to the thermal radiation characteristics of living organisms, thus obtaining candidate locations.

[0011] Based on the above technical solutions, preferably, the step of vertically stitching together the heat conduction information of different depth layers using a thermal tomography connection algorithm to construct a four-dimensional spatiotemporal data structure containing spatial coordinates, temperature values, and time dimensions includes: The compensated temporal infrared image is divided into multiple thermal conduction layers according to the spatial depth direction, and each thermal conduction layer corresponds to a different material occlusion depth range. The temperature transfer relationship between adjacent layers is calculated using the interlayer heat conduction attenuation model, and the heat conduction connection matrix between layers is established. Based on the physical laws of heat conduction, the temperature distribution at each level is reconstructed in reverse to compensate for the temperature attenuation loss caused by material shading. The temperature data at each level after inverse reconstruction are registered in three-dimensional space according to depth order to obtain a four-dimensional spatiotemporal data structure, which includes location coordinates, temperature intensity, material type and timestamp.

[0012] Based on the above technical solutions, preferably, the step of performing multi-domain noise detection and suppression on the time-series features, identifying the frequency components of impulse noise, and performing filtering processing in conjunction with the frequency range of physiological signals to obtain the reconstructed time-series features includes: The time-series characteristics are subjected to multi-domain analysis processing in the time domain, frequency domain, and time-frequency domain. The noise type is matched and identified using a physical simulation noise template library. The frequency components and amplitude characteristics of non-periodic impulse noise are detected through an impulse detection gating mechanism to obtain the noise distribution characteristics. Based on the noise distribution characteristics, adaptive filter parameters are designed, and a filter passband is established in combination with the preset physiological signal frequency range. The time series characteristics are then subjected to frequency domain filtering, and weak signals are reconstructed using the heart rate-respiration coupling constraint to obtain the reconstructed time series characteristics.

[0013] Based on the above technical solutions, preferably, the step of associating and matching the spatial coordinate information of the candidate location with the physiological signal information of the reconstructed temporal features to output the vital sign perception result includes: The spatial-signal correlation matching module is used to perform spatiotemporal correspondence analysis on the three-dimensional spatial coordinate information of the candidate positions and the physiological signal information of the reconstructed temporal features. The correlation score between each candidate position and the corresponding physiological signal intensity is calculated by the distance weighting algorithm, and a mapping relationship table between candidate positions and physiological signals is established. Threshold filtering is performed based on the correlation scores in the mapping table. Valid life forms with correlation scores exceeding a preset threshold are extracted. Spectral analysis is performed on the physiological signals of the valid life forms to extract respiratory rate parameters. Combined with spatial coordinates and signal intensity information, vital sign perception results are generated. The vital sign perception results include vital sign intensity, respiratory rate, and precise three-dimensional location.

[0014] On the other hand, the present invention also provides a vital sign sensing system based on time-series infrared images, the system comprising: The image acquisition and preprocessing module is used to acquire a time-series infrared image sequence of a disaster relief scene, and to use an image processor to perform brightness and contrast standardization processing on the time-series infrared image sequence to obtain a standardized time-series infrared image. The dual-channel feature extraction and fusion module is used to separate and process the standardized temporal infrared image through a dual-channel convolutional network model to obtain short-time features and long-time features. Based on the variance of the ruin temperature detected by the temperature sensor, the dual-channel weight coefficients are calculated, and the short-time features and long-time features are weighted and fused to obtain temporal features. The multi-domain noise suppression and signal reconstruction module is used to perform multi-domain noise detection and suppression on the time-series features, identify the frequency components of impulse noise, and perform filtering processing in combination with the frequency range of physiological signals to obtain the reconstructed time-series features. The material identification and thermal conduction compensation module is used to detect and classify the material types in the standardized time-series infrared image, calculate the thermal conduction attenuation coefficient according to the preset material thermal conduction parameter database, and compensate and correct the image temperature value to obtain a compensated time-series infrared image. The three-dimensional thermal distribution modeling and candidate region screening module is used to perform spatial structure analysis on the compensated temporal infrared image, construct a three-dimensional spatial thermal distribution model by combining multi-frame temporal information, and screen candidate regions of living organisms according to the heat source intensity threshold to obtain candidate locations. The spatial signal association matching and result output module is used to associate and match the spatial coordinate information of the candidate location with the physiological signal information of the reconstructed temporal features, and output the vital sign perception result.

[0015] The present invention provides a vital sign sensing method and system based on time-series infrared images, which has the following advantages over existing technologies: (1) By integrating a dual-channel convolutional network model with material thermal conduction compensation, a spatiotemporal coupled convolutional processor is used to extract multi-scale spatial features and model three-dimensional thermal distribution. Combined with multi-domain noise detection and adaptive filtering, physiological signals and environmental interference are dynamically separated. The thermal conduction attenuation effect is accurately compensated according to the material distribution characteristics of the ruins, which improves the accuracy of the perception of vital signs of trapped people in disaster relief scenarios. (2) By integrating the dual-channel convolutional network model with cascaded attenuation compensation, the fast and slow dual channels are used to extract short-cycle physiological signal features and long-cycle background signal features respectively. The dual-channel weight coefficients are dynamically adjusted in combination with the temperature variance of the ruin environment, and the weight allocation strategy is adaptively adjusted according to the environmental stability level. This improves the accuracy of temporal feature extraction in complex ruin environments. At the same time, the adaptive weight fusion strategy meets the actual needs of weak physiological signal detection under different environmental conditions. (3) By integrating the material identification module and the interface thermal bridge modeling algorithm, the pixel-level material classification method is used to identify the material type of the ruins and extract the thermal conduction parameters. The change law of thermophysical parameters is dynamically adjusted by combining the specific material library of the ruins. The topological connection relationship of the material interface is accurately modeled according to the multi-layer coupled heat transfer model, which improves the accuracy of the perception of the body temperature distribution of the obscured personnel. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a vital signs perception method based on time-series infrared images according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a method for sensing vital signs based on time-series infrared images, comprising the following steps: A time-series infrared image sequence of a disaster relief scene is acquired, and the brightness and contrast of the time-series infrared image sequence are normalized using an image processor to obtain a normalized time-series infrared image. The standardized temporal infrared image is separated and processed by a dual-channel convolutional network model to obtain short-time features and long-time features. The dual-channel weight coefficients are calculated based on the variance of the ruin temperature detected by the temperature sensor, and the short-time features and long-time features are weighted and fused to obtain temporal features. The timing features are subjected to multi-domain noise detection and suppression using a signal analysis processor, the frequency components of impulse noise are identified by a spectrum analyzer, and filtering is performed in combination with the frequency range of physiological signals to obtain the reconstructed timing features. The material type in the standardized time-series infrared image is detected and classified using a material identification sensor. The thermal conduction attenuation coefficient is calculated based on a preset material thermal conduction parameter database. The image temperature value is then compensated and corrected to obtain a compensated time-series infrared image. The spatial structure of the compensated temporal infrared image is analyzed using a 3D reconstruction processor. A 3D spatial heat distribution model is constructed by combining multi-frame temporal information. Candidate regions of living organisms are screened based on heat source intensity thresholds to obtain candidate locations. The spatial coordinate information of the candidate location is associated and matched with the physiological signal information of the reconstructed temporal features using a data fusion processor, and the vital sign perception results are output. The vital sign perception results include vital sign intensity, respiratory rate and precise three-dimensional location.

[0020] Specifically, this embodiment integrates a dual-channel convolutional network model with material thermal conductivity compensation. It utilizes a spatiotemporally coupled convolutional processor for multi-scale spatial feature extraction and three-dimensional thermal distribution modeling. Multi-domain noise detection and adaptive filtering dynamically separate physiological signals from environmental interference, and precisely compensate for thermal conductivity attenuation effects based on the material distribution characteristics of the ruins. This embodiment solves the problem of balancing weak physiological signal extraction and accurate spatial positioning in complex ruin environments through a spatial-signal correlation matching mechanism, improving the accuracy of vital sign perception for trapped personnel in disaster relief scenarios.

[0021] The acquisition of the time-series infrared image sequence of the disaster relief scene involves using an image processor to perform brightness and contrast normalization processing on the time-series infrared image sequence to obtain a normalized time-series infrared image, including: Infrared thermal imagers are used to perform multi-angle scanning in disaster relief scenarios. Image acquisition frequency and spatial resolution parameters are set to obtain original temporal infrared image sequences, which include the background of the ruins and the area where potential trapped personnel may be located.

[0022] In one specific embodiment, the use of an infrared thermal imager to perform multi-angle scanning in a disaster relief scenario includes: The infrared thermal imager's scanning trajectory is set as a spiral path covering the target ruin area. The scanning interval distance is dynamically adjusted according to the spatial complexity of the ruins. Continuous time-series image frames are acquired at each scanning position, and the corresponding spatial coordinates and timestamp information are recorded simultaneously to obtain the original time-series infrared image sequence containing three-dimensional information of space, time, and temperature.

[0023] The original time-series infrared image sequence is preprocessed using an image processor. Sensor noise is eliminated through bad pixel detection and repair. The image brightness is normalized according to the ambient temperature reference value. The image contrast is adaptively adjusted according to the reflective characteristics of the ruin material to obtain a standardized time-series infrared image.

[0024] In one specific embodiment, the preprocessing of the original temporal infrared image sequence using an image processor includes: The method of neighboring pixel difference analysis is used to detect bad pixels with abnormal high or low temperatures. The bad pixels are repaired by bilinear interpolation. A dynamic temperature mapping range is established based on the lowest and highest temperature values ​​of the ruin environment. The image brightness value is mapped to the standard range. A contrast enhancement matrix is ​​established according to the difference in infrared emissivity of different materials. The image is then adaptively adjusted for contrast.

[0025] In one specific embodiment, infrared image acquisition and preprocessing in an earthquake rubble environment includes: At the site of an earthquake rescue operation, the rubble area was approximately 50 meters long, 30 meters wide, and 15 meters high, mainly formed by the collapse of concrete structures. Multi-angle scanning was performed using an infrared thermal imager, with the specific parameters set as follows:

[0026] Scan trajectory settings: The infrared thermal imager was mounted on the drone platform, and a spiral scanning path was set. The initial height was 20 meters above the ruins, and the spiral radius gradually increased from 5 meters to 25 meters. The height decreased by 2 meters with each rotation, and the scanning interval was set to 3 meters. Considering the high complexity of the ruins, the scanning interval was shortened to 1.5 meters in areas with dense building structures.

[0027] Image acquisition parameters: The image acquisition frequency is set to 10 frames / second, the spatial resolution is 640×480 pixels, the temperature measurement range is -20℃ to +60℃, and the temperature accuracy is ±0.1℃. 30 frames of images are continuously acquired at each scanning position, and GPS coordinates (accuracy set to ±1 meter) and UTC timestamps (accuracy set to ±1 millisecond) are recorded simultaneously.

[0028] The preprocessing process includes: Defective pixel detection and repair: A 3×3 neighborhood window is used for pixel difference analysis, with a threshold set at ±5℃. For example, if a pixel is detected with a temperature of 45℃, while the average temperature of its eight neighboring pixels is 18℃, it is determined to be a high-temperature defective pixel. Using bilinear interpolation, the repaired temperature value of this defective pixel is:

[0029] Brightness normalization: Temperature sensors detected the lowest temperature in the ruins environment to be -5℃ and the highest to be 35℃. A dynamic temperature mapping range of 40℃ was established. The original image brightness values ​​(0-255) were mapped to the standard range (0-1) using the following formula: For example, when a pixel temperature is 20°C, the standardized brightness value is... .

[0030] Adaptive Contrast Adjustment: Based on a material thermal conductivity parameter database, the emissivity of concrete is 0.95, and that of steel reinforcement is 0.85. A contrast enhancement matrix is ​​established. For concrete areas, the contrast enhancement coefficient is set to 1.2; for steel reinforcement areas, it is set to 1.4. The adjusted pixel values ​​are: The offset is dynamically determined based on the material type.

[0031] The separation process of the standardized temporal infrared image using a dual-channel convolutional network model includes: The dual-channel convolutional network model includes a fast-channel convolutional module, a slow-channel convolutional module, and a cascaded attenuation compensation module; The standardized temporal infrared image is input into the fast channel convolution module, and the time window of the fast channel is set to 1 to 10 seconds. Short-period physiological signal features in the image are extracted to obtain short-time features. The standardized temporal infrared image is input into the slow channel convolution module, and the time window of the slow channel is set to 1 to 6 hours. Long-period background signal features in the image are extracted to obtain long-time features. A cascaded attenuation compensation module is used to perform inverse signal attenuation compensation for both short-time and long-time characteristics.

[0032] The calculation of the dual-channel weight coefficients and the weighted fusion of short-term and long-term features include: Temperature sensors are used to detect the temperature values ​​at multiple monitoring points in the ruins environment in real time, and the variance of the temperature values ​​is calculated as the temperature variance of the ruins. The environmental stability level is determined based on the magnitude of the temperature variance of the ruins. When the temperature variance of the ruins exceeds the preset threshold, the weight coefficient of the fast channel is set to a high weight value and the weight coefficient of the slow channel is set to a low weight value. The short-term features are multiplied by the fast channel weight coefficient, and the long-term features are multiplied by the slow channel weight coefficient. The weighted features are then numerically summed to obtain the time-series features.

[0033] In one specific embodiment, the formula for calculating the dual-channel weighting coefficient is: ; ; in, This refers to the fast-track weighting coefficient; This refers to the slow channel weighting coefficient. The temperature variance of the ruins; This is the fast channel attenuation adjustment factor, with a value range of 0.1-0.3; This is a slow channel enhancement adjustment factor, with a value range of 0.05-0.15; This is the ratio of the time window lengths, i.e., the ratio of the current window length to the standard window length; The fast channel baseline variance threshold, with a value range of [value range missing]. ; This is the baseline variance threshold for the slow channel, with a value range of [value range missing]. ; This is the normalization coefficient for the fast channel, with a value ranging from 0.6 to 0.8; This is the normalization coefficient for the slow channel, with a value ranging from 0.2 to 0.4.

[0034] Specifically, traditional dual-channel weight calculation methods use fixed weight allocation, which cannot adapt to the dynamic changes in the ruin environment. This embodiment innovatively introduces the variance of ruin temperature. As an indicator for environmental stability assessment, it is used through an exponential decay function. and Dynamic adjustment of the weights for fast and slow channels is achieved. When ambient temperature fluctuates significantly, the weight of the fast channel is automatically increased to capture transient physiological signals; when the environment is stable, the weight of the slow channel is increased to extract long-term trend features. Normalization coefficient. and This ensures the rationality of weight allocation and avoids the feature loss problem caused by weight imbalance in traditional methods.

[0035] The formula for calculating the time series features is: ; in, The fused temporal characteristics; It is a short-term characteristic; It is a long-term characteristic; For fast-channel time decay function, , For slow channel time enhancement function, , This is the cross-term coupling coefficient, with a value ranging from 0.1 to 0.3. This represents the fast-channel time decay rate, with a value ranging from 0.02 to 0.05. This represents the slow channel time enhancement rate, with a value ranging from 0.01 to 0.03. For reference time points.

[0036] Specifically, traditional feature fusion methods use simple linear weighting, ignoring the dynamic correlation of the time dimension. This embodiment innovatively introduces a time decay function. and time augmentation function This achieves accurate modeling of temporal correlations. (Interaction terms) For the first time, the nonlinear coupling effect between fast and slow channels was considered, and the synergistic effect between channels was captured by geometric averaging, which significantly improved the expressive power of fused features.

[0037] This embodiment effectively solves the balance problem between temporal resolution and anti-interference capability in physiological signal detection under complex ruin environments by using an adaptive dual-channel weight allocation driven by the temperature variance of the ruins and a feature fusion mechanism that enhances temporal correlation. While maintaining rapid response capability, it significantly improves the detection accuracy of weak physiological signals. Compared with the traditional fixed weight method, it improves the signal detection accuracy and reduces the false detection rate.

[0038] The process of performing multi-domain noise detection and suppression on the time-series features, identifying the frequency components of impulse noise, and performing filtering processing in conjunction with the frequency range of physiological signals to obtain reconstructed time-series features includes: The time-series characteristics are subjected to multi-domain analysis in the time domain, frequency domain, and time-frequency domain. The noise type is matched and identified using a physical simulation noise template library. The frequency components and amplitude characteristics of non-periodic impulse noise are detected through an impulse detection gating mechanism to obtain the noise distribution characteristics.

[0039] In one specific embodiment, the step of performing multi-domain analysis processing on the time-series features in the time domain, frequency domain, and time-frequency domain, and using a physical simulation noise template library for noise type matching and identification, includes: The time-series features are input into the time-domain analysis module, frequency-domain analysis module, and time-frequency domain analysis module, respectively. The signal statistical features of the corresponding domains are extracted. Standard templates for aftershock noise, environmental interference noise, and equipment noise are retrieved from the physical simulation noise template library. The similarity score is calculated by the template matching algorithm. The impact noise components that exceed the range of normal physiological signals are identified by the gated threshold judgment mechanism.

[0040] Based on the noise distribution characteristics, adaptive filter parameters are designed, and a filter passband is established in combination with the preset physiological signal frequency range. The time series characteristics are then subjected to frequency domain filtering, and weak signals are reconstructed using the heart rate-respiration coupling constraint to obtain the reconstructed time series characteristics.

[0041] In one specific embodiment, the step of performing frequency domain filtering on the time-series features and reconstructing weak signals using heart rate-respiration coupling constraints includes: Based on the frequency location and amplitude information in the noise distribution characteristics, the cutoff frequency and attenuation coefficient of the filter are calculated. The frequency range of the physiological signal is set to 0.1 to 5 Hz as the filter passband. Frequency components outside the passband range are attenuated. At the same time, the weak physiological signal that was mistakenly deleted is reconstructed probabilistically using a heavy-tailed distributed variational encoder. The reliability of the reconstructed signal is evaluated by Bayesian uncertainty quantization.

[0042] In one specific embodiment, multi-domain noise suppression and physiological signal reconstruction in a fire ruin environment includes: At a fire rescue site, the ruins environment contained multiple noise sources, including aftershocks, wind noise, and equipment vibration. The time-series characteristic signal was 600 seconds long, with a sampling frequency of 50Hz, and contained a total of 30,000 data points. The specific processing procedure is as follows:

[0043] The time-domain analysis module includes a statistical feature calculator, an anomaly detector, a signal feature extractor, and an output interface. The statistical feature calculator includes a mean calculation unit, a variance calculation unit, a peak detection unit, and a higher-order moment calculation unit. The anomaly detector integrates the Z-score detection algorithm, the Grubbs test algorithm, and a sliding window detector. The signal feature extractor includes an amplitude analyzer, a duration analyzer, and a rate of change analyzer. The output interface provides a standardized time-domain feature vector output. The calculated mean of the time-series feature signal is 0.025, the standard deviation is 0.18, the peak factor is 4.2, and the skewness coefficient is 1.8. Fifteen abnormal peaks were detected in the time-series feature signal, with peak amplitudes exceeding five times the standard deviation of the mean.

[0044] The frequency domain analysis module includes an FFT transformer, a power spectral density calculator, a peak detector, and a frequency feature extractor. The FFT transformer employs a radix-2 Cooley-Tukey algorithm and supports 2048-point transformation. The power spectral density calculator includes Welch and Bartlett method estimators. The peak detector integrates an adaptive threshold detection algorithm and a peak merging algorithm. The frequency feature extractor includes a dominant frequency detector, a bandwidth analyzer, and a harmonic analyzer. Fast Fourier Transform analysis revealed that the main frequency components of the signal are distributed in the range of 0.1-8 Hz. A strong peak with an amplitude of 0.12 exists at 0.2 Hz, representing a respiratory signal; a peak with an amplitude of 0.08 exists at 1.1 Hz, representing a heart rate signal; and an abnormal peak with an amplitude of 0.25 exists at 7.5 Hz, representing device noise.

[0045] The time-frequency domain analysis module includes a short-time Fourier transform, a wavelet transform, and an instantaneous feature extractor. The short-time Fourier transform includes Hanning window, Hamming window, and Gaussian window selectors. The wavelet transform supports Morlet, Daubechies, and Mexican hat wavelets. The instantaneous feature extractor integrates an instantaneous frequency calculator and an instantaneous amplitude calculator. Using a short-time Fourier transform with a window length of 2 seconds and an overlap rate of 75%, intermittent energy spikes were observed in the 2.5-4.0 Hz frequency band within the 120-135 second timeframe, with an amplitude reaching 0.35, which was identified as aftershock noise.

[0046] Noise type matching and identification includes: The system retrieves a physical simulation noise template library, which includes a noise model database and a template generator. The noise model database stores mathematical model parameters for 15 typical disaster environmental noises. The template generator automatically generates standard noise templates based on the physical parameters. The similarity calculation engine integrates Pearson correlation coefficient, cosine similarity, and Euclidean distance calculators. The threshold manager dynamically adjusts the matching threshold and confidence parameters. Three types of standard noise templates are retrieved from the template library:

[0047] Aftershock noise template: frequency range 2-6Hz, duration 10-20 seconds, amplitude attenuation index -0.8; Environmental interference noise template: frequency range 5-15Hz, continuous, amplitude variation coefficient <0.3; Equipment noise template: The frequency is concentrated at 7.5Hz±0.5Hz, the amplitude is stable, and it appears periodically.

[0048] Template matching calculation includes: calculating similarity scores using a normalized cross-correlation algorithm. The similarity to the aftershock noise template was 0.87 during the 120-135 second time period. The 7.5Hz frequency component has a similarity of 0.92 with the equipment noise template. It did not exceed the gating threshold of 0.7, and the similarity with the environmental interference noise template was 0.34.

[0049] The adaptive filter is designed with parameters including a filter designer, a parameter optimizer, a multi-stage cascaded structure, and a real-time monitor. The filter designer includes Butterworth, Chebyshev, and elliptic filter design algorithms. The parameter optimizer uses genetic algorithms and particle swarm optimization to optimize filter parameters. The multi-stage cascaded structure includes a low-pass filter stage, a high-pass filter stage, a band-pass filter stage, and a notch filter stage. The real-time monitor monitors the filtering effect and dynamically adjusts the filter coefficients.

[0050] First, set the filter passband: establish the filter passband based on the physiological signal frequency range of 0.1-5Hz, where the respiratory signal passband is 0.1-0.8Hz and the heart rate signal passband is 0.8-5.0Hz.

[0051] Next, the filter parameters were calculated: the low-pass cutoff frequency is 5.0Hz, the attenuation coefficient is -40dB / decade, the high-pass cutoff frequency is 0.1Hz, and the attenuation coefficient is -20dB / decade.

[0052] For the notch filter, the center frequency is 7.5Hz, the quality factor is Q=10, and the attenuation depth is -60dB.

[0053] Re-establish heart rate-respiration coupling constraint: The heavy-tailed variational encoder parameters are designed. The heavy-tailed variational encoder includes an encoder network, a decoder network, a variational inference module, and an uncertainty quantizer. The encoder network contains three fully connected layers and a ReLU activation function, with 512-256-16 nodes respectively. The decoder network contains three fully connected layers and a Sigmoid activation function, with 16-256-512 nodes respectively. The variational inference module integrates reparameterization techniques and a KL divergence calculator. The uncertainty quantizer includes a Monte Carlo sampler and a confidence interval calculator. The Student-t distribution is used as the heavy-tailed distribution, with 3 degrees of freedom and a scaling parameter of 0.05. The encoder latent variable dimension is set to 16, and the decoder output dimension is the same as the original signal.

[0054] The probability reconstruction process includes: reconstructing the 0.3Hz and 1.8Hz frequency components that were mistakenly deleted during the filtering process. The posterior distribution parameters are calculated through variational inference.

[0055] The component reconstruction amplitude at 0.3Hz is: , ; The component reconstruction amplitude at 1.8Hz is: , CC

[0056] Re-quantifying Bayesian uncertainty: Achieved by calculating the confidence level of the reconstructed signal. The confidence level of the 0.3Hz component is: 95% confidence interval [0.039, 0.051], confidence score 0.89; The confidence level of the 1.8Hz component is: 95% confidence interval [0.025, 0.039], confidence score 0.83.

[0057] The detection and classification of material types in the standardized time-series infrared image and the compensation and correction of image temperature values ​​include: The material recognition module is used to classify the materials of the ruins in the standardized time-series infrared images at the pixel level and identify the material types, which include concrete, steel bars, bricks, wood and soil, to obtain a material distribution map. Based on the spatial distribution of each material type in the material distribution map, the corresponding thermal conductivity coefficient, thermal diffusivity and heat capacity parameters are extracted from the material thermal conductivity parameter database. The thermal bridge modeling algorithm is used to calculate the thermal conduction coupling effect of multi-material interfaces, and the temperature attenuation degree of different material regions is quantitatively analyzed to generate a thermal conduction attenuation coefficient matrix. The thermal conduction attenuation coefficient matrix is ​​multiplied pixel-by-pixel with the standardized time-series infrared image to obtain the compensated time-series infrared image.

[0058] The calculation of the multi-material interface thermal conduction coupling effect using the interface thermal bridge modeling algorithm includes: Identify the boundary pixel positions between different materials in the material distribution map and establish the topological connection relationship of the material interface; Based on the variation law of thermophysical parameters of various materials under disaster environment recorded in the specific material library of ruins, the effective thermal conductivity at the interface is calculated; The heat transfer path and attenuation law at the material interface are analyzed by using a multilayer coupled heat transfer model. The influence of material thickness, density distribution and porosity on heat conduction is considered to obtain the heat conduction attenuation coefficient matrix.

[0059] In one specific embodiment, the formula for calculating the interface thermal conduction coupling effect is: ; in, For material interface and The effective thermal conductivity coefficient between them; For material interface thermal conductivity coefficient; For material interface thermal conductivity coefficient; For material interface and The interface roughness coupling factor between them ranges from 0.05 to 0.25. For material interface and The equivalent density at that location, ; and Material interface and Density distribution; This represents the maximum density value of the material in the ruins. This is the porosity influence coefficient, with a value ranging from 0.3 to 0.7. Porosity of the material interface; The void attenuation index has a value range of 1.2-1.8.

[0060] Specifically, traditional heat conduction calculations use the heat conduction coefficient of a single material, neglecting the complex physical effects of material interfaces. This embodiment innovatively introduces an interface roughness coupling factor. The influence of interface microstructure on heat conduction is modeled using a sine function. Porosity correction term. Using a power function, the nonlinear attenuation effect of void structures on heat conduction in a ruin environment is accurately described, solving the problem that traditional methods cannot handle the thermal bridging effect at multi-material interfaces.

[0061] The formula for generating the heat conduction attenuation coefficient matrix is: ; in, For position The thermal conductivity attenuation coefficient at that location; For position The number of material types involved; For the first The effective thermal conductivity of the material; For reference thermal conductivity, the value is taken as... ; To reach the first The distance between the centers of the materials; For the first The radius of influence of thermal conductivity of a certain material; For the first Weighting factors for different materials ; It is a step function; The threshold temperature for heat source detection; The ambient reference temperature; It is an exponential function.

[0062] Specifically, traditional attenuation coefficient calculation methods are based on single-layer material models and cannot handle complex situations involving multiple materials stacked together. This embodiment innovatively adopts a multiplication method. To achieve composite calculation of the influence of multiple materials, a Gaussian distance weighting function is used. Accurately modeling the effect of spatial distance on heat conduction. Step function. By introducing a temperature threshold judgment mechanism, attenuation compensation is only performed when a valid heat source is detected, thus avoiding interference from background noise and significantly improving calculation accuracy and stability.

[0063] This embodiment accurately quantifies the physical process of heat conduction in complex ruin environments by using heat conduction coupling modeling with dual correction of interface roughness and porosity, as well as an algorithm for generating attenuation coefficient matrices based on the influence of multiple materials. This effectively compensates for the temperature attenuation loss caused by material obstruction, improves the detection accuracy of deep heat sources, reduces positioning errors, and provides a reliable physical basis for accurately identifying the location of trapped personnel.

[0064] The step of performing spatial structure analysis on the compensated temporal infrared image and constructing a three-dimensional spatial thermal distribution model by combining multi-frame temporal information includes: The compensated temporal infrared image is subjected to multi-scale spatial feature extraction using a spatiotemporally coupled convolutional processor, and spatiotemporal correlation analysis is performed by combining the temperature change gradient information of adjacent time frames. The thermal tomography algorithm is used to vertically stitch together the heat conduction information of different depth layers to construct a four-dimensional spatiotemporal data structure containing spatial coordinates, temperature values ​​and time dimensions. The four-dimensional spatiotemporal data structure is converted into a three-dimensional spatial heat distribution model using a three-dimensional reconstruction algorithm. A threshold range for heat source intensity is set, and spatial clustering analysis is performed on heat source regions that exceed the threshold to screen out candidate regions that conform to the thermal radiation characteristics of living organisms, thus obtaining candidate locations.

[0065] The method involves vertically stitching together heat conduction information from different depth layers using a thermal tomography connection algorithm to construct a four-dimensional spatiotemporal data structure containing spatial coordinates, temperature values, and a time dimension, including: The compensated temporal infrared image is divided into multiple thermal conduction layers according to the spatial depth direction, and each thermal conduction layer corresponds to a different material occlusion depth range. The temperature transfer relationship between adjacent layers is calculated using the interlayer heat conduction attenuation model, and the heat conduction connection matrix between layers is established. Based on the physical laws of heat conduction, the temperature distribution at each level is reconstructed in reverse to compensate for the temperature attenuation loss caused by material shading. The temperature data at each level after inverse reconstruction are registered in three-dimensional space according to depth order to obtain a four-dimensional spatiotemporal data structure, which includes location coordinates, temperature intensity, material type and timestamp.

[0066] In one specific embodiment, the calculation formula for the interlayer heat conduction attenuation model is: ; in, To from the depth layer Passed to the deep layer Calories; The initial heat source intensity; For the first The thickness of the layer material; For the first The thermal conductivity characteristic length of the layer material; For the first thermal resistance of the layer ; For the first The equivalent heat transfer area of ​​the layer; For the first The tilt angle of the layer relative to the vertical direction; It is an exponential function.

[0067] Specifically, traditional interlayer heat conduction models employ simple exponential decay, which cannot accurately describe the heat transfer process in multilayer heterogeneous materials. This embodiment innovatively introduces a cumulative decay term. This involves comprehensively considering the thickness and feature length of each material layer. Thermal resistance coupling term. A harmonic averaging approach is used to accurately model the thermal resistance matching effect between adjacent layers. Angle correction factor. For the first time, the influence of the inclination angle of the ruin material on vertical heat transfer was considered, which significantly improved the accuracy of interlayer heat transfer calculation.

[0068] The reconstruction formula for the four-dimensional spatiotemporal data structure is as follows: ; in, The reconstructed four-dimensional temperature distribution; This represents the total number of heat conduction layers. For the first Time-varying weighting coefficients for the layer; For the first The measured temperature value of the layer; For the first The depth response function of the layer ; For the first The center depth coordinates of the layer For the first Layer thickness parameters; For the first The temperature decay length constant of the layer; This is a combined noise term based on location coordinates, temperature intensity, material type, and timestamp. It is an exponential function.

[0069] Specifically, traditional 3D reconstruction methods, based on linear interpolation, cannot handle complex temperature distributions along the depth direction. This embodiment employs a layered weighted summation method. Time-varying weighting coefficient Dynamic adjustment of the importance of different depth layers was implemented. Depth response function A sigmoid function is used to smoothly model the nonlinear characteristics of depth influence. (Exponential decay term) According to the physical laws describing the natural decay of temperature with depth, the noise term is considered. Uncertainties in location coordinates, temperature intensity, material type, and timestamps are handled uniformly.

[0070] This embodiment achieves accurate temperature distribution reconstruction of deep structures in ruins by using multi-layer heat transfer modeling with interlayer thermal resistance matching and angle correction, as well as a four-dimensional spatiotemporal reconstruction algorithm with layer weights and depth response functions. This enables depth positioning accuracy to reach ±0.15 meters and controls the reconstruction error of the three-dimensional heat distribution model within ±0.8℃. This provides high-precision technical support for accurate spatial positioning and vital sign intensity assessment of trapped personnel in complex multi-layer ruin environments.

[0071] In one specific embodiment, the construction of a three-dimensional spatial heat distribution model for an explosion rubble environment includes: At the rescue site of an explosion, the rubble was approximately 12 meters deep and contained a complex layered structure consisting of multiple concrete slabs, steel mesh, and broken bricks and stones. The compensated temporal infrared images had a resolution of 640×480 pixels, a temporal length of 360 seconds, a sampling frequency of 5 frames / second, and a total of 1800 frames. The composition of each module and the specific processing procedure are as follows:

[0072] The spatiotemporally coupled convolutional processor includes a multi-scale feature extractor, a temporal correlation analyzer, a gradient calculator, and a feature fusion processor, wherein: The multi-scale feature extractor contains parallel convolutional layers with three types of convolutional kernels: 3×3, 5×5, and 7×7, with 64, 32, and 16 output channels per layer, respectively. The temporal correlation analyzer uses a 3×3×5 3D convolutional kernel to capture spatiotemporal correlations and includes a temporal dimension pooling layer and a spatial dimension pooling layer. The gradient calculator integrates the Sobel operator, Prewitt operator, and Roberts operator to perform multi-directional gradient calculations; The feature fusion unit employs an attention mechanism to adaptively weight and fuse multi-scale features.

[0073] The thermal tomographic connection algorithm module includes a depth layerer, an interlayer connection matrix generator, an inverse reconstruction unit, and a spatial registration unit, wherein: The depth layerer automatically divides the depth layers according to the material distribution and thermal conductivity coefficient, supporting 1-20 adjustable layers; The interlayer connection matrix generator calculates the thermal conduction coupling coefficient between adjacent layers based on the finite element method. The inverse reconstructor uses an iterative least squares method to solve the inverse heat conduction problem, including a regularization term and a convergence test. The spatial registration device integrates the ICP algorithm and the feature point matching algorithm to perform multi-layer data alignment.

[0074] The 3D reconstruction algorithm module includes a point cloud generator, a surface reconstructor, a heat source clusterer, and a threshold optimizer, wherein: The point cloud generator converts two-dimensional image pixels into three-dimensional point clouds and supports multiple interpolation methods. The surface reconstructor uses Delaunay triangulation and Poisson reconstruction algorithms to construct a three-dimensional surface; The heat source clusterer integrates the DBSCAN algorithm and the K-means algorithm for heat source region segmentation. The threshold optimizer determines the optimal heat source intensity threshold based on ROC curve analysis.

[0075] The step of associating and matching the spatial coordinate information of the candidate locations with the physiological signal information of the reconstructed temporal features, and outputting the vital sign perception results, includes: The spatial-signal correlation matching module is used to perform spatiotemporal correspondence analysis on the three-dimensional spatial coordinate information of the candidate positions and the physiological signal information of the reconstructed temporal features. The correlation score between each candidate position and the corresponding physiological signal intensity is calculated by the distance weighting algorithm, and a mapping relationship table between candidate positions and physiological signals is established.

[0076] In one specific embodiment, the spatial-signal correlation matching module performs spatiotemporal correspondence analysis on the three-dimensional spatial coordinate information of the candidate locations and the physiological signal information of the reconstructed temporal features, and calculates the correlation score between each candidate location and the corresponding physiological signal intensity using a distance-weighted algorithm, including: The three-dimensional spatial coordinates of the candidate positions are synchronized with the signal sampling points at the corresponding time points in the reconstructed temporal features. A spatial neighborhood search algorithm is used to set a search radius around each candidate position, and physiological signal data within the search range is extracted. A weighting coefficient is calculated based on the inverse relationship between signal strength and spatial distance. The correlation score is obtained by multiplying the weighting coefficient by the signal strength.

[0077] Threshold filtering is performed based on the correlation scores in the mapping table. Valid life forms with correlation scores exceeding a preset threshold are extracted. Spectral analysis is performed on the physiological signals of the valid life forms to extract respiratory rate parameters. Combined with spatial coordinates and signal intensity information, vital sign perception results are generated. The vital sign perception results include vital sign intensity, respiratory rate, and precise three-dimensional location.

[0078] In one specific embodiment, the generation of vital sign perception results includes: A correlation score threshold is set to a preset value. Candidate locations with correlation scores greater than the correlation score threshold are selected as valid vital signs locations. The reconstructed temporal features corresponding to the valid vital signs locations are processed by Fast Fourier Transform. The peak frequency in the spectrum is identified as the respiratory frequency parameter, and the amplitude value of the physiological signal is used as the vital sign intensity parameter. Combined with three-dimensional spatial coordinates, a structured vital sign perception result is formed.

[0079] In one specific embodiment, the output of vital sign perception results at the factory collapse site includes: At the factory collapse rescue site, after obtaining three candidate locations, the reconstruction time-series characteristic signal with a length of 600 seconds was used. The composition of each module and the specific processing procedure are as follows:

[0080] The space-signal association matching module includes a spatiotemporal synchronization aligner, a spatial neighborhood searcher, a distance-weighted calculator, and an association scorer, wherein: The spatiotemporal synchronization aligner includes a timestamp parser, a frame rate converter, and a synchronization error compensator, supporting a synchronization accuracy of ±0.02 seconds. The spatial neighborhood searcher integrates KD tree spatial indexing, spherical search algorithm and ellipsoidal search algorithm, and supports adaptive search radius adjustment; The distance weighting calculator includes an Euclidean distance calculation unit, a Mahalanobis distance calculation unit, and a weighting function generator; The correlation scorer employs a multi-dimensional feature fusion algorithm, integrating signal strength, frequency consistency, and time stability scores.

[0081] The spectrum analysis module includes, wherein: The FFT transformer supports 1024-point, 2048-point, and 4096-point transformations and employs Hanning and Blackman window functions. The peak identifier integrates adaptive threshold detection, peak accuracy refinement, and harmonic suppression algorithms. The respiratory rate extractor includes a physiological rate range filter and a rate stability verifier; The parameter confidence calculator calculates parameter confidence based on signal-to-noise ratio and spectral purity.

[0082] The present invention also provides a vital signs sensing system based on time-series infrared images, the system comprising: The image acquisition and preprocessing module is used to acquire a time-series infrared image sequence of a disaster relief scene, and to use an image processor to perform brightness and contrast standardization processing on the time-series infrared image sequence to obtain a standardized time-series infrared image. The dual-channel feature extraction and fusion module is used to separate and process the standardized temporal infrared image through a dual-channel convolutional network model to obtain short-time features and long-time features. Based on the variance of the ruin temperature detected by the temperature sensor, the dual-channel weight coefficients are calculated, and the short-time features and long-time features are weighted and fused to obtain temporal features. The multi-domain noise suppression and signal reconstruction module is used to perform multi-domain noise detection and suppression on the time-series features, identify the frequency components of impulse noise, and perform filtering processing in combination with the frequency range of physiological signals to obtain the reconstructed time-series features. The material identification and thermal conduction compensation module is used to detect and classify the material types in the standardized time-series infrared image, calculate the thermal conduction attenuation coefficient according to the preset material thermal conduction parameter database, and compensate and correct the image temperature value to obtain a compensated time-series infrared image. The three-dimensional thermal distribution modeling and candidate region screening module is used to perform spatial structure analysis on the compensated temporal infrared image, construct a three-dimensional spatial thermal distribution model by combining multi-frame temporal information, and screen candidate regions of living organisms according to the heat source intensity threshold to obtain candidate locations. The spatial signal association matching and result output module is used to associate and match the spatial coordinate information of the candidate location with the physiological signal information of the reconstructed temporal features, and output the vital sign perception result.

[0083] Specifically, this embodiment of a vital sign perception system based on temporal infrared images integrates a dual-channel convolutional network model with multi-domain noise suppression, utilizes material recognition algorithms for thermal conduction compensation and three-dimensional spatial thermal distribution modeling, dynamically adjusts feature extraction and signal reconstruction strategies using a spatiotemporal correlation matching mechanism, and adaptively optimizes the multi-module processing flow according to the complexity of the ruin environment. Through spatial-signal-temporal three-dimensional correlation analysis, it solves the problem of accurately locating trapped personnel and detecting weak physiological signals in complex ruin environments, improving the accuracy and reliability of vital sign perception in disaster relief scenarios. Through a complete multi-level feature fusion and confidence assessment mechanism, it meets the actual needs of rapid and accurate rescue under different disaster types.

[0084] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for sensing vital signs based on temporal infrared images, characterized in that, Includes the following steps: A time-series infrared image sequence of a disaster relief scene is acquired, and the brightness and contrast of the time-series infrared image sequence are normalized using an image processor to obtain a normalized time-series infrared image. The standardized temporal infrared image is separated and processed by a dual-channel convolutional network model to obtain short-time features and long-time features. The dual-channel weight coefficients are calculated based on the variance of the ruin temperature detected by the temperature sensor, and the short-time features and long-time features are weighted and fused to obtain temporal features. Multi-domain noise detection and suppression are performed on the time-series features to identify the frequency components of impulse noise. Filtering is then performed in conjunction with the frequency range of physiological signals to obtain the reconstructed time-series features. The material types in the standardized time-series infrared images are detected and classified. The thermal conduction attenuation coefficient is calculated based on a preset material thermal conduction parameter database. The image temperature values ​​are then compensated and corrected to obtain a compensated time-series infrared image. Spatial structure analysis is performed on the compensated temporal infrared image, and a three-dimensional spatial heat distribution model is constructed by combining multi-frame temporal information. Candidate regions of living organisms are screened according to the heat source intensity threshold to obtain candidate locations. The spatial coordinate information of the candidate location is correlated and matched with the physiological signal information of the reconstructed temporal features to output the vital sign perception result.

2. The vital signs perception method based on time-series infrared images as described in claim 1, characterized in that, The separation process of the standardized temporal infrared image using a dual-channel convolutional network model includes: The dual-channel convolutional network model includes a fast-channel convolutional module, a slow-channel convolutional module, and a cascaded attenuation compensation module; The standardized temporal infrared image is input into the fast channel convolution module, and the time window of the fast channel is set to 1 to 10 seconds. Short-period physiological signal features in the image are extracted to obtain short-time features. The standardized temporal infrared image is input into the slow channel convolution module, and the time window of the slow channel is set to 1 to 6 hours. Long-period background signal features in the image are extracted to obtain long-time features. A cascaded attenuation compensation module is used to perform inverse signal attenuation compensation for both short-time and long-time characteristics.

3. The vital signs perception method based on time-series infrared images as described in claim 2, characterized in that, The calculation of the dual-channel weight coefficients and the weighted fusion of short-term and long-term features include: Temperature sensors are used to detect the temperature values ​​at multiple monitoring points in the ruins environment in real time, and the variance of the temperature values ​​is calculated as the temperature variance of the ruins. The environmental stability level is determined based on the magnitude of the temperature variance of the ruins. When the temperature variance of the ruins exceeds the preset threshold, the weight coefficient of the fast channel is set to a high weight value and the weight coefficient of the slow channel is set to a low weight value. The short-term features are multiplied by the fast channel weight coefficient, and the long-term features are multiplied by the slow channel weight coefficient. The weighted features are then numerically summed to obtain the time-series features.

4. The vital signs perception method based on time-series infrared images as described in claim 1, characterized in that, The detection and classification of material types in the standardized time-series infrared image and the compensation and correction of image temperature values ​​include: The materials of the ruins in the standardized time-series infrared images are classified at the pixel level to identify the material types, which include concrete, steel bars, bricks, wood and soil, and a material distribution map is obtained. Based on the spatial distribution of each material type in the material distribution map, the corresponding thermal conductivity coefficient, thermal diffusivity and heat capacity parameters are extracted from the material thermal conductivity parameter database. The thermal bridge modeling algorithm is used to calculate the thermal conduction coupling effect of multi-material interfaces, and the temperature attenuation degree of different material regions is quantitatively analyzed to generate a thermal conduction attenuation coefficient matrix. The thermal conduction attenuation coefficient matrix is ​​multiplied pixel-by-pixel with the standardized time-series infrared image to obtain the compensated time-series infrared image.

5. The vital signs perception method based on time-series infrared images as described in claim 4, characterized in that, The calculation of the multi-material interface thermal conduction coupling effect using the interface thermal bridge modeling algorithm includes: Identify the boundary pixel positions between different materials in the material distribution map and establish the topological connection relationship of the material interface; Based on the variation law of thermophysical parameters of various materials under disaster environment recorded in the specific material library of ruins, the effective thermal conductivity at the interface is calculated; The heat transfer path and attenuation law at the material interface are analyzed by using a multilayer coupled heat transfer model. The influence of material thickness, density distribution and porosity on heat conduction is considered to obtain the heat conduction attenuation coefficient matrix.

6. The vital signs perception method based on time-series infrared images as described in claim 1, characterized in that, The step of performing spatial structure analysis on the compensated temporal infrared image and constructing a three-dimensional spatial thermal distribution model by combining multi-frame temporal information includes: The compensated temporal infrared image is subjected to multi-scale spatial feature extraction using a spatiotemporally coupled convolutional processor, and spatiotemporal correlation analysis is performed by combining the temperature change gradient information of adjacent time frames. The thermal tomography algorithm is used to vertically stitch together the heat conduction information of different depth layers to construct a four-dimensional spatiotemporal data structure containing spatial coordinates, temperature values ​​and time dimensions. The four-dimensional spatiotemporal data structure is converted into a three-dimensional spatial heat distribution model using a three-dimensional reconstruction algorithm. A threshold range for heat source intensity is set, and spatial clustering analysis is performed on heat source regions that exceed the threshold to screen out candidate regions that conform to the thermal radiation characteristics of living organisms, thus obtaining candidate locations.

7. The vital signs perception method based on time-series infrared images as described in claim 6, characterized in that, The method involves vertically stitching together heat conduction information from different depth layers using a thermal tomography connection algorithm to construct a four-dimensional spatiotemporal data structure containing spatial coordinates, temperature values, and a time dimension, including: The compensated temporal infrared image is divided into multiple thermal conduction layers according to the spatial depth direction, and each thermal conduction layer corresponds to a different material occlusion depth range. The temperature transfer relationship between adjacent layers is calculated using the interlayer heat conduction attenuation model, and the heat conduction connection matrix between layers is established. Based on the physical laws of heat conduction, the temperature distribution at each level is reconstructed in reverse to compensate for the temperature attenuation loss caused by material shading. The temperature data at each level after inverse reconstruction are registered in three-dimensional space according to depth order to obtain a four-dimensional spatiotemporal data structure, which includes location coordinates, temperature intensity, material type and timestamp.

8. The vital signs perception method based on time-series infrared images as described in claim 1, characterized in that, The process of performing multi-domain noise detection and suppression on the time-series features, identifying the frequency components of impulse noise, and performing filtering processing in conjunction with the frequency range of physiological signals to obtain reconstructed time-series features includes: The time-series characteristics are subjected to multi-domain analysis processing in the time domain, frequency domain, and time-frequency domain. The noise type is matched and identified using a physical simulation noise template library. The frequency components and amplitude characteristics of non-periodic impulse noise are detected through an impulse detection gating mechanism to obtain the noise distribution characteristics. Based on the noise distribution characteristics, adaptive filter parameters are designed, and a filter passband is established in combination with the preset physiological signal frequency range. The time series characteristics are then subjected to frequency domain filtering, and weak signals are reconstructed using the heart rate-respiration coupling constraint to obtain the reconstructed time series characteristics.

9. The vital signs perception method based on temporal infrared images as described in claim 1, characterized in that, The step of associating and matching the spatial coordinate information of the candidate locations with the physiological signal information of the reconstructed temporal features, and outputting the vital sign perception results, includes: The spatial-signal correlation matching module is used to perform spatiotemporal correspondence analysis on the three-dimensional spatial coordinate information of the candidate positions and the physiological signal information of the reconstructed temporal features. The correlation score between each candidate position and the corresponding physiological signal intensity is calculated by the distance weighting algorithm, and a mapping relationship table between candidate positions and physiological signals is established. Threshold filtering is performed based on the correlation scores in the mapping table. Valid life forms with correlation scores exceeding a preset threshold are extracted. Spectral analysis is performed on the physiological signals of the valid life forms to extract respiratory rate parameters. Combined with spatial coordinates and signal intensity information, vital sign perception results are generated. The vital sign perception results include vital sign intensity, respiratory rate, and precise three-dimensional location.

10. A vital signs sensing system based on temporal infrared images, used to execute a vital signs sensing method based on temporal infrared images as described in any one of claims 1-9, characterized in that, The system includes: The image acquisition and preprocessing module is used to acquire a time-series infrared image sequence of a disaster relief scene, and to use an image processor to perform brightness and contrast standardization processing on the time-series infrared image sequence to obtain a standardized time-series infrared image. The dual-channel feature extraction and fusion module is used to separate and process the standardized temporal infrared image through a dual-channel convolutional network model to obtain short-time features and long-time features. Based on the variance of the ruin temperature detected by the temperature sensor, the dual-channel weight coefficients are calculated, and the short-time features and long-time features are weighted and fused to obtain temporal features. The multi-domain noise suppression and signal reconstruction module is used to perform multi-domain noise detection and suppression on the time-series features, identify the frequency components of impulse noise, and perform filtering processing in combination with the frequency range of physiological signals to obtain the reconstructed time-series features. The material identification and thermal conduction compensation module is used to detect and classify the material types in the standardized time-series infrared image, calculate the thermal conduction attenuation coefficient according to the preset material thermal conduction parameter database, and compensate and correct the image temperature value to obtain a compensated time-series infrared image. The three-dimensional thermal distribution modeling and candidate region screening module is used to perform spatial structure analysis on the compensated temporal infrared image, construct a three-dimensional spatial thermal distribution model by combining multi-frame temporal information, and screen candidate regions of living organisms according to the heat source intensity threshold to obtain candidate locations. The spatial signal association matching and result output module is used to associate and match the spatial coordinate information of the candidate location with the physiological signal information of the reconstructed temporal features, and output the vital sign perception result.

Images