A data acquisition processing method and system for optoelectronic measurement

Abstract

The application relates to the technical field of photoelectric measurement, and discloses a data acquisition and processing method and system for photoelectric measurement. The method comprises the following steps: acquiring photoelectric data of a target region, realizing multi-source data space-time registration through preprocessing, and extracting space-time fusion photoelectric signals; obtaining spectral characteristics according to the signals, dividing a preliminary detection region, and completing fine target segmentation and feature optimization; constructing a feature extraction model based on time-frequency analysis to perform target classification, and generating an acquisition result; and obtaining energy distribution changes by using multi-period data to generate dynamic calibration parameters. The system comprises a memory and a processor, and executes corresponding programs to realize the above steps. The application solves the problems of the prior art in aspects such as multi-source data fusion, target identification and dynamic calibration, improves the precision and reliability of photoelectric measurement, and is suitable for photoelectric measurement scenes in multiple fields.

Description

Technical Field

[0001] This invention relates to the field of photoelectric measurement technology, specifically to a data acquisition and processing method and system for photoelectric measurement. Background Technology

[0002] In the course of modern technological development, photoelectric measurement technology, with its high precision, non-contact nature, and fast response, has been widely applied in numerous fields such as industrial inspection, environmental monitoring, and aerospace. However, the current data acquisition and processing processes in photoelectric measurement still face many challenges, severely restricting the further development and application of this technology.

[0003] Photoelectric measurements often require acquiring data from multiple sources, including data from photoelectric sensor arrays and data from mobile measuring devices. These different types of data differ in time and space, making direct fusion difficult. Traditional data acquisition methods typically employ fixed sensor layouts and single data acquisition modes, which cannot adequately adapt to complex and changing measurement environments, resulting in poor spatiotemporal consistency of the data. For example, when measuring in complex terrain environments, traditional fixed sensor arrays struggle to fully cover the target area, and data acquired by mobile measuring devices often lacks effective calibration, making it difficult to match with sensor array data. This leads to significant errors during data fusion, affecting the accuracy of the measurement results.

[0004] In the data processing stage, existing data processing algorithms have shortcomings in spectral feature extraction and target segmentation. Traditional spectral feature extraction methods are sensitive to noise and interference in the signal, making it difficult to accurately extract the key spectral features of the measured object, resulting in low target recognition accuracy. In the target segmentation process, segmentation methods based on simple thresholds or fixed rules cannot adapt to the diversity and complexity of different targets, easily leading to oversegmentation or undersegmentation, and failing to achieve fine target segmentation. For example, when measuring small targets or targets with similar spectral features, traditional methods struggle to accurately distinguish the target from the background, reducing the reliability of the measurement.

[0005] Furthermore, existing data acquisition and processing systems lack dynamic calibration mechanisms. During measurement, environmental factors (such as changes in light intensity and temperature fluctuations) and variations in the equipment's own performance can cause drift and deviations in the measurement data. However, traditional systems cannot calibrate the data in a timely manner according to the actual measurement conditions, leading to the continuous accumulation of measurement errors. This makes it difficult to guarantee the accuracy and stability of long-term measurements, limiting the application of photoelectric measurement technology in long-term monitoring and other scenarios.

[0006] In summary, existing optoelectronic measurement data acquisition and processing technologies have significant shortcomings in areas such as multi-source data fusion, spectral feature extraction, target segmentation, and dynamic calibration. There is an urgent need for a new technical solution to address these issues and improve the accuracy, reliability, and adaptability of optoelectronic measurements. Summary of the Invention

[0007] The purpose of this invention is to provide a data acquisition and processing method and system for photoelectric measurement to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a data acquisition and processing method for photoelectric measurement, the method comprising:

[0009] Acquire photoelectric data of the target area, including photoelectric sensor array data and mobile measurement device data; preprocess the acquired photoelectric data and perform spatiotemporal registration of multi-source data to extract spatiotemporal fused photoelectric signals of the target area;

[0010] The spectral features of the measured object are obtained based on the spatiotemporal fusion photoelectric signal. The spectral features are used to divide the preliminary detection area. The photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation. The spectral features are then optimized.

[0011] A feature extraction model is constructed based on time-frequency analysis. The optimized spectral features are used as the model input. Target classification is performed through multi-scale decomposition. Based on the classification results and the corresponding spectral features, photoelectric measurement data acquisition results of the target area are generated.

[0012] The energy distribution changes corresponding to the fine target segmentation results are obtained by using photoelectric measurement data of the target area over multiple time periods, and dynamic calibration parameters are generated based on the energy distribution changes.

[0013] Preferably, the acquired photoelectric data is preprocessed and multi-source data spatiotemporal registration is performed to extract spatiotemporally fused photoelectric signals of the target region, specifically as follows:

[0014] The photoelectric data is aligned by spatial grids and time series to obtain the spatiotemporal correlation of the photoelectric data corresponding to the grids or sequences, and the grids or sequences whose correlation meets the preset conditions are fused.

[0015] The phase error in the photoelectric data fusion process is obtained, and the fusion accuracy of the grid or sequence is determined based on the phase error. When the phase error is within a preset tolerance range, the next grid is fused until all grids or sequences are fused.

[0016] The fused photoelectric signal is baseline corrected by multi-scale decomposition, and a dynamic threshold is set according to the signal amplitude fluctuation to suppress noise. The corrected photoelectric signal is then obtained, and background interference components are removed and normalized.

[0017] The preprocessed photoelectric signal is acquired, local frequency domain features are extracted through an adaptive filtering network, the local frequency domain features are compared for coherence, a preset number of feature points are set, and the point where the local frequency domain feature with the highest coherence is located is obtained as a reference point based on the number of feature points.

[0018] The preprocessed photoelectric data is used as the reference signal and the target signal. The reference point features in the reference signal and the target signal are mapped to obtain the spatiotemporal correlation features. The spatiotemporal correlation features are then matched with the local frequency domain features.

[0019] Obtain the matching weight matrix, perform signal registration of photoelectric data based on the weight matrix, and extract the spatiotemporal fusion photoelectric signal of the registered target area.

[0020] Preferably, the spectral characteristics of the measured object are obtained based on the spatiotemporal fusion photoelectric signal, and the preliminary detection area is delineated using the spectral characteristics, specifically as follows:

[0021] The spatiotemporal fusion photoelectric signal of the target area is acquired, and after time-frequency transformation, a corresponding spectrum image is generated. Based on the spectrum image, the frequency band is segmented to obtain the main frequency component and energy distribution characteristics.

[0022] Based on the energy distribution characteristics, the spectrum image is cropped to suppress high-frequency noise. The target core frequency band is obtained by cropping the optimized spectrum image. The temporal distribution of the spatiotemporal fusion photoelectric signal corresponding to the core frequency band is obtained. The temporal distribution is then converted into a grayscale spectrum.

[0023] Sub-pixel-level frequency domain positioning is performed in the grayscale spectrum to obtain frequency band edge information. Based on the edge information, the frequency band is iteratively optimized by least squares ellipse fitting to obtain the energy center of the optimized ellipse. The frequency band width is determined based on the energy center to generate frequency domain features.

[0024] Spectral features are generated based on the main frequency component, energy distribution characteristics, and frequency domain characteristics. The corresponding spectral features are extracted using historical spatiotemporal fusion photoelectric signals to construct a training dataset. A convolutional neural network is trained using the training dataset to segment the measurement target.

[0025] The sliding window regression generates various spectral features, and the corresponding feature regions are labeled in the frequency domain to achieve the initial division of the detection area.

[0026] Preferably, the photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation, and the spectral features are optimized, specifically as follows:

[0027] A preliminary detection area is obtained, and clustering is performed in different frequency domain intervals of the preliminary detection area according to the energy density threshold. The peak point of each frequency domain interval is selected.

[0028] The energy density threshold is set by extracting the distance between the energy centers of the target optimization ellipse by the spectral features corresponding to the target, and by using the peak point to calculate the distance with the nearest frequency domain signal from high frequency to low frequency, and determining whether the distance is less than the energy density threshold.

[0029] If the value is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated; if the value is greater than the preset tolerance, the distance difference is obtained; if the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated.

[0030] The boundary range of the detection area is updated based on the energy distribution information in the frequency domain interval. The clustering of all signals is completed through iteration. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

[0031] Preferably, a feature extraction model is constructed based on time-frequency analysis methods, specifically as follows:

[0032] A feature extraction model is constructed based on wavelet transform to obtain optimized spectral features. These features are then imported into the feature extraction model. Convolutional kernels are used to extract time-frequency features in different frequency domain intervals. Hidden nodes of corresponding layers are set according to the time-frequency features in different frequency domain intervals, and fusion is performed through fully connected layers.

[0033] In the feature extraction model, a weight allocation layer is set up, and an attention mechanism is used to obtain the dynamic weights of time-frequency features in different frequency domain intervals, which characterize the contribution of time-frequency features.

[0034] The weighted time-frequency features are compressed again through convolution and average pooling operations. The compressed time-frequency features are then dimensionality reduced and fused before being imported into a fully connected layer.

[0035] The probability distribution of the target category is calculated by a normalized exponential function, the target classification result is determined based on the probability distribution, and the target classification result of the target area is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target area.

[0036] Preferably, the energy distribution change corresponding to the fine target segmentation result is obtained by acquiring photoelectric measurement data of the target area over multiple time periods, and dynamic calibration parameters are generated based on the energy distribution change, specifically as follows:

[0037] Based on the optimized spectral features, the optimized dominant frequency component and energy distribution features are extracted, and calibration reference points are obtained using the optimized dominant frequency component and energy distribution features to generate a calibration reference point set;

[0038] Fine target segmentation results are obtained by using photoelectric measurement data of the target area over multiple time periods, and data interpolation is performed on the calibration reference point set based on the fine target segmentation results;

[0039] The data change trend of the calibration reference point set is obtained to extract the energy distribution change. The energy distribution change is compared with the preset calibration range. When the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated.

[0040] Preferably, the present invention further includes a data acquisition and processing system for photoelectric measurement, the system comprising: a memory and a processor, wherein the memory includes a data acquisition and processing program for photoelectric measurement, and the data acquisition and processing program, when executed by the processor, performs the following steps:

[0041] Acquire photoelectric data of the target area, including photoelectric sensor array data and mobile measurement device data; preprocess the acquired photoelectric data and perform spatiotemporal registration of multi-source data to extract spatiotemporal fused photoelectric signals of the target area;

[0042] The spectral features of the measured object are obtained based on the spatiotemporal fusion photoelectric signal. The spectral features are used to divide the preliminary detection area. The photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation. The spectral features are then optimized.

[0043] A feature extraction model is constructed based on time-frequency analysis. The optimized spectral features are used as the model input. Target classification is performed through multi-scale decomposition. Based on the classification results and the corresponding spectral features, photoelectric measurement data acquisition results of the target area are generated.

[0044] The energy distribution changes corresponding to the fine target segmentation results are obtained by using photoelectric measurement data of the target area over multiple time periods, and dynamic calibration parameters are generated based on the energy distribution changes.

[0045] Preferably, the spectral characteristics of the measured object are obtained based on the spatiotemporal fusion photoelectric signal, and the preliminary detection area is delineated using the spectral characteristics, specifically as follows:

[0046] The spatiotemporal fusion photoelectric signal of the target area is acquired, and after time-frequency transformation, a corresponding spectrum image is generated. Based on the spectrum image, the frequency band is segmented to obtain the main frequency component and energy distribution characteristics.

[0047] Based on the energy distribution characteristics, the spectrum image is cropped to suppress high-frequency noise. The target core frequency band is obtained by cropping the optimized spectrum image. The temporal distribution of the spatiotemporal fusion photoelectric signal corresponding to the core frequency band is obtained. The temporal distribution is then converted into a grayscale spectrum.

[0048] Sub-pixel-level frequency domain positioning is performed in the grayscale spectrum to obtain frequency band edge information. Based on the edge information, the frequency band is iteratively optimized by least squares ellipse fitting to obtain the energy center of the optimized ellipse. The frequency band width is determined based on the energy center to generate frequency domain features.

[0049] Spectral features are generated based on the main frequency component, energy distribution characteristics, and frequency domain characteristics. The corresponding spectral features are extracted using historical spatiotemporal fusion photoelectric signals to construct a training dataset. A convolutional neural network is trained using the training dataset to segment the measurement target.

[0050] The sliding window regression generates various spectral features, and the corresponding feature regions are labeled in the frequency domain to achieve the initial division of the detection area.

[0051] Preferably, the photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation, and the spectral features are optimized, specifically as follows:

[0052] A preliminary detection area is obtained, and clustering is performed in different frequency domain intervals of the preliminary detection area according to the energy density threshold. The peak point of each frequency domain interval is selected.

[0053] The energy density threshold is set by extracting the distance between the energy centers of the target optimization ellipse by the spectral features corresponding to the target, and by using the peak point to calculate the distance with the nearest frequency domain signal from high frequency to low frequency, and determining whether the distance is less than the energy density threshold.

[0054] If the value is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated; if the value is greater than the preset tolerance, the distance difference is obtained; if the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated.

[0055] The boundary range of the detection area is updated based on the energy distribution information in the frequency domain interval. The clustering of all signals is completed through iteration. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

[0056] Preferably, a feature extraction model is constructed based on time-frequency analysis methods, specifically as follows:

[0057] A feature extraction model is constructed based on wavelet transform to obtain optimized spectral features. These features are then imported into the feature extraction model. Convolutional kernels are used to extract time-frequency features in different frequency domain intervals. Hidden nodes of corresponding layers are set according to the time-frequency features in different frequency domain intervals, and fusion is performed through fully connected layers.

[0058] In the feature extraction model, a weight allocation layer is set up, and an attention mechanism is used to obtain the dynamic weights of time-frequency features in different frequency domain intervals, which characterize the contribution of time-frequency features.

[0059] The weighted time-frequency features are compressed again through convolution and average pooling operations. The compressed time-frequency features are then dimensionality reduced and fused before being imported into a fully connected layer.

[0060] The probability distribution of the target category is calculated by a normalized exponential function, the target classification result is determined based on the probability distribution, and the target classification result of the target area is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target area.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] During the data acquisition phase, by acquiring data from the photoelectric sensor array and the mobile measuring device, combined with an adaptive spatiotemporal registration strategy, the system can adapt to complex and ever-changing measurement environments. The spatial grid employs an adaptive partitioning method based on the characteristics of the target area, and the time series is dynamically adjusted according to changes in the measurement environment. This ensures accurate alignment of multi-source data under different terrain and lighting conditions, significantly improving the spatiotemporal consistency and accuracy of the data. Compared to traditional fixed layouts and single acquisition modes, the data acquisition method of this invention can more comprehensively and accurately acquire photoelectric information of the target area, providing a high-quality data foundation for subsequent processing.

[0063] In the data processing, this invention employs a series of innovative methods for spectral feature extraction and target segmentation. Spectral features are obtained through operations such as time-frequency transformation, frequency band segmentation, and ellipse fitting, and optimized using deep learning technology. This effectively suppresses noise and interference, accurately extracting the key spectral features of the measured object. In the fine target segmentation stage, based on the analysis of energy density thresholds and frequency domain intervals, high-precision segmentation of targets of different types and characteristics is achieved through multiple iterations of optimization. This avoids over-segmentation or under-segmentation problems, significantly improving the accuracy and reliability of target recognition. Compared with traditional methods, the ability to identify small targets and targets with similar spectral characteristics is greatly enhanced.

[0064] In terms of feature extraction and target classification, the feature extraction model based on time-frequency analysis, combined with wavelet transform, attention mechanism, and other techniques, can deeply mine time-frequency features and dynamically adjust model parameters according to the features in different frequency domain intervals to achieve accurate target classification. This approach fully considers the time-frequency characteristics of photoelectric signals and, compared with traditional fixed-parameter classification models, has stronger processing capabilities for complex signals, more accurate classification results, and can be effectively applied to various complex measurement scenarios.

[0065] Furthermore, this invention introduces a dynamic calibration mechanism. By analyzing the energy distribution changes corresponding to the fine target segmentation results in multi-time period photoelectric measurement data of the target area, dynamic calibration parameters are generated in a timely manner. This mechanism can monitor the impact of environmental factors and equipment performance changes on the measurement data in real time and automatically perform calibration, effectively avoiding the accumulation of measurement errors, ensuring the accuracy and stability of long-term measurements, and making photoelectric measurement technology more practical and adaptable in fields with extremely high precision requirements such as long-term monitoring and industrial automation. Attached Figure Description

[0066] Figure 1 This is a schematic diagram illustrating the working principle of the data acquisition and processing method for photoelectric measurement described in this invention.

[0067] Figure 2 A flowchart for spatiotemporal registration and preprocessing of multi-source data;

[0068] Figure 3 Flowchart for spectral feature acquisition and preliminary detection area division;

[0069] Figure 4 This is a flowchart of the construction process for a feature extraction model based on time-frequency analysis. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Please see Figures 1-4 This invention relates to a data acquisition and processing method for photoelectric measurement, the specific implementation steps of which are as follows:

[0072] The process involves acquiring photoelectric data from the target area, including data from photoelectric sensor arrays and mobile measurement devices. Next, the acquired photoelectric data is preprocessed and multi-source data spatiotemporal registration is performed. The photoelectric data is aligned using spatial grids and time series, and the spatiotemporal correlation between corresponding photoelectric data in each grid or sequence is calculated. Grids or sequences with correlations meeting preset conditions are fused. The phase error during the photoelectric data fusion process is acquired, and the fusion accuracy of the grids or sequences is determined based on the phase error. When the phase error is within a preset tolerance range, the next grid is fused until all grids or sequences are fused. The fused photoelectric signal is then subjected to baseline correction using multi-scale decomposition, and noise suppression is performed by setting a dynamic threshold based on signal amplitude fluctuations. The corrected photoelectric signal is then acquired. Background interference components are removed and normalized. The preprocessed photoelectric signal is acquired, and local frequency domain features are extracted using an adaptive filtering network. The local frequency domain features are compared for coherence. A preset number of feature points is used, and the point with the highest coherence in the local frequency domain feature is selected as the reference point. The preprocessed photoelectric data is used as both the reference and target signals. The reference point features in the reference and target signals are mapped to obtain spatiotemporal correlation features. The spatiotemporal correlation features are matched with the local frequency domain features. A matching weight matrix is ​​obtained, and the photoelectric data is registered based on the weight matrix to extract the spatiotemporally fused photoelectric signal of the target region.

[0073] Based on the aforementioned spatiotemporal fusion photoelectric signals, the spectral characteristics of the measured object are obtained, and the spectral characteristics are used to delineate the preliminary detection area. Specifically, the spatiotemporal fusion photoelectric signals of the target area are obtained, and after time-frequency transformation, a corresponding spectrum image is generated. Based on the spectrum image, frequency band segmentation is performed to obtain the main frequency component and energy distribution characteristics. Based on the energy distribution characteristics, the spectrum image is truncated to suppress high-frequency noise. The core frequency band of the target is obtained by truncating and optimizing the spectrum image. The temporal distribution of the spatiotemporal fusion photoelectric signals corresponding to the core frequency band is obtained, and the temporal distribution is converted into a grayscale spectrum. Subpixel-level frequency domain positioning is performed in the grayscale spectrum to obtain frequency band edge information. Based on the edge information, the frequency band is iteratively optimized by least-squares ellipse fitting to obtain the energy center of the optimized ellipse. The frequency band width is determined based on the energy center, and frequency domain features are generated. Spectral features are generated based on the main frequency component, energy distribution characteristics, and frequency domain features. The corresponding spectral features are extracted using historical spatiotemporal fusion photoelectric signals to construct a training dataset. A convolutional neural network is trained using the training dataset to segment the measured target. Each spectral feature is generated according to sliding window regression, and the corresponding feature regions are labeled in the frequency domain to achieve the preliminary division of the detection area.

[0074] The photoelectric signal of the detection area is captured for pattern recognition to achieve fine target segmentation, and the spectral features are optimized. Specifically, a preliminary detection area is obtained, and clustering is performed in different frequency domain intervals within the preliminary detection area based on an energy density threshold. Peak points are selected in each frequency domain interval. The distance between the energy centers of the target optimization ellipse is extracted using the spectral features corresponding to the target, and an energy density threshold is set. The distance to the nearest frequency domain signal is calculated using the peak points from high to low frequency, and it is determined whether the distance is less than the energy density threshold. If it is less, the energy distribution information of the frequency domain interval is updated; if it is greater, the distance difference is obtained. If the distance difference is less than a preset tolerance, the energy distribution information of the frequency domain interval is updated. The boundary range of the detection area is updated based on the energy distribution information of the frequency domain interval. Clustering of all signals is completed iteratively, and fine target segmentation is achieved based on the updated boundary range of the detection area, with spectral features optimized.

[0075] A feature extraction model based on time-frequency analysis is constructed. Optimized spectral features are used as input to the model, and target classification is performed through multi-scale decomposition. Based on the classification results and corresponding spectral features, photoelectric measurement data acquisition results for the target region are generated. Specifically, the construction process involves building a feature extraction model based on wavelet transform, obtaining optimized spectral features, importing them into the feature extraction model, extracting time-frequency features in different frequency domains using convolutional kernels, setting corresponding hidden nodes based on the time-frequency features in different frequency domains, and fusing them through a fully connected layer. A weight allocation layer is set in the feature extraction model, using an attention mechanism to obtain dynamic weights of time-frequency features in different frequency domains, representing the contribution of the time-frequency features. The weighted time-frequency features are then compressed again through convolution and average pooling operations, and the compressed time-frequency features are dimensionality-reduced and fused before being imported into a fully connected layer. The probability distribution of the target category is calculated using a normalized exponential function, and the target classification result is determined based on the probability distribution. The target classification result of the target region is then combined with the corresponding spectral features to generate photoelectric measurement data acquisition results for the target region.

[0076] Finally, the energy distribution changes corresponding to the fine target segmentation results are obtained through multi-time period photoelectric measurement data of the target area, and dynamic calibration parameters are generated based on the energy distribution changes. The specific steps are as follows: Optimized dominant frequency components and energy distribution features are extracted based on the optimized spectral characteristics; calibration reference points are obtained using these optimized dominant frequency components and energy distribution features to generate a calibration reference point set; fine target segmentation results are obtained through multi-time period photoelectric measurement data of the target area; data interpolation is performed on the calibration reference point set based on the fine target segmentation results; the energy distribution changes are extracted by obtaining the data change trend of the calibration reference point set; the energy distribution changes are compared with a preset calibration range; and dynamic calibration parameters are generated when the energy distribution changes exceed the preset calibration range. Example 1:

[0077] During the data acquisition phase, the photoelectric sensor array collects data from the target area with a fixed layout and sampling frequency, while the mobile measuring device synchronously records information such as position and time, thereby obtaining photoelectric data of the target area that includes data from the photoelectric sensor array and data from the mobile measuring device.

[0078] In the multi-source data spatiotemporal registration stage, the spatial grid uses regular square grids to divide the target area. During division, the side length of each square grid is determined based on the actual size of the target area and the required accuracy. For example, if the target area is 100 square meters and a measurement accuracy of 1 square meter is required, the target area is divided into 10×10 square grids. The time series is divided with a fixed sampling interval, which is set according to the frequency characteristics of photoelectric data changes to ensure effective capture of data changes. When aligning photoelectric data, the photoelectric data is accurately mapped to the corresponding grids and time points based on the grid coordinate information and the time series timestamps. When calculating spatiotemporal correlation, a correlation coefficient algorithm is used. For every two grids or data sequences, the correlation coefficient between them is calculated to determine if the correlation meets preset conditions. Preset conditions can be set according to actual measurement needs and data characteristics; for example, a correlation coefficient greater than 0.8 is considered to meet the conditions. During the fusion process, the phase error is calculated based on the phase information of the data, obtaining the phase error by comparing the phase differences of different data. When the phase error is within the preset tolerance range, the next grid is fused until all grids or sequences are fused. The preset tolerance is also determined based on the measurement accuracy requirements.

[0079] In the processing of the fused optoelectronic signal, wavelet decomposition is used for baseline correction. Appropriate wavelet basis functions and decomposition levels are selected to decompose the signal and remove baseline drift. A dynamic threshold is set based on signal amplitude fluctuations for noise suppression. Statistical methods, such as calculating the mean and standard deviation of the signal amplitude, are used to determine a suitable threshold range. Components exceeding the threshold are considered noise and removed, thus obtaining the corrected optoelectronic signal. Simultaneously, background interference components are eliminated using a specific algorithm, and the signal is normalized to ensure it falls within a suitable numerical range.

[0080] After acquiring the preprocessed photoelectric signal, local frequency domain features are extracted using an adaptive filtering network. The adaptive filtering network employs an adaptive least mean square algorithm, which automatically adjusts the filter parameters based on the characteristics of the input signal to adapt to different signal environments. The extracted local frequency domain features are compared for coherence, with a preset number of feature points, for example, 10. By calculating the coherence between each local frequency domain feature and these 10 feature points, the point containing the local frequency domain feature with the highest coherence is selected as the reference point. Next, the preprocessed photoelectric data is used as both the reference and target signals. Spatiotemporal correlation features are obtained by mapping the reference point features between the reference and target signals, and then matched with the local frequency domain features. A matching weight matrix is ​​obtained using a specific matching algorithm, and signal registration of the photoelectric data is performed based on the weight matrix, thereby extracting an accurate spatiotemporally fused photoelectric signal.

[0081] In acquiring spectral features and dividing the detection region, a time-frequency transformation is performed on the spatiotemporal fusion optoelectronic signal using a short-time Fourier transform to convert the signal from the time domain to the frequency domain, generating a corresponding spectral image. Frequency band segmentation is performed based on the frequency distribution characteristics of the spectral image to obtain the dominant frequency component and energy distribution features. Based on the energy distribution features, the spectral image is truncated to suppress high-frequency noise. The spectral image is optimized by repeatedly adjusting the truncated range to obtain the target core frequency band. The time-domain distribution of the spatiotemporal fusion optoelectronic signal corresponding to the core frequency band is obtained and converted into a grayscale spectrum. Sub-pixel-level frequency domain localization is performed in the grayscale spectrum, and a specific localization algorithm is used to obtain the frequency band edge information. Based on the edge information, the frequency band is iteratively optimized using least-squares ellipse fitting, adjusting the ellipse parameters in each iteration to obtain the energy center of the optimized ellipse. The bandwidth is determined based on the energy center, generating frequency domain features. Spectral features are generated based on the dominant frequency component, energy distribution features, and frequency domain features. A large number of historical spatiotemporal fusion optoelectronic signals and their corresponding spectral features are collected to construct a training dataset, ensuring that the training dataset covers various measurement situations and target features, possessing diversity and representativeness. A convolutional neural network is trained using a training dataset to segment the measurement target. Spectral features are generated based on sliding window regression, and frequency domain annotations are applied to the corresponding feature regions to achieve preliminary segmentation of the detection area.

[0082] In the fine target segmentation and spectral feature optimization stage, after obtaining the initial detection area, energy density thresholds are reasonably set in different frequency domain intervals of the initial detection area based on the target's spectral characteristics and actual measurement conditions. Clustering operations are then performed, and peak points are selected in each frequency domain interval. The distance between the energy centers of the target optimization ellipse is extracted using the spectral features corresponding to the target, and this distance is used to set the energy density threshold. The distance between the peak point and the nearest frequency domain signal is calculated from high frequency to low frequency, and it is determined whether this distance is less than the energy density threshold. If it is less, the energy distribution information of the frequency domain interval is updated; if it is greater, the distance difference is obtained. If the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is also updated. The preset tolerance is set according to the measurement accuracy requirements. The boundary range of the detection area is updated based on the energy distribution information of the frequency domain interval. Clustering of all signals is completed through multiple iterations. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

[0083] When constructing the feature extraction model, wavelet transform is used. The size and number of convolutional kernels are appropriately set according to the time-frequency characteristics of different frequency domain intervals to effectively extract these features. Based on the extracted time-frequency characteristics, hidden nodes of corresponding layers are accurately set and fused through fully connected layers. A weight allocation layer is set in the feature extraction model, using an attention mechanism to obtain dynamic weights for time-frequency characteristics in different frequency domain intervals. These weights accurately represent the contribution of time-frequency characteristics. The weighted time-frequency characteristics are then compressed again through convolution and average pooling operations to reduce feature dimensionality. The compressed time-frequency characteristics are then fused and imported into fully connected layers. The probability distribution of the target category is calculated using a normalized exponential function. Based on the probability distribution, the target classification result is determined. The target classification result of the target region is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target region.

[0084] When generating dynamic calibration parameters, the optimized dominant frequency component and energy distribution characteristics are extracted based on the optimized spectral features. These characteristics are then used to obtain calibration reference points, generating a calibration reference point set. Fine target segmentation results are obtained through multi-time-segment photoelectric measurement data of the target area. Based on these fine target segmentation results, data interpolation is performed on the calibration reference point set to supplement missing data points. The data variation trend of the calibration reference point set is meticulously analyzed, energy distribution changes are extracted, and the energy distribution changes are compared with the preset calibration range. When the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated. Example 2:

[0085] The data acquisition process is consistent with the overall scheme, using an array of photoelectric sensors and a moving measuring device to acquire photoelectric data of the target area.

[0086] In the multi-source data spatiotemporal registration stage, the spatial grid uses an irregular triangular grid to divide the target area. This division method is determined based on factors such as the topography of the target area and the distribution of key measurement areas. For example, for target areas with complex terrain and many irregular boundaries, triangular grids can be flexibly used according to the actual terrain contours to ensure that each triangular grid better fits the actual measurement area. The time series is non-uniformly divided according to the needs of the measurement task. Smaller time intervals are used in periods of drastic changes in photoelectric data or periods of key focus; larger time intervals are used in periods of gradual data change, thereby reducing the amount of data processing while ensuring data validity. When aligning photoelectric data, coordinate transformation and time interpolation methods are used. For spatial coordinates, data collected by different sensors are unified to the same coordinate system through coordinate transformation algorithms; for the time dimension, time interpolation algorithms are used to interpolate data with different sampling frequencies to a unified time node, thereby accurately mapping the data to the corresponding grid and time point. When calculating spatiotemporal correlation, mutual information algorithms are used to determine the correlation by calculating the mutual information values ​​between different data. Mutual information reflects the degree of dependence between two data sequences. Setting an appropriate mutual information threshold, a value greater than the threshold indicates that the correlation between the data meets the requirements. During fusion, phase error calculation considers the frequency and phase variation trends of the data. By analyzing the phase differences at different frequencies and the phase changes over time, the phase error is accurately calculated. For baseline correction, multi-scale decomposition employs empirical mode decomposition (EMD), which decomposes the signal into multiple intrinsic mode functions based on its characteristics. The dynamic threshold setting combines local signal features. By analyzing the signal's characteristics within local time periods or spatial regions, such as amplitude variation range and frequency distribution, a suitable dynamic threshold is determined to effectively suppress noise. When extracting local frequency domain features, the adaptive filtering network uses a recursive least squares algorithm. This algorithm continuously updates the filter parameters, enabling the filter to quickly adapt to signal changes and accurately extract local frequency domain features. During signal registration, the calculation method of the matching weight matrix is ​​optimized, for example, by introducing a regularization term, to avoid overfitting during matrix calculation and improve the accuracy of signal registration.

[0087] In the stages of acquiring spectral features and dividing the detection region, a time-frequency transformation is performed on the spatiotemporally fused optoelectronic signal using Gabor transform. Gabor transform can analyze the signal simultaneously in the time and frequency domains, generating a spectral image with good time-frequency resolution. Frequency band segmentation is performed based on the energy distribution of the spectral image. By analyzing the concentrated energy regions and distribution in the spectral image, the frequency band division boundaries are determined, thereby obtaining the dominant frequency component and energy distribution characteristics. Based on the energy distribution characteristics, the spectral image is cropped, focusing on preserving areas of concentrated energy and suppressing high-frequency noise. The spectral image is optimized by repeatedly adjusting the cropping range and parameters to obtain the target core frequency band. The time-domain distribution of the spatiotemporally fused optoelectronic signal corresponding to the core frequency band is obtained and converted into a grayscale spectrum. Sub-pixel-level frequency domain localization is performed in the grayscale spectrum using a specific localization algorithm, such as a gradient-based localization algorithm, to obtain frequency band edge information. Based on the edge information, the frequency band is iteratively optimized using least-squares ellipse fitting. Each iteration adjusts the ellipse parameters based on the fitting error to obtain the energy center of the optimized ellipse. The bandwidth is determined based on the energy center, generating frequency domain features. Spectral features are generated based on the dominant frequency component, energy distribution characteristics, and frequency domain features. When constructing the training dataset, historical data is filtered and preprocessed to remove outliers, such as erroneous data caused by sensor malfunctions or external interference, ensuring dataset quality. The filtered historical spatiotemporal fused photoelectric signals are used to extract corresponding spectral features to construct the training dataset. A convolutional neural network is then trained on this dataset to segment the measurement target. Each spectral feature is generated using sliding window regression, and the corresponding feature regions are frequency domain labeled to achieve preliminary segmentation of the detection area.

[0088] In the fine target segmentation and spectral feature optimization stage, after obtaining the initial detection area, the energy density threshold is adjusted according to the shape and size of the target in different frequency domain intervals of the initial detection area, and clustering is performed. Peak points are selected in each frequency domain interval. The distance between the energy centers of the target optimization ellipse is extracted using the spectral features corresponding to the target, and this distance is used to set the energy density threshold. The distance between the peak point and the nearest frequency domain signal is calculated from high frequency to low frequency, and it is determined whether this distance is less than the energy density threshold. If it is less, the energy distribution information of the frequency domain interval is updated; if it is greater, the distance difference is obtained. If the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is also updated. The preset tolerance is set according to the measurement accuracy requirements. The boundary range of the detection area is updated according to the energy distribution information of the frequency domain interval. Through multiple iterations and optimizations, the boundary of the detection area is gradually refined, completing the clustering of all signals. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

[0089] When constructing the feature extraction model, improvements are made based on wavelet transform. The shape and parameters of the convolutional kernels are adjusted, for example, using kernels of different sizes and shapes to adapt to the time-frequency feature extraction requirements of different frequency domain intervals. The connection method of hidden nodes is optimized, employing a more complex connection structure to enhance the model's ability to express time-frequency features. An attention mechanism is utilized to better highlight important time-frequency features; by calculating the weights of different time-frequency features, the model focuses more on important features during training and prediction. The weighted time-frequency features are then compressed again through convolution and average pooling operations. The compressed time-frequency features are then dimensionality-reduced and fused before being imported into a fully connected layer. The probability distribution of the target category is calculated using a normalized exponential function. Based on the probability distribution, the target classification result is determined, and the target classification result of the target region is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target region.

[0090] When generating dynamic calibration parameters, the optimized dominant frequency component and energy distribution characteristics are extracted based on the optimized spectral features. These features are then used to obtain calibration reference points, generating a calibration reference point set. Fine target segmentation results are obtained through multi-time-segment photoelectric measurement data of the target area. Based on these fine target segmentation results, data interpolation is performed on the calibration reference point set to supplement missing data points. Taking into account multiple factors, such as measurement time and environmental conditions, the calibration reference point set is analyzed to more accurately extract energy distribution changes. These energy distribution changes are compared with a preset calibration range. When the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated. Example 3:

[0091] Starting with data acquisition, the photoelectric sensor array and mobile measuring device are still used to obtain photoelectric data of the target area. The layout of the sensor array and the moving path of the measuring device are pre-planned and set according to the characteristics of the target area.

[0092] In the multi-source data spatiotemporal registration stage, the target area is divided using a hexagonal grid. Compared to other conventional grids, hexagonal grids offer better balance and symmetry in spatial filling, making them particularly suitable for measurement scenarios with high isotropic requirements. During division, the side length of the hexagonal grid is determined based on the scope of the target area and the required measurement accuracy. For example, if the target area is a circular region, to ensure omnidirectional uniform measurement, it can be divided into a closely spaced hexagonal grid. The side length of each hexagonal grid is determined according to the required measurement resolution; for instance, a hexagonal grid with a 1-meter side length can achieve a measurement accuracy of approximately 1 square meter. The time series is divided according to event triggering; that is, when there is a significant change in photoelectric data, the measuring device enters a specific area, or preset measurement conditions are met, a new time series node is triggered. When aligning photoelectric data, spatial geometric relationships are utilized to calculate the coordinate positions of each data point in the hexagonal grid, accurately mapping them to the corresponding grid cells; simultaneously, based on the logical relationships of time events, the data is associated with the corresponding time nodes. When calculating spatiotemporal correlation, a dynamic time warping algorithm is employed. This algorithm effectively calculates the similarity between different data sequences even when the time series data exhibits nonlinear variations. During the calculation, a distance matrix is ​​constructed to find the optimal time warping path, thereby obtaining the correlation between data. In the fusion process, phase error is calculated based on the phase change rate of the data. The phase error value is calculated by analyzing the phase change trends of adjacent time points or adjacent grid data. For baseline correction, a combination of Fourier transform and wavelet transform is used. First, the Fourier transform is used to convert the signal from the time domain to the frequency domain, analyzing and preliminarily processing the overall frequency components of the signal. Then, wavelet transform is used to perform multi-scale decomposition of the signal, further refining the signal characteristics and removing baseline drift and high-frequency noise. The dynamic threshold is dynamically adjusted according to the signal's changing trend. By monitoring the amplitude fluctuation range of the signal in real time and combining statistical methods such as moving average and standard deviation, a suitable threshold is determined to achieve effective noise suppression. When extracting local frequency domain features, an adaptive filtering network is constructed using a neural network-based adaptive algorithm. This network learns the characteristic patterns of signals through training and automatically adjusts the filter parameters to adapt to the local frequency domain feature extraction requirements under different signal environments. During signal registration, constraints such as spatial location constraints and temporal order constraints are introduced to optimize the calculation process of the matching weight matrix, ensuring accurate registration of optoelectronic data and thus extracting accurate spatiotemporal fused optoelectronic signals.

[0093] In acquiring spectral features and delineating the detection region, a time-frequency transformation is performed on the spatiotemporal fusion photoelectric signal using fractional Fourier transform. As a generalized Fourier transform, fractional Fourier transform can analyze signals at different fractional orders, making it suitable for signal processing with nonlinear frequency modulation characteristics. Compared to traditional Fourier transform, it can more effectively capture the time-frequency features of the signal. Frequency band segmentation is performed based on the frequency characteristics of the spectral image and the characteristics of the target. By analyzing the distribution of different frequency components in the spectral image and combining it with the inherent characteristics of the target in the frequency domain, such as energy concentration areas at specific frequencies, the boundary of the frequency band is determined, thereby obtaining the dominant frequency component and energy distribution characteristics. Based on the energy distribution characteristics, the spectral image is truncated, focusing on retaining frequency regions related to the target and suppressing high-frequency noise. The spectral image is optimized by repeatedly adjusting the truncating parameters to obtain the target's core frequency band. The time-domain distribution of the spatiotemporal fusion photoelectric signal corresponding to the core frequency band is obtained and converted into a grayscale spectrum. Subpixel-level frequency domain localization is performed in the grayscale spectral image using high-precision localization algorithms, such as those based on image edge detection and morphological processing, to obtain frequency band edge information. Based on this edge information, the frequency band is iteratively optimized using least-squares ellipse fitting. Each iteration adjusts the ellipse parameters according to the fitting error until a satisfactory fit is achieved, obtaining the energy center of the optimized ellipse. The bandwidth is determined based on the energy center, generating frequency domain features. Spectral features are generated based on the dominant frequency component, energy distribution characteristics, and frequency domain features. When constructing the training dataset, data augmentation techniques are used to expand it. For example, translation, scaling, and rotation operations are performed on historical spatiotemporal fused photoelectric signals to simulate different measurement angles and positions; or appropriate Gaussian noise is added to simulate interference in actual measurements, thereby increasing the diversity of the dataset and improving the model's generalization ability. A convolutional neural network is trained using the expanded training dataset to segment the measurement target. Spectral features are generated based on sliding window regression, and frequency domain annotations are performed on the corresponding feature regions to achieve preliminary segmentation of the detection area.

[0094] In the process of fine target segmentation and spectral feature optimization, after obtaining the initial detection area, energy density thresholds are set in different frequency domain intervals of the initial detection area according to the target's material and optical properties. Clustering is then performed, and peak points are selected in each frequency domain interval. Targets of different materials have unique energy distribution characteristics in their spectra. By analyzing the spectral characteristics corresponding to the target material, energy density thresholds are reasonably set to accurately identify the target. The distance between the energy centers of the target optimization ellipse is extracted using the spectral features corresponding to the target, and this distance is used to set the energy density threshold. The distance between the peak point and the nearest frequency domain signal is calculated from high frequency to low frequency, and it is determined whether this distance is less than the energy density threshold. If it is less, the energy distribution information of the frequency domain interval is updated; if it is greater, the distance difference is obtained. If the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is also updated. The preset tolerance is set according to the measurement accuracy requirements. The boundary range of the detection area is updated according to the energy distribution information of the frequency domain interval. Through multiple complex clustering and iterative algorithms, the boundary of the detection area is gradually refined, completing the clustering of all signals. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

[0095] When constructing the feature extraction model, a deep neural network model is built based on wavelet transform. The number of layers and complexity of the network are increased; for example, a multi-layer neural network structure containing multiple convolutional layers, pooling layers, and fully connected layers is constructed to enhance the model's ability to extract and represent time-frequency features. Convolutional kernels are used to extract time-frequency features in different frequency domains, and hidden nodes of corresponding numbers are set according to the time-frequency features in different frequency domains, which are then fused through fully connected layers. A weight allocation layer is set in the feature extraction model, using an attention mechanism to more accurately allocate the weights of time-frequency features. By calculating the importance of different time-frequency features in target classification, corresponding weights are assigned, making the model pay more attention to key features during training and prediction. The weighted time-frequency features are compressed again through convolution and average pooling operations, and the compressed time-frequency features are dimensionality reduced and fused before being imported into fully connected layers. The probability distribution of the target category is calculated using a normalized exponential function, and the target classification result is determined based on the probability distribution. The target classification result of the target region is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target region.

[0096] When generating dynamic calibration parameters, optimized dominant frequency components and energy distribution characteristics are extracted based on optimized spectral features. These features are then used to obtain calibration reference points, generating a calibration reference point set. Fine target segmentation results are obtained through multi-time-segment photoelectric measurement data of the target area. Data interpolation is performed on the calibration reference point set based on these fine target segmentation results to supplement missing data points. Machine learning algorithms, such as regression analysis and decision trees, are used to analyze the calibration reference point set, uncovering potential relationships and patterns of change among the data to more accurately predict energy distribution changes. The energy distribution changes are compared with a preset calibration range. When the energy distribution changes exceed the preset calibration range, dynamic calibration parameters are generated. Example 4:

[0097] During the data acquisition phase, the photoelectric sensor array is arranged around or inside the target area according to a specific distribution pattern. The mobile measuring device moves and measures the target area according to a pre-planned path, thereby acquiring photoelectric data of the target area that includes data from the photoelectric sensor array and data from the mobile measuring device.

[0098] In the multi-source data spatiotemporal registration stage, a hybrid grid structure is adopted for the spatial grid, including square, triangular, and hexagonal grids. During the partitioning process, the appropriate grid type is selected based on the shape, complexity, and measurement requirements of different parts of the target area. For example, square grids are used for areas with regular shapes and uniform measurement accuracy requirements; triangular grids are used for areas with irregular boundaries to better fit the boundaries; and hexagonal grids are used for areas requiring isotropic measurements. The time series employs a hybrid partitioning method, combining fixed intervals and event-triggered sampling. Fixed-interval sampling is used during periods of relatively stable data change, while event-triggered sampling is used during periods where data changes may occur or during periods of key interest, such as when the sensor detects a sudden change in environmental parameters.

[0099] When aligning photoelectric data, multiple mapping methods are used comprehensively. For spatial coordinates, corresponding coordinate transformation relationships are established for different types of grids to map sensor data to a unified coordinate system. For the time dimension, a time interpolation algorithm is used to unify fixed-interval sampling data and event-triggered sampling data. When calculating spatiotemporal correlation, a combined algorithm is used, combining correlation coefficient algorithm, mutual information algorithm, and dynamic time warping algorithm. First, the correlation coefficient algorithm is used to initially screen out data pairs with high correlation. Then, the mutual information algorithm is used to further measure the degree of dependence between data. For data with large time series differences, the dynamic time warping algorithm is used to calculate their similarity.

[0100] During the fusion process, the calculation of phase error comprehensively considers multiple dimensions of data information. For the first The spatial grid in the ... Phase value and phase error of data at each time point The calculation formula is:

[0101]

[0102] in The total number of spatial grids, The total number of time points. For the first The average phase values ​​of all spatial grid data at each time point. For baseline correction, multi-scale decomposition employs a multi-scale morphological decomposition method, using structuring elements of different scales to perform opening and closing operations on the signal to remove baseline drift. The dynamic threshold is set based on the local and global characteristics of the signal; first, the mean and variance of the local signal region are calculated, and then the threshold is determined by combining this with the statistical characteristics of the global signal.

[0103] When extracting local frequency domain features, the adaptive filtering network employs a hybrid adaptive algorithm, combining the advantages of the least mean square algorithm and the recursive least squares algorithm, automatically switching algorithms based on the signal's stationarity and rate of change. During signal registration, the accuracy of the matching weight matrix is ​​improved by optimizing the algorithm and adding spatial location constraints and temporal order constraints. For example, spatially adjacent grid data are given higher weights during registration, and temporally continuous data should maintain consistent order.

[0104] When acquiring spectral features and dividing the detection area, time-frequency transformation is performed on the spatiotemporal fusion photoelectric signal, using a combination of multiple transformation methods. For example, a short-time Fourier transform is first performed to obtain the initial time-frequency distribution of the signal, and then a fractional Fourier transform is used to refine the analysis of specific frequency components. Frequency band segmentation is performed based on various characteristics of the target, such as the spectral features corresponding to the target's material, shape, and motion state. Based on the energy distribution characteristics, the spectral image is regionally truncated, high-frequency noise is suppressed, and the core frequency band of the target is obtained. The temporal domain distribution of the spatiotemporal fusion photoelectric signal corresponding to the core frequency band is converted into a grayscale spectrum. Sub-pixel-level frequency domain localization is performed in the grayscale spectrum to obtain frequency band edge information. The frequency band is iteratively optimized using least-squares ellipse fitting to obtain the energy center of the optimized ellipse. The frequency band width is determined based on the energy center, generating frequency domain features.

[0105] When constructing the training dataset, feature selection and dimensionality reduction are performed. Principal Component Analysis (PCA) and other methods are used to remove redundant features, retaining only the most valuable features for target classification, reducing data dimensionality, and improving training efficiency. The corresponding spectral features are extracted from the processed historical spatiotemporal fused photoelectric signals to construct the training dataset. A convolutional neural network is then trained on this dataset to segment the measurement target. Each spectral feature is generated based on sliding window regression, and frequency domain annotation is performed on the corresponding feature regions to achieve preliminary segmentation of the detection area.

[0106] In the fine target segmentation and spectral feature optimization stage, after obtaining the initial detection area, the energy density threshold is dynamically adjusted based on various target attributes (such as material, size, reflectivity, etc.) in different frequency domain intervals of the initial detection area. Clustering is then performed, and interval peak points are selected. The distance between the energy centers of the target optimization ellipse is extracted using the spectral features corresponding to the target to set the energy density threshold. The distance between the peak points and the nearest frequency domain signal is calculated from high to low frequency, and it is determined whether this distance is less than the energy density threshold. If it is less, the energy distribution information of the frequency domain interval is updated; if it is greater, the distance difference is obtained. If the distance difference is less than a preset tolerance, the energy distribution information of the frequency domain interval is updated. The boundary range of the detection area is updated based on the energy distribution information of the frequency domain interval. Through complex iteration and optimization strategies, clustering of all signals is completed. Fine target segmentation is performed based on the updated boundary range of the detection area, and the spectral features are optimized.

[0107] When constructing the feature extraction model, a heterogeneous neural network model is built based on wavelet transform, combining the advantages of convolutional neural networks (CNN) and recurrent neural networks (RNN). CNN is used to extract the spatial information of time-frequency features, while RNN is used to process the time-series information of time-frequency features. Time-frequency features in different frequency domain intervals are extracted using convolutional kernels, and hidden nodes of corresponding layers are set according to the features. Fusion is performed through fully connected layers. A weight allocation layer is set in the feature extraction model, and the attention mechanism is used to more comprehensively capture the importance of time-frequency features. The weighted time-frequency features are then subjected to convolution, pooling, dimensionality reduction, and fusion operations. The probability distribution of the target category is calculated using a normalized exponential function, and the target classification result is determined based on the probability distribution. The target classification result of the target region is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target region.

[0108] When generating dynamic calibration parameters, the optimized dominant frequency component and energy distribution characteristics are extracted based on the optimized spectral features. These features are then used to obtain calibration reference points, generating a calibration reference point set. Fine target segmentation results are obtained through multi-time-segment photoelectric measurement data of the target area, and data interpolation is performed on the calibration reference point set based on these results. Big data analytics techniques, such as association rule mining and time-series pattern analysis, are used to conduct in-depth analysis of the calibration reference point set, more accurately determining energy distribution changes. These energy distribution changes are compared with a preset calibration range; when the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated. Example 5:

[0109] Data acquisition is conducted based on the actual characteristics of the target area. The deployment of the photoelectric sensor array is not fixed; rather, it is tailored to the specific characteristics of the target area, considering its potential location, activity patterns, and environmental factors. For example, in mountainous terrain with complex topography, to comprehensively capture the target's photoelectric information, the sensor array is dispersed along key locations such as valleys and ridges. The mobile measuring device also plans a flexible movement path based on the pre-set measurement tasks and the target's activity range, ensuring that photoelectric data from different angles and locations is acquired during movement. Ultimately, the data from both the photoelectric sensor array and the mobile measuring device are fully acquired and used as the raw data for subsequent processing.

[0110] In the multi-source data spatiotemporal registration stage, the spatial grid employs an adaptive grid generation method based on the characteristics of the target region. This method utilizes advanced image recognition and analysis technology to process the image data of the target region. First, it identifies different terrain features, object distributions, and other characteristics within the target region. Then, based on the complexity of these features and the required measurement accuracy, a grid is dynamically generated. In areas with dense objects and rich feature variations, the grid is finer to ensure accurate capture of changes in photoelectric data; while in relatively flat areas with simple features, the grid is sparser, thus reducing data processing volume while maintaining measurement accuracy. The time series is also adaptively adjusted according to changes in the measurement environment. By monitoring environmental parameters in real time, such as light intensity, temperature, and wind speed, when these parameters change significantly and may affect photoelectric data acquisition, the system automatically adjusts the sampling interval of the time series. For example, when light intensity changes drastically, the sampling interval is shortened to ensure timely capture of photoelectric data fluctuations caused by changes in light intensity.

[0111] When aligning photoelectric data, a smart algorithm is used for dynamic mapping. This algorithm, based on a deep learning model, learns from a large amount of historical data to master the mapping patterns of photoelectric data at different spatial locations and time points. For newly acquired data, the model can automatically and accurately map it to the corresponding spatial grid and time point based on the data's characteristics. When calculating spatiotemporal correlation, a deep learning-based correlation calculation method is used. A dedicated deep learning network model is constructed, taking different photoelectric data as input. Through multi-layered neurons and feature extraction within the network, the correlation values ​​between data are output. During training, the network model continuously optimizes its parameters to improve the accuracy of data correlation calculation. During the fusion process, the phase error is calculated based on the deep learning model's prediction and analysis of the data phase. The model learns the phase change patterns in historical data, predicts the phase of new data, and then compares the predicted value with the actual value to calculate the phase error.

[0112] During baseline correction, an adaptive multi-scale decomposition method is employed. This method automatically determines and selects the most suitable decomposition method and parameters based on the characteristics of the signal itself. For example, for signals with obvious periodicity, Fourier transform correlation decomposition is chosen; for non-stationary signals, wavelet transform and other methods are used. Dynamic thresholding is optimized using machine learning algorithms. A machine learning model is established, taking various signal features, such as amplitude, frequency, and trend, as input and outputting a dynamic threshold. During training, the model continuously adjusts parameters based on the relationship between different signal features and the optimal threshold to achieve precise threshold optimization, thereby effectively suppressing noise. When extracting local frequency domain features, an adaptive filtering network based on deep learning is used. This network consists of multiple deep learning modules. Through training on a large amount of data, the network can automatically learn the frequency domain features of the signal and adjust the filtering parameters in real time according to changes in the input signal, accurately extracting local frequency domain features. During signal registration, a matching weight matrix is ​​generated using a deep learning model. By learning from successful signal registration cases in historical data, the model understands the relationship between different signal features and weight matrices, thereby generating accurate matching weight matrices for new optoelectronic data, achieving high-precision signal registration, and extracting high-quality spatiotemporal fusion optoelectronic signals.

[0113] In terms of acquiring spectral features and delineating detection regions, a deep learning-based time-frequency transformation method is used for time-frequency transformation of spatiotemporally fused photoelectric signals. A deep learning network is constructed and trained on a large number of time-frequency transformation samples, enabling it to directly perform time-frequency transformation on the input photoelectric signals and generate accurate spectral images. Frequency band segmentation is also performed based on the analysis results of the spectral images by the deep learning model. By learning the frequency band features of different targets in a large number of spectral images, the model can automatically identify the frequency bands related to the targets in the spectral images and determine their boundaries, thereby obtaining the dominant frequency component and energy distribution features. Based on the energy distribution features, the spectral images are cropped to suppress high-frequency noise and obtain the target's core frequency band. The temporal distribution of the spatiotemporally fused photoelectric signal corresponding to the core frequency band is obtained and converted into a grayscale spectrum. Sub-pixel-level frequency domain localization is performed in the grayscale spectrum to obtain frequency band edge information. The deep learning model is used to analyze the grayscale spectrum, accurately obtaining the frequency band edge information by identifying features such as edges and textures in the image. The frequency band is iteratively optimized using least-squares elliptic fitting based on edge information to obtain the energy center of the optimized ellipse. The bandwidth is then determined based on the energy center, generating frequency domain features. Spectral features are generated based on the dominant frequency component, energy distribution characteristics, and frequency domain features.

[0114] When constructing the training dataset, an online learning approach is used to continuously update the dataset. During system operation, each new set of valid data is added to the training dataset, and the original dataset is filtered and reorganized based on the characteristics of the new data, removing redundant or no longer representative data. The updated training dataset is used to train a convolutional neural network, enabling the network to continuously adapt to new data features and improve the accuracy of target segmentation. Spectral features are generated based on sliding window regression, and frequency domain annotation is performed on the corresponding feature regions to achieve preliminary segmentation of the detection region.

[0115] In the fine target segmentation and spectral feature optimization stage, after obtaining the initial detection region, the energy density threshold in different frequency domain intervals of the initial detection region is dynamically set based on the analysis results of the target by the deep learning model. The deep learning model determines the appropriate energy density threshold for each frequency domain interval through comprehensive analysis of the target's spectral features, shape, material, and other information. Then, clustering is performed in each frequency domain interval to select the interval peak points. The distance between the peak points and the nearest frequency domain signal is calculated from high to low frequency, and it is determined whether this distance is less than the energy density threshold. If it is less, the energy distribution information of the frequency domain interval is updated; if it is greater, the distance difference is obtained. If the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated. The boundary range of the detection region is updated based on the energy distribution information of the frequency domain interval. Through multiple iterations and optimizations using the deep learning algorithm, the clustering of all signals is completed. Fine target segmentation is then performed based on the updated boundary range of the detection region, and the spectral features are optimized.

[0116] When constructing the feature extraction model, a complex neural network model is built based on deep learning. The model contains multiple network layers with different functions, such as convolutional layers for extracting time-frequency features, pooling layers for reducing data dimensionality, and recurrent layers for processing time-series information. Attention mechanisms and other deep learning techniques, such as self-attention and residual connections, are utilized to better extract and fuse time-frequency features. During model training, the network parameters are continuously adjusted using the backpropagation algorithm to ensure the model can accurately extract features from various targets. The weighted time-frequency features are then subjected to a series of convolution, pooling, and dimensionality reduction operations before being fed into a fully connected layer. The probability distribution of the target category is calculated using a normalized exponential function, and the target classification result is determined based on the probability distribution. The target classification result of the target region is then combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result for the target region.

[0117] When generating dynamic calibration parameters, the optimized dominant frequency component and energy distribution characteristics are extracted based on the optimized spectral features. These features are then used to obtain calibration reference points, generating a calibration reference point set. Fine target segmentation results are obtained through multi-time-segment photoelectric measurement data of the target area, and data interpolation is performed on the calibration reference point set based on these results. A deep learning prediction model is used to analyze the calibration reference point set. This model predicts future energy distribution changes by learning the variation patterns of the calibration reference point set in historical data. The predicted energy distribution changes are compared with a preset calibration range. When the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated.

[0118] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0119] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data acquisition and processing method for photoelectric measurement, characterized in that, Includes the following steps: Acquire photoelectric data of the target area, including photoelectric sensor array data and mobile measurement device data; preprocess the acquired photoelectric data and perform spatiotemporal registration of multi-source data to extract spatiotemporal fused photoelectric signals of the target area; The spectral features of the measured object are obtained based on the spatiotemporal fusion photoelectric signal. The spectral features are used to divide the preliminary detection area. The photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation. The spectral features are then optimized. A feature extraction model is constructed based on time-frequency analysis. The optimized spectral features are used as the model input. Target classification is performed through multi-scale decomposition. Based on the classification results and the corresponding spectral features, photoelectric measurement data acquisition results of the target area are generated. The energy distribution change corresponding to the fine target segmentation result is obtained by acquiring photoelectric measurement data of the target area over multiple time periods, and dynamic calibration parameters are generated based on the energy distribution change. The energy distribution changes corresponding to the fine target segmentation results are obtained by acquiring photoelectric measurement data of the target area over multiple time periods. Dynamic calibration parameters are generated based on the energy distribution changes, specifically: Based on the optimized spectral features, the optimized dominant frequency component and energy distribution features are extracted, and calibration reference points are obtained using the optimized dominant frequency component and energy distribution features to generate a calibration reference point set; Fine target segmentation results are obtained by using photoelectric measurement data of the target area over multiple time periods, and data interpolation is performed on the calibration reference point set based on the fine target segmentation results; The data change trend of the calibration reference point set is obtained to extract the energy distribution change. The energy distribution change is compared with the preset calibration range. When the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated.

2. The data acquisition and processing method for photoelectric measurement according to claim 1, characterized in that, The acquired photoelectric data is preprocessed and multi-source data spatiotemporal registration is performed to extract spatiotemporally fused photoelectric signals of the target region, specifically as follows: The photoelectric data is aligned by spatial grids and time series to obtain the spatiotemporal correlation of the photoelectric data corresponding to the grids or sequences, and the grids or sequences whose correlation meets the preset conditions are fused. The phase error in the photoelectric data fusion process is obtained, and the fusion accuracy of the grid or sequence is determined based on the phase error. When the phase error is within a preset tolerance range, the next grid is fused until all grids or sequences are fused. The fused photoelectric signal is baseline corrected by multi-scale decomposition, and a dynamic threshold is set according to the signal amplitude fluctuation to suppress noise. The corrected photoelectric signal is then obtained, and background interference components are removed and normalized. The preprocessed photoelectric signal is acquired, local frequency domain features are extracted through an adaptive filtering network, the local frequency domain features are compared for coherence, a preset number of feature points are set, and the point where the local frequency domain feature with the highest coherence is located is obtained as a reference point based on the number of feature points. The preprocessed photoelectric data is used as the reference signal and the target signal. The reference point features in the reference signal and the target signal are mapped to obtain the spatiotemporal correlation features. The spatiotemporal correlation features are then matched with the local frequency domain features. Obtain the matching weight matrix, perform signal registration of photoelectric data based on the weight matrix, and extract the spatiotemporal fusion photoelectric signal of the registered target area.

3. The data acquisition and processing method for photoelectric measurement according to claim 1, characterized in that, The spectral characteristics of the measured object are obtained based on the spatiotemporal fusion photoelectric signal, and the preliminary detection area is delineated using the spectral characteristics, specifically as follows: The spatiotemporal fusion photoelectric signal of the target area is acquired, and after time-frequency transformation, a corresponding spectrum image is generated. Based on the spectrum image, the frequency band is segmented to obtain the main frequency component and energy distribution characteristics. Based on the energy distribution characteristics, the spectrum image is cropped to suppress high-frequency noise. The target core frequency band is obtained by cropping the optimized spectrum image. The temporal distribution of the spatiotemporal fusion photoelectric signal corresponding to the core frequency band is obtained. The temporal distribution is then converted into a grayscale spectrum. Sub-pixel-level frequency domain positioning is performed in the grayscale spectrum to obtain frequency band edge information. Based on the edge information, the frequency band is iteratively optimized by least squares ellipse fitting to obtain the energy center of the optimized ellipse. The frequency band width is determined based on the energy center to generate frequency domain features. Spectral features are generated based on the main frequency component, energy distribution characteristics, and frequency domain characteristics. The corresponding spectral features are extracted using historical spatiotemporal fusion photoelectric signals to construct a training dataset. A convolutional neural network is trained using the training dataset to segment the measurement target. The sliding window regression generates various spectral features, and the corresponding feature regions are labeled in the frequency domain to achieve the initial division of the detection area.

4. The data acquisition and processing method for photoelectric measurement according to claim 1, characterized in that, The photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation, and the spectral features are optimized, specifically as follows: A preliminary detection area is obtained, and clustering is performed in different frequency domain intervals of the preliminary detection area according to the energy density threshold. The peak point of each frequency domain interval is selected. The energy density threshold is set by extracting the distance between the energy centers of the target optimization ellipse by the spectral features corresponding to the target, and by using the peak point to calculate the distance with the nearest frequency domain signal from high frequency to low frequency, and determining whether the distance is less than the energy density threshold. If the value is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated; if the value is greater than the preset tolerance, the distance difference is obtained; if the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated. The boundary range of the detection area is updated based on the energy distribution information in the frequency domain interval. The clustering of all signals is completed through iteration. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

5. The data acquisition and processing method for photoelectric measurement according to claim 1, characterized in that, A feature extraction model is constructed based on time-frequency analysis, specifically as follows: A feature extraction model is constructed based on wavelet transform to obtain optimized spectral features. These features are then imported into the feature extraction model. Convolutional kernels are used to extract time-frequency features in different frequency domain intervals. Hidden nodes of corresponding layers are set according to the time-frequency features in different frequency domain intervals, and fusion is performed through fully connected layers. In the feature extraction model, a weight allocation layer is set up, and an attention mechanism is used to obtain the dynamic weights of time-frequency features in different frequency domain intervals, which characterize the contribution of time-frequency features. The weighted time-frequency features are compressed again through convolution and average pooling operations. The compressed time-frequency features are then dimensionality reduced and fused before being imported into a fully connected layer. The probability distribution of the target category is calculated by a normalized exponential function, the target classification result is determined based on the probability distribution, and the target classification result of the target area is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target area.

6. A data acquisition and processing system for photoelectric measurement, characterized in that, The system includes: a memory and a processor. The memory includes a data acquisition and processing program for photoelectric measurement. When the data acquisition and processing program is executed by the processor, it performs the following steps: Acquire photoelectric data of the target area, including photoelectric sensor array data and mobile measurement device data; preprocess the acquired photoelectric data and perform spatiotemporal registration of multi-source data to extract spatiotemporal fused photoelectric signals of the target area; The spectral features of the measured object are obtained based on the spatiotemporal fusion photoelectric signal. The spectral features are used to divide the preliminary detection area. The photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation. The spectral features are then optimized. A feature extraction model is constructed based on time-frequency analysis. The optimized spectral features are used as the model input. Target classification is performed through multi-scale decomposition. Based on the classification results and the corresponding spectral features, photoelectric measurement data acquisition results of the target area are generated. The energy distribution change corresponding to the fine target segmentation result is obtained by acquiring photoelectric measurement data of the target area over multiple time periods, and dynamic calibration parameters are generated based on the energy distribution change. The energy distribution changes corresponding to the fine target segmentation results are obtained by acquiring photoelectric measurement data of the target area over multiple time periods. Dynamic calibration parameters are generated based on the energy distribution changes, specifically: Based on the optimized spectral features, the optimized dominant frequency component and energy distribution features are extracted, and calibration reference points are obtained using the optimized dominant frequency component and energy distribution features to generate a calibration reference point set; Fine target segmentation results are obtained by using photoelectric measurement data of the target area over multiple time periods, and data interpolation is performed on the calibration reference point set based on the fine target segmentation results; The data change trend of the calibration reference point set is obtained to extract the energy distribution change. The energy distribution change is compared with the preset calibration range. When the energy distribution change exceeds the preset calibration range, dynamic calibration parameters are generated.

7. The data acquisition and processing system for photoelectric measurement according to claim 6, characterized in that, The spectral characteristics of the measured object are obtained based on the spatiotemporal fusion photoelectric signal, and the preliminary detection area is delineated using the spectral characteristics, specifically as follows: The spatiotemporal fusion photoelectric signal of the target area is acquired, and after time-frequency transformation, a corresponding spectrum image is generated. Based on the spectrum image, the frequency band is segmented to obtain the main frequency component and energy distribution characteristics. Based on the energy distribution characteristics, the spectrum image is cropped to suppress high-frequency noise. The target core frequency band is obtained by cropping the optimized spectrum image. The temporal distribution of the spatiotemporal fusion photoelectric signal corresponding to the core frequency band is obtained. The temporal distribution is then converted into a grayscale spectrum. Sub-pixel-level frequency domain positioning is performed in the grayscale spectrum to obtain frequency band edge information. Based on the edge information, the frequency band is iteratively optimized by least squares ellipse fitting to obtain the energy center of the optimized ellipse. The frequency band width is determined based on the energy center to generate frequency domain features. Spectral features are generated based on the main frequency component, energy distribution characteristics, and frequency domain characteristics. The corresponding spectral features are extracted using historical spatiotemporal fusion photoelectric signals to construct a training dataset. A convolutional neural network is trained using the training dataset to segment the measurement target. The sliding window regression generates various spectral features, and the corresponding feature regions are labeled in the frequency domain to achieve the initial division of the detection area.

8. The data acquisition and processing system for photoelectric measurement according to claim 7, characterized in that, The photoelectric signal of the detection area is extracted for pattern recognition to complete fine target segmentation, and the spectral features are optimized, specifically as follows: A preliminary detection area is obtained, and clustering is performed in different frequency domain intervals of the preliminary detection area according to the energy density threshold. The peak point of each frequency domain interval is selected. The energy density threshold is set by extracting the distance between the energy centers of the target optimization ellipse by the spectral features corresponding to the target, and by using the peak point to calculate the distance with the nearest frequency domain signal from high frequency to low frequency, and determining whether the distance is less than the energy density threshold. If the value is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated; if the value is greater than the preset tolerance, the distance difference is obtained; if the distance difference is less than the preset tolerance, the energy distribution information of the frequency domain interval is updated. The boundary range of the detection area is updated based on the energy distribution information in the frequency domain interval. The clustering of all signals is completed through iteration. Fine target segmentation is completed based on the updated boundary range of the detection area, and the spectral features are optimized.

9. The data acquisition and processing system for photoelectric measurement according to claim 6, characterized in that, A feature extraction model is constructed based on time-frequency analysis, specifically as follows: A feature extraction model is constructed based on wavelet transform to obtain optimized spectral features. These features are then imported into the feature extraction model. Convolutional kernels are used to extract time-frequency features in different frequency domain intervals. Hidden nodes of corresponding layers are set according to the time-frequency features in different frequency domain intervals, and fusion is performed through fully connected layers. In the feature extraction model, a weight allocation layer is set up, and an attention mechanism is used to obtain the dynamic weights of time-frequency features in different frequency domain intervals, which characterize the contribution of time-frequency features. The weighted time-frequency features are compressed again through convolution and average pooling operations. The compressed time-frequency features are then dimensionality reduced and fused before being imported into a fully connected layer. The probability distribution of the target category is calculated by a normalized exponential function, the target classification result is determined based on the probability distribution, and the target classification result of the target area is combined with the corresponding spectral features to generate the photoelectric measurement data acquisition result of the target area.

Images