Photon echo power detection method and system based on cat's eye effect

Abstract

The present invention discloses a photon echo power detection method and system based on the cat's eye effect; the present invention relates to the field of laser detection technology; a three-dimensional Hough transform (3DHough Transform) is applied to the image or point cloud data in the data set D to identify the shape and position of the target, and the Hessian normal vector →n is output. Combined with the Markov matrix, the future state of the target is predicted according to its historical state to achieve fuzzy tracking; the present invention can more accurately predict the reflection distribution through the fusion of the SVM regression model and multiple features, thereby more accurately correcting the photon echo power. It is helpful to reduce errors in the ranging process and improve the ranging accuracy. And it has stronger adaptability to changes in different environmental conditions and target characteristics. For example, factors such as the roughness or shape of the target surface will affect the reflection distribution, and the present invention can comprehensively consider these factors, thereby improving the robustness of the system.

Description

Technical Field

[0001] The present invention relates to the field of laser detection technology, specifically to the field of satellite laser ranging (SLR) or precision orbit determination (POD), and in particular to a photon echo power detection method and system based on cat's eye effect. Background Art

[0002] At night, cats' eyes appear exceptionally bright. This is because when light penetrates the pupil of a cat's eye, is projected onto the fundus and is reflected by the fundus, the light beam will bend back along the original path, making the cat's eyes appear exceptionally bright at night. This phenomenon can also be compared to the working principle of some optoelectronic devices. In the field of optoelectronic devices, a photodetector is installed on the focal plane of its optical system. When the laser is incident on the detector surface, a light beam returning along the original path is generated based on the principle of reversibility of light. This return light beam has good directionality and high brightness, which produces a retroreflected laser similar to the cat's eye effect.

[0003] Specifically, the lens with focusing function and the photosensitive surface with mirror reflection characteristics work together to reversely reflect the incident laser, which has the effect of "collimation", making the intensity of the laser echo significantly higher than the reflected echo from other targets or backgrounds. The optical system of this type of optoelectronic device that uses the "cat's eye" effect is mainly composed of an objective lens and a photosensitive surface located in the focal plane. Ideally, Figure 2 As shown in the figure, this optical system is actually a system without aberration and diffraction, and parallel light beams far from the optical axis (such as AB or A1B1) will converge at point F on the focal plane. In view of the axial symmetry of the optical system, the light reflected from point F on the focal plane will return in the form of mirror reflection, strictly along the original path of the incident light (such as A1B1 or AB). This feature makes this type of optoelectronic equipment have excellent applications in distance measurement.

[0004] In actual operation, when the laser beam hits the optical window of the target, the laser will enter the field of view of the optical system due to the reversibility of the optical path and be reflected by the system. After the reflected light is collimated by the optical system, it will form a beam of high-brightness reflected light. By utilizing the cat's eye effect, accurate detection and ranging of distant targets can be achieved.

[0005] The laser ranging algorithm based on the cat's eye effect can be found in the literature cited in this background technology. [4]-[5], but in practice it usually relies on accurately measuring the round-trip time of the laser pulse to determine the distance between the target and the rangefinder. However, if the reflected signal is too weak, that is, the photon echo power is too low, then the rangefinder may not be able to accurately detect the arrival time of the reflected pulse, or the detected signal may be interfered by noise, resulting in ranging errors. Therefore, the photon echo power P needs to be calculated in the ranging algorithm. r When the power of the received reflected signal is lower than the threshold, the rangefinder will discard the measurement or adopt other strategies to improve the accuracy of the ranging, such as taking the average of multiple measurements to reduce the impact of noise.

[0006] Because when the laser is incident vertically, the distribution of reflected light is the most uniform and the energy loss is minimal, so in order to simplify calculations and analysis, it is usually assumed in practice that the laser incident direction is perpendicular to the target surface. However, the actual laser incident direction may not be completely perpendicular to the target surface. This can easily lead to errors in the photon echo power and the actual value; there is a deviation between the measurement result and the true value. Because the intensity and distribution of reflected light are affected by the incident angle, ignoring this may introduce systematic errors. Therefore, many existing technologies have adopted different methods to improve it, such as:

[0007] A prior art [1] Based on 3D Hough transform and multi-hypothesis tracking, the point information obtained before information fusion is fully utilized, and further combined with the target tracking method to effectively reduce the computational complexity. However, this scheme is highly dependent on the point information obtained before information fusion. If this information is inaccurate or incomplete, the tracking performance may be significantly affected. And it is difficult to adapt to dynamically changing environments.

[0008] Another existing technology [2] Generate a reference low-coherence light radiation beam, and guide the reference beam to the optical interference measurement sensor arrangement in a second incident direction at a preset incident angle relative to the first incident direction of the measurement beam; superimpose the measurement beam and the reference beam on the common incident area of ​​the sensor arrangement; detect the position of the interference fringe pattern between the measurement beam and the reference beam on the incident area; and determine the optical path difference between the measurement light path and the reference light path based on the position of the interference fringe pattern along the illumination axis of the incident area, and then correct the error. However, the interpretation capability of this scheme is insufficient because it requires the generation of a reference low-coherence light radiation beam and the setting of a complex interference measurement sensor arrangement, which first increases the complexity and cost of the system. And determining the optical path difference between the measurement light path and the reference light path by detecting the position of the interference fringe pattern is easily limited by the clarity and resolution of the interference fringes and the sensor's ability to interpret the interference pattern. At the same time, for rapidly changing or dynamic environments, the detection and interpretation of interference fringes may become more difficult because the system needs to quickly and accurately capture and process changes in the interference pattern.

[0009] Another existing technology [3] Based on random matrix theory, the construction of multi-extended target tracking framework and complex shape estimation problems are studied, and a new algorithm is proposed to correct it and improve the computational efficiency. However, due to the framework limitations of random matrices, this scheme focuses more on tracking shape estimation, but is not strong enough in data fusion and prediction, and cannot make full use of multiple information sources for comprehensive analysis.

[0010] To this end, the present invention proposes a photon echo power detection method and system based on the cat's eye effect.

[0011] The prior art citation information listed in this background technology is as follows:

[0012] [1] Wang Jun, Liu Dongli. Research on target tracking method before multi-radar point fusion[J]. Electroacoustic Technology, 2021(010):045.

[0013] [2] Simone Donaldo, Barbara Previtalli, Daniel Colombo. Method and system for determining and controlling the separation distance between the working head of a laser processing machine and the surface of the object being processed by low-coherence optical interferometry technology [P]. CN202080095290.6: 2022-09-13.

[0014] [3] Hu Qi. Research on extended target tracking algorithm based on random matrix[D]. Xidian University, 2019.

[0015] [4] Long Mingliang, Zhang Haifeng, Lin Haisheng, et al. Satellite laser ranging based on the “cat’s eye” effect of the optical system of the SiCH-2 remote sensing satellite [J]. Reviews in Earth and Planetary Physics, 2023, 54(5): 581-586.

[0016] [5] Ma Haozhou. Research on target reflection characteristics and "cat's eye" effect [J]. Laser Technology, 27.6 (2003): 600-602. Summary of the invention

[0017] In view of this, the embodiments of the present invention hope to provide a photon echo power detection method and system based on the cat's eye effect to solve or alleviate the technical problems existing in the prior art, namely, difficulty in adapting to dynamically changing environments, insufficient interpretation capabilities and tracking shape estimation, and at least provide a beneficial option for this; the technical solution of the present invention is implemented as follows:

[0018] First, the photon echo power detection method based on the cat's eye effect:

[0019] 1. Overview:

[0020] This solution aims to solve the above technical problems. By considering different information sources, especially the real data of reflection distribution corresponding to different divergence angles, a SVM regression model (Support Vector Machine) is established. While the SVM regression model learns the nonlinear relationship between reflection distribution and different information sources, it introduces point cloud data of three-dimensional Hough transform and fuzzy tracking of Markov matrix. The voting algorithm (Boyer-Moore VotingAlgorithm) is used to fully utilize and fuse information, reduce dependence on a single information source, and estimate the reflection distribution during ranging, thereby optimizing resource utilization and improving ranging efficiency and effect.

[0021] (II) Analysis of target reflection characteristics:

[0022] 2.1 Laser energy under cat's eye effect:

[0023] According to the method in reference (Wei Guanghui, Yang Peigen. Application of Laser Technology in Ordnance Industry [J]. Beijing Ordnance Industry Press, 1995.2-5), the laser energy E1 per unit area of ​​the laser spot at a distance R is obtained as follows:

[0024]

[0025] Among them, E t is the laser pulse energy (J), τ t is the transmittance of the transmitting optical system; θ t is the outgoing beam divergence, R is the distance (m); τ is the atmospheric transmittance of the laser; θ is the angle between the average surface normal of the irradiated part and the incident light;

[0026] Then the laser energy E2 intercepted by the target can be obtained:

[0027]

[0028] Among them, A s is the target area irradiated by the laser beam (m 2 );

[0029] The total energy E3 of the diffuse reflection target uniformly diffusely reflected in the hemispherical space in the incident direction is:

[0030]

[0031] Where ρ is the diffuse reflectance;

[0032] Based on the above three energies, the transmission power P can be obtained t ; When it is reflected back to the rangefinder, the effective receiving area is A r , the transmittance is τ rThe receiving system receives the photon echo power P of the target laser radar equation when it reaches the photosensitive surface of the photosensitive element. r for:

[0033]

[0034] 2.2 Power Algorithm of Traditional Cat's Eye Effect:

[0035] In practice, the reflection of the incident laser from a rough surface is not uniform in all directions. Therefore, the traditional approach is to qualitatively regard it as a Lambertian source, using a cosine distribution instead of a backward uniform distribution (i.e., the reflection intensity in the incident direction is considered to be at least 1 times greater than the average value), and the laser incident direction is considered to be perpendicular to the target surface (θ = 0):

[0036]

[0037] At this time, the target reflection area A s :

[0038]

[0039] Then the photon echo power P of the target lidar equation caused by diffuse reflection will be r Corrected to:

[0040]

[0041] 2.3 Improvement ideas:

[0042] In summary, 2.1 to 2.2 are traditional laser ranging algorithms for large diffuse reflective targets. Although the Lambertian source is an idealized light source, its light intensity in all directions is uniform, stable and independent of the observation angle. This assumption makes subsequent calculations and analysis easier. And because when the laser is incident vertically, the distribution of the reflected light is most uniform and the energy loss is minimal. In actual situations, the laser incident direction may not be completely perpendicular to the target surface. If the simplified model assumes that θ = 0, but in actual situations θ ≠ 0, this will lead to a deviation between the measurement result and the true value. Because the intensity and distribution of the reflected light are affected by the angle of incidence, ignoring this may introduce systematic errors.

[0043] However, this traditional technical model also provides corresponding improvement ideas: more accurately tracking blurred or uncertain targets in complex backgrounds, and establishing a more accurate physical reflection model to describe the reflection behavior of light on surfaces of different roughness.

[0044] (III) Technical solution:

[0045] To achieve the technical goal in 2.3, this solution is implemented using the following steps S1 to S5:

[0046] 3.1 Data preprocessing:

[0047] According to the historical data, the real value features are extracted, including different bands λ, different divergence angles α and different target surface roughness r, and the data set D is established:

[0048] D=[(λ1,α1,r1),(λ2,α2,r2),…,(λ n ,α n ,r n )];

[0049] Among them, (λ i , α i , r i ) is any sample in the data set D, and n is the number of samples.

[0050] Then the dataset D is divided into training set D train , validation set D val and the test set D test :

[0051] D=[D train ∪D val ∪D test ];

[0052] The data set D can also be further cleaned and organized to ensure the accuracy and consistency of the data.

[0053] 3.2 Step S1, 3D Hough transform and fuzzy tracking:

[0054] Apply 3D Hough Transform to the image or point cloud data in dataset D to identify the shape and position of the target and output the Hessian normal vector Combined with the Markov matrix, the future state of the target is predicted according to its historical state to achieve fuzzy tracking; the result of each fuzzy tracking is state X t+1 .

[0055] 3.2.1 Three-dimensional Hough transform method:

[0056] 3D Hough transform is a technique for recognizing shapes from images or point cloud data. Read the target 3D point cloud dataset P = (p1, p2, ..., p n ), where each point p i All have three-dimensional coordinates (x i ,y i ,z i). The 3D Hough transform converts these data points from the original space to the parameter space and recognizes the shape through a voting mechanism. The conversion process is selected as the Hessian normal vector form, traversing and voting on the point cloud data to form the Hessian normal vector:

[0057]

[0058] in, is the normal vector of the plane (Hessian normal vector), (n x ,n y ,n z ) is a point p in the image or point cloud data i The normal vector x-axis, y-axis, and z-axis coordinates.

[0059] 3.2.2 Fuzzy tracking:

[0060] Assume that the historical state sequence of the target is (X1, X2, ..., X t ), where each state X i Contains different position features and speed features. The Markov matrix assumes a state X t+1 Depends only on the current state X t The state transition is represented by the Markov matrix, forming the state X t+1 The probability of generating ij :

[0061] P(X t+1 =j|X t =i)=p ij ;

[0062] Among them, p ij From a state X i Transition to another state X j This matrix can be used to predict the next state based on the current state.

[0063] 3.3 Step S2, feature vector construction:

[0064] The output of the Hough transform, that is, the Hessian normal vector And the output of the Markov matrix, that is, the state X t+1 , combined with different bands λ, different divergence angles α and different target surface roughness r, a comprehensive feature vector FV is constructed:

[0065]

[0066] Due to status X t+1 Containing multiple features, the feature vector FV is further expanded as:

[0067]

[0068] Among them, (x t+1 ,y t+1 ,z t+1 ) represents the position predicted by the Markov matrix, represents the speed predicted by the Markov matrix.

[0069] 3.4 Step S3, SVM regression model:

[0070] Use the SVM regression model to construct a regression function f(x), and use the regression function f(x) to learn the relationship between the feature vector FV and the reflection distribution. It also includes appropriate kernel functions and model parameters (penalty coefficient C value, insensitive ε value, etc.). And train the SVM regression model on the training set.

[0071] 3.4.1 Step S300, regression function f(x):

[0072] Given a training set D train :D train ={(x1,y1),(x2,y2),...,(x m ,y m )};

[0073] Among them, x i is the input feature, that is, the feature vector FV; y i is the real data of the corresponding reflection distribution, and m is the training set D train The goal of the SVM regression model is to find a regression function f(x) to fit these data. The regression function f(x) is constructed as:

[0074]

[0075] Where: ω is the coefficient vector of the regression function; b is the bias term, is a function that maps input data to a high-dimensional space (implemented by a kernel function); ξ i and is a slack variable used to deal with outliers or noise in the data, C is a penalty coefficient used to control the complexity of the model and the degree of fit to the training data, and ε is a parameter of the insensitive loss function that defines the approximation accuracy of the model to the training data. T represents the transposition of the vector. When ω is dot-producted with the input feature vector (after kernel function mapping), the transposition operation ensures that the dimensions match and that the correct linear combination can be performed. m represents the number of samples in the training dataset D.

[0076] 3.4.2 Step S301, the kernel function is the radial basis kernel function K(x,x'):

[0077] K(x,x′)=exp(-γ|xx′| 2 );

[0078] Where γ is a positive hyperparameter. The negative sign indicates that as the distance between x and x' increases, the value of the kernel function decreases exponentially, reflecting that the similarity between data points decreases with increasing distance. x and x' are two vector data points in the input space. |xx′| 2 Represents the square of the Euclidean distance between vectors x and x'. exp is the exponential function of the base of the natural logarithm, e, which is used to calculate the value of the Gaussian function. -γ|xx′| 2 Used to control the shape of the Gaussian kernel.

[0079] 3.4.3 Step S302, introducing ε-insensitive loss function:

[0080] For the previous regression function f(x), the function must have the following characteristics: i The predicted value of y i There is a certain tolerance for the difference between the predicted value and the actual value, that is, when the difference between the predicted value and the actual value is within a small range, it is considered that there is no loss. Its form is as follows:

[0081] L(x,y,f(x))=max{0,|yf(x)|-ε};

[0082] Where y is the actual value, f(x) is the predicted value output of the regression function for the input x, and the parameter ε is a pre-given positive number that represents our tolerance for prediction errors. L(x,y,f(x)) is the ε-insensitive loss function.

[0083] 3.5 Comprehensive model training:

[0084] Although the above steps S2 to S3 construct a regression model, it still needs to be trained before actual deployment. The training method is as follows:

[0085] 3.5.1 The first step, initialization:

[0086] Set the initial values ​​of the parameters of the SVM regression model. Prepare training data (training set D train ), including the label of the feature vector FV and the label of the corresponding reflection distribution.

[0087] D train ={(FV1,y1),(FV2,y2),...,(FV m ,y m )};

[0088] Among them, FV i represents the feature vector of the i-th sample, yi represents the corresponding reflection distribution label, and m is the number of training samples.

[0089] 3.5.2 The second step is to train the SVM regression model:

[0090] The optimization problem is set to solve the regression function f(x), that is, the regression function f(x) is regarded as the training function:

[0091]

[0092] Then use the training set D train The SVM regression model is trained. The model penalty coefficient C and hyperparameter γ are adjusted by cross-validation to achieve the best performance (i.e., obtain their optimal values).

[0093] 3.5.3 The third step is evaluation and verification:

[0094] On the validation set D val The performance of the SVM regression model is evaluated on the y-axis. The optimal coefficient vector ω and bias term b are determined based on the evaluation results, and the slack variable ξ is automatically determined. i and The key formula in the evaluation step is the decision function F(x) of the SVM regression model, which is used to predict new data points:

[0095] F(x)=ω T φ(x)+b.

[0096] 3.5.4 Step 4: Iteration and Optimization:

[0097] The model is iterated and optimized according to the test results, including adjusting the model parameters (coefficient vector ω, bias term b, penalty coefficient C, hyperparameter γ and slack variables ξi and ). Repeat the above steps until satisfactory prediction performance is achieved or the maximum number of training iterations is reached.

[0098] 3.5.5 Step 5: Model fusion and testing:

[0099] The Hessian normal vector Predicted future state X t+1 and other relevant features such as the position predicted by the Markov matrix (x t+1 ,y t+1 ,z t+1 ) and the speed predicted by the Markov matrix ) are combined into a new feature vector FV'. Ensure that this new feature vector FV' can capture the spatial shape, dynamic behavior of the target and the future state information predicted by the Markov matrix.

[0100] 3.5.6 Step 6: Deployment and Application:

[0101] Deploy the trained comprehensive model to the actual application environment.

[0102] 3.6 Step S4, fusion output:

[0103] The prediction result of the Markov matrix (probability p ij ), shape recognition results of Hough transform (Hessian normal vector ) and the prediction results of the SVM regression model (the output of the regression function f(x)), are fused together using a voting mechanism to obtain a comprehensive prediction value of the reflection distribution

[0104] Based on the fusion function G, a comprehensive reflection distribution prediction value can be given according to these three inputs

[0105] The way to fuse the function G is to use weighted average:

[0106]

[0107] Among them, (w1,w2,w3) are weight coefficients, which can be determined by cross-validation or other optimization methods; function of (p ij )and It converts the probability and Hessian normal vector of the Markov matrix into the predicted value y of the reflection distribution i Related mapping functions. The specific forms of these functions need to be defined according to the actual data and problems.

[0108] For example, because this step uses a voting mechanism, a rule can be designed for these functions: a majority vote is performed based on the predicted results (or converted results) of each component to determine the final reflection distribution prediction value; based on this, the function function of (p ij )and will be converted into a kind of "vote" or "selection". In this case, the two functions can be designed to return some kind of "decision" or "category" label based on the input, which will then be used in the voting process. However, in order to integrate these different information sources (Markov matrix probabilities, Hessian normal vectors, outputs of SVM regression models) into a voting system, it is necessary to first define how each information source "votes" for a specific reflection distribution category:

[0109] 1) For probability p ij:The Markov matrix gives the probability of state transition. Because different states are associated with different reflection distributions, then p ij Can be used to "vote" for the most likely state (or reflection distribution):

[0110]

[0111] Among them, it means that p is selected ij The largest j state. This can be thought of as a "vote" for the next state.

[0112] 2) For the Hessian normal vector

[0113]

[0114] Category is an assignment function that assigns each normal vector to a reflectance distribution category based on the properties of the Hessian normal vector.

[0115] In the voting mechanism, each component (Markov matrix, 3D Hough transform, SVM regression model) will "vote" for a certain reflectance distribution category. The final reflectance distribution prediction value can be determined by counting the votes for each category:

[0116]

[0117] Where c represents different reflectance distribution categories, and Votes(c) represents the number of votes for category c. This expression defines that the final predicted reflectance distribution is the category with the most votes.

[0118] 3.7 Step S5, obtaining the photon echo power P r :

[0119] In the traditional Lambertian source model, the reflected intensity is assumed to be cosine distributed and independent of the observation angle. However, in reality, the distribution of reflected light is affected by factors such as surface roughness, incident angle, and material. Therefore, a correction factor is introduced This factor relies on the fitting function f to synthesize the predicted value. and the incident angle θ, thereby adjusting the distribution of the reflection intensity.

[0120]

[0121] Among them, the correction factor The form is:

[0122]

[0123] Among them, a and b are parameters that need to be fitted by experimental data and can be determined by the least square method, θ0 is the reference incident angle (for example, the angle at vertical incidence), It is the average predicted value or a reference value of the SVM regression model on the training data.

[0124] This correction factor represents the decrease in reflected intensity when the incident angle deviates from the reference angle or the predicted value deviates from the average value. The exponential function form ensures that the correction factor is always positive and decreases smoothly as the deviation increases.

[0125] (IV) Mechanisms for resolving technical issues:

[0126] 4.1 Improvements for those that are highly dependent on point information:

[0127] By introducing the SVM regression model and voting mechanism, the present invention can more effectively fuse multiple information, reduce the dependence on complex three-dimensional Hough transform, and thus reduce computational complexity. Moreover, by combining the prediction of the SVM regression model, the state prediction of the Markov matrix, and the shape recognition of the Hough transform, our method can more comprehensively utilize and fuse information, reduce the dependence on a single information source, and improve robustness. At the same time, through the SVM regression model and voting mechanism, it can better adapt to dynamically changing environments and target characteristics. The SVM regression model can be retrained based on new data to adapt to different scenarios and targets. Moreover, through the prediction of the SVM regression model, our method can estimate the reflection distribution before ranging, thereby optimizing the utilization of resources, such as parameters such as the power and angle of laser emission, and improving ranging efficiency.

[0128] 4.2 Improvement of insufficient interpretation ability:

[0129] The present invention is based on the SVM regression model and multiple feature fusion, and does not require a complex interference measurement system and reference beam generation, thereby simplifying the system complexity and reducing costs. At the same time, by using a comprehensive machine learning model composed of an SVM regression model, Markov matrix prediction, and Hough transform, it is itself based on the sensitivity of environmental factors for targeted training, so it can provide accurate predictions and corrections more stably. Furthermore, the present invention does not rely on the measurement of interference fringes, but predicts and corrects the photon echo power through a machine learning model and feature fusion, thereby avoiding the limitations of optical path difference measurement in traditional technologies.

[0130] 4.3 No longer limited to tracking shape estimation:

[0131] The technical solution of the present invention integrates multiple features (such as the predicted state change of the Markov matrix, the Hessian normal vector of the Hough transform, different bands, divergence angles and target surface roughness, etc.) through the SVM regression model, so that the model can understand the target state more comprehensively and improve the processing ability of complex shapes and dynamic targets. In addition, by combining the prediction results of the Markov matrix, the shape recognition results of the Hough transform and the prediction results of the SVM regression model, a comprehensive reflection distribution prediction value can be provided, which is used as a correction factor to correct the photon echo power, thereby enhancing the accuracy and robustness of the prediction.

[0132] Second, the photon echo power detection system based on the cat's eye effect:

[0133] like Figure 5 As shown, the system includes a processor, a wireless receiving device connected to the processor, a wireless transmitting device connected to the optical transmitting system and the optical receiving system, and a register connected to the processor, wherein program instructions are stored in the register;

[0134] The optical transmitting system and the optical receiving system realize data transmission with the processor based on the wireless connection between the wireless transmitting device and the wireless receiving device;

[0135] When the program instructions are executed by the processor, the processor is caused to execute the photon echo power detection method as described above.

[0136] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0137] 1. Improve the ranging accuracy: The present invention can more accurately predict the reflection distribution through the fusion of the SVM regression model and multiple features, thereby more accurately correcting the photon echo power. It helps to reduce the error in the ranging process and improve the ranging accuracy. It also has stronger adaptability to different environmental conditions and changes in target characteristics. For example, factors such as the roughness or shape of the target surface will affect the reflection distribution, and the present invention can comprehensively consider these factors, thereby improving the robustness of the system.

[0138] 2. Improve anti-interference ability: The present invention fuses multiple prediction results through a voting mechanism, which can reduce the impact of individual prediction errors on the final result. Therefore, the fusion method of the present invention can suppress the interference of noise and outliers to a certain extent, making the corrected photon echo power more reliable.

[0139] 3. Optimizing resource utilization: The present invention can estimate the reflection distribution before ranging through the prediction of the SVM regression model, thereby optimizing the parameters such as the power and angle of the laser emission. It helps to save energy and improve the efficiency of ranging. At the same time, since multiple features and prediction methods are taken into account, it can perform ranging in more complex and changeable environments. For example, when the target surface is unevenly rough, irregularly shaped, or the material is changeable, this method can still maintain a high ranging accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0140] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0141] Figure 2 This is a schematic diagram of the optical system of an optoelectronic device that utilizes the "cat's eye" effect under traditional ideal conditions;

[0142] Figure 3 Schematic diagram of the training process of the SVM regression model of the present invention;

[0143] Figure 4 is a flowchart of the voting mechanism of the present invention;

[0144] Figure 5 It is a schematic diagram of the system composition of the present invention;

[0145] Figure 6 It is a training recall curve diagram of the SVM regression model in a specific implementation mode;

[0146] Figure 7 It is a schematic diagram of the distribution of prediction results of the voting fusion method in a specific implementation manner. DETAILED DESCRIPTION

[0147] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0148] Example 1: Please refer to Figure 1 This embodiment discloses a photon echo power detection method based on the cat's eye effect; its core is to use a support vector machine (SVM) regression model to learn and predict the complex relationship between reflection distribution and different information sources; steps S1 to S5 in this embodiment need to be executed to achieve the above requirements.

[0149] In this embodiment, firstly, the real features are extracted according to the historical data, including different bands λ, different divergence angles α and different target surface roughness r, and the initial data set D is established:

[0150] D=[(λ1,α1,r1),(λ2,α2,r2),…,(λ n ,α n ,r n )];

[0151] Among them, (λ i , α i , r i ) is any sample in the dataset D, representing a data point under a specific band, divergence angle, and surface roughness conditions. n is the number of samples. At the same time, in order to train and optimize the SVM regression model, the dataset D is divided into three parts: training set D train , validation set D val and the test set D test , which are used to train the model, adjust the model parameters and evaluate the final performance of the model respectively:

[0152] D=[D train ∪D val ∪D test ] ；

[0153] Furthermore, the C language program for establishing the above dataset D and cleaning (normalizing) it is as follows:

[0154]

[0155]

[0156]

[0157] In the above program, a Sample structure is defined to represent each sample, which contains three features: band (lambda), divergence angle (alpha), and surface roughness (r). In the main function, an array raw_data containing the original data is created. The minimum and maximum values ​​of each feature are found by traversing the original data set. Then the normalize function is used to normalize each feature value to the range [0,1]. This function achieves normalization by subtracting the minimum value and dividing by the range (maximum value minus minimum value). Finally, the normalized data set is output.

[0158] In this embodiment, regarding step S1, 3D Hough transform and fuzzy tracking: applying 3D Hough transform to the image or point cloud data in the data set D to identify the shape and position of the target and output the Hessian normal vector Combined with the Markov matrix, the future state of the target is predicted according to its historical state to achieve fuzzy tracking; the result of each fuzzy tracking is state X t+1 .

[0159] Specifically, the 3D Hough transform method: In this method, we first read the target 3D point cloud dataset P = (p1, p2, ..., p n ), where each point p i All have three-dimensional coordinates (x i ,y i ,z i ). The core idea of ​​3D Hough transform is to transform data points from the original 3D space to parameter space. In this process, each data point "votes" for possible shapes in parameter space. During the transformation, the form of Hessian normal vector is chosen to represent the identified shape:

[0160]

[0161] in, It is the normal vector of the plane (Hessian normal vector), which describes the normal vector of a point p in the image or point cloud data. i The normal direction, (n x ,n y ,n z ) is a point p in the image or point cloud data i The normal vector x-axis, y-axis and z-axis coordinates. The three-dimensional Hough transform method can use the algorithm encapsulated by the OpenCV library to directly calculate the Hessian normal vector we need

[0162] Specifically, the fuzzy tracking of Markov matrix: Let the historical state sequence of the target be (X1, X2, ..., X t ), each state X i contains position features and velocity features. These states form the basis of the Markov chain. The Markov matrix assumes that the next state X t+1 Depends only on the current state X t . And from state X i Transfer to state X j The probability is expressed as p ij . Through this matrix, we can predict the probability distribution of the next state based on the current state:

[0163] P(X t+1 =j|X t =i) = p ij ;

[0164] Combining the historical state sequence and the Markov matrix, the fuzzy tracking of the target is realized. The result of each tracking is a new state X t+1 , which reflects the possible position and speed of the target at the future moment. Therefore, by utilizing the historical state information, the accuracy and stability of tracking are improved. Moreover, the introduction of the Markov matrix simplifies the complexity of state transfer, making real-time tracking possible. At the same time, the fuzzy tracking method can adapt to the uncertainty of target motion and improve the robustness of the system.

[0165] In summary, the three-dimensional Hough transform and fuzzy tracking technology in step S1 provide powerful shape recognition and tracking capabilities for the photon echo power detection method based on the cat's eye effect. These technologies not only improve the accuracy of detection, but also enhance the adaptability and robustness of the system in various complex environments.

[0166] In this embodiment, regarding step S2, feature vector construction: the output result of the Hough transform, that is, the Hessian normal vector (provides important information about the shape and orientation of the target surface, which helps us understand the interaction between the laser beam and the target surface), and the output of the Markov matrix, i.e., the state X t+1 (Understanding the dynamic behavior of the target), combined with different bands λ, different divergence angles α and different target surface roughness r, a comprehensive feature vector FV is constructed:

[0167]

[0168] Lasers of different wavelengths will have different propagation characteristics in the atmosphere and interactions with different materials; the divergence angle affects the concentration and irradiation range of the laser beam; and the surface roughness of the target directly affects the reflection and scattering of the laser beam.

[0169] And because of the state X t+1 Containing multiple features, the feature vector FV is further expanded as:

[0170]

[0171] Among them, (x t+1 ,y t+1 ,z t+1 ) represents the position predicted by the Markov matrix, represents the speed predicted by the Markov matrix.

[0172] Then, based on the corresponding band λ, divergence angle α and target surface roughness r, the corresponding feature vector FV is filled into the original data set D. The feature vector FV contains all-round information from Hessian normal vector, Markov state prediction to laser parameters and target surface characteristics. This feature vector will serve as an important input for subsequent photon echo power detection, which will help improve the accuracy and reliability of detection.

[0173] In this embodiment, regarding step S3, SVM regression model: a regression function f(x) is constructed using the SVM regression model, and the relationship between the feature vector FV and the reflectance distribution is learned through the regression function f(x). It also includes a suitable kernel function and model parameters. And the SVM regression model is trained on the training set. It includes the following steps S300-S302.

[0174] Specifically, step S300 constructs a regression function f(x): The purpose of this function is to predict the corresponding reflection distribution based on the given input features (i.e., the feature vector FV). To achieve this goal, we use the support vector machine (SVM) method and determine the parameters of the model by optimizing a specific objective function. First, a training set D is given. train , which contains m sample pairs (x i ,y i ). Each sample pair consists of an input feature x i (i.e., the eigenvector FV) and a corresponding reflection distribution y i composition:

[0175] D train ={(x1,y1),(x2,y2),...,(x m ,y m )};

[0176] The goal of the SVM regression model is to find a regression function f(x) to fit the data, which consists of two parts: one is a measure of model complexity (i.e., half of the L2 norm of ω), and the other is a penalty for the error in fitting the training data (through the slack variable ξ i and Our goal is to find a set of model parameters (coefficient vector ω of the regression function, bias term b, slack variable ξ i and ), so that the objective function achieves the minimum value:

[0177]

[0178] Among them: ε is the parameter of the insensitive loss function, which defines the approximation accuracy of the model to the training data. T represents the transpose of the vector. When ω is dot-producted with the input feature vector (after kernel function mapping), the transpose operation ensures that the dimensions match and the correct linear combination can be performed. In this formula, it refers to the kernel function, which is used to map the input data to a high-dimensional space in order to better capture the nonlinear relationship in the data. This mapping is achieved in an implicit way (i.e., the concept of "kernel trick" in the field of SVM), and does not require explicit calculation of the vector representation in the high-dimensional space.

[0179] By solving this optimization problem, we can obtain a set of optimal model parameters (ω, b, ξi and ξi*), thereby constructing the final regression function f(x). This function can predict the corresponding reflection distribution based on the new input features, thereby achieving accurate detection of the photon echo power.

[0180] Specifically, in step S301, the above kernel function is preferably a radial basis kernel function K(x, x'): it has a characteristic that it can map data points to an infinite-dimensional feature space, which enables SVM to handle very complex nonlinear relationships:

[0181] K(x,x′)=exp(-γ|xx′| 2 );

[0182] Among them, γ is a positive hyperparameter that controls the similarity measure of the data and how the similarity of data points changes with the distance between them. The more sensitive the kernel function is to the distance between data points, the faster the similarity decreases. The negative sign indicates that as the distance between x and x' increases, the value of the kernel function decreases exponentially, reflecting that the similarity between data points decreases with the increase of distance. x and x' are two vector data points in the input space, which are the original feature vectors or the feature vectors after some transformation, and are also the feature vectors corresponding to the training set D train Samples in |xx′| 2 Represents the square of the Euclidean distance between vectors x and x'. exp is the exponential function of the base of the natural logarithm, e, which is used to calculate the value of the Gaussian function. -γ|xx′| 2 Used to control the shape of the Gaussian kernel. When two data points are very close (i.e. |xx′| 2 As the distance increases, the value of the kernel function drops rapidly to close to 0, indicating that the similarity decreases.

[0183] It can be understood that the radial basis kernel function selected in step S301 provides the SVM regression model with the ability to handle complex nonlinear relationships, and controls the similarity measure between data points by adjusting the value of the hyperparameter γ, thereby optimizing the prediction performance of the model. This is crucial to improving the accuracy and stability of laser detection, laser ranging, and photon echo power detection.

[0184] Specifically, in step S302, an ε-insensitive loss function is introduced: the function needs to have the following characteristics: i The predicted value of y i There is a certain tolerance for the difference between the predicted value and the actual value, that is, when the difference between the predicted value and the actual value is within a small range, it is considered that there is no loss. Its form is as follows:

[0185] L(x,y,f(x))=max{0,|yf(x)|-ε};

[0186] Among them, y is the actual value, that is, the true target value we want to predict, f(x) is the predicted value output of the regression function for the input x, and the parameter ε in this formula is a pre-given positive number, representing our tolerance for prediction errors. L(x,y,f(x)) is the ε-insensitive loss function.

[0187] When the difference between the predicted value and the actual value is in the range [-ε,ε], the loss function value is 0. This is equivalent to the fact that points within this range will not affect the training of the model, making the model more sparse, that is, only some data points have a practical impact on the model. Since the ε-insensitive loss function is tolerant to small prediction errors, it is relatively insensitive to noise and outliers in the data, which makes the model more robust.

[0188] In addition, since the ε-insensitive loss function is tolerant to small prediction errors, it is relatively insensitive to noise and outliers in the data. This makes the model more robust when dealing with data sets containing noise or outliers, and better able to capture the main trends and features of the data.

[0189] It is understandable that in applications such as laser detection, laser ranging, and photon echo power detection, the SVM regression model with the ε-insensitive loss function can more effectively handle complex nonlinear relationships and noise problems in the data. By adjusting the tolerance parameter ε, we can find a balance between the complexity of the model and the degree of fit to the data, thereby achieving more accurate predictions and more stable performance. This is of great significance for improving the accuracy and reliability of laser detection systems.

[0190] Furthermore, the C++ program framework of the SVM regression model constructed in the above steps S300 to S302 is as follows:

[0191]

[0192]

[0193]

[0194]

[0195] The above C++ program defines a SVMModel class, which is used to encapsulate the relevant data and operations of the support vector machine (SVM) model. The member variable class defines the parameters of the SVM regression model, such as gamma (parameter for the RBF kernel function) and epsilon (parameter for the ε-insensitive loss function). These parameters are crucial to the performance and generalization ability of the SVM regression model. The constructor is used to initialize the SVM regression model instance and set the initial gamma and epsilon parameter values. The train function is a key part of SVM regression model training. The predict function is used to classify or regress new data points based on the trained SVM regression model. This function uses the trained support vector and decision function F(x) to calculate the output category or value of the input data.

[0196] In this example, see Figure 1 and Figure 4 , about step S4, fusion output: after the SVM regression model is trained, the prediction result of the Markov matrix (probability p ij ), shape recognition results of Hough transform (Hessian normal vector ) and the prediction results of the SVM regression model (the output of the regression function f(x)), are fused together using a voting mechanism to obtain a comprehensive prediction value of the reflection distribution

[0197] Based on the fusion function G, a comprehensive reflection distribution prediction value can be given according to these three inputs

[0198] The way to fuse the function G is to use weighted average:

[0199]

[0200] Among them, (w1, w2, w3) are weight coefficients, which can be determined by subjective assignment based on conventional knowledge in this field, cross-validation or other optimization methods (such as genetic algorithm, iterative method); if there is no special requirement, the average of the three weight coefficients can be assigned; function of (p ij )and It converts the probability and Hessian normal vector of the Markov matrix into the predicted value y of the reflection distribution i Related mapping functions. The specific forms of these functions need to be defined according to the actual data and problems.

[0201] It is understandable that by fusing information from multiple sources (the probability prediction of the Markov matrix, the shape recognition results of the Hough transform, and the prediction results of the SVM regression model), we can obtain a more comprehensive and accurate reflection distribution prediction value. This fusion method takes advantage of the complementarity between different methods, thereby improving the accuracy of the overall prediction. And because of the fusion of multiple information sources, this method is more robust to noise or errors in a single information source. Even if there is a problem or a large error in one information source, other information sources can still provide valuable information to correct or compensate for this error.

[0202] Preferably, because this step adopts a voting mechanism, the mapping function function of (p ij )and Design a rule: perform majority voting based on the prediction results (or converted results) of each component to determine the final reflection distribution prediction value; based on this, function of (p ij )and will be converted into a kind of "vote" or "selection". In this case, the two functions can be designed to return some kind of "decision" or "category" label based on the input, which will then be used in the voting process. However, in order to integrate these different information sources (Markov matrix probabilities, Hessian normal vectors, outputs of SVM regression models) into a voting system, it is necessary to first define how each information source "votes" for a specific reflection distribution category:

[0203] 1) function of (p ij ) implementation method:

[0204] The Markov matrix gives the probability of state transition, that is, the possibility of transitioning from one state to another. Different states are associated with different reflection distributions. Therefore, we can vote for the most likely state (i.e., reflection distribution) through pij:

[0205]

[0206] This expression states that for a given row in the state transition probability matrix (i.e., the transition probability from state i to all possible states j), we choose the one that makes p ijThe largest state j. This can be viewed as the system's vote on the most likely next state. In this way, we convert the state transition probability into a specific state choice, which is in turn associated with a specific reflection distribution.

[0207] 2) For the Hessian normal vector provides important information about the object's shape and orientation. In our approach, this normal vector is used to vote for the reflectance distribution class whose characteristics best match:

[0208]

[0209] Category is an assignment function that assigns each normal vector to a reflectance distribution category based on the properties of the Hessian normal vector. In this way, we assign the Hessian normal vector The information is converted into a specific reflection distribution category selection.

[0210] Preferably, the form of the allocation function Category is:

[0211]

[0212] in: is the representative vector or center point of the kth reflection distribution category, which can be determined by clustering the Hessian normal vectors in the training data. Is a Euclidean distance function used to measure the Hessian normal vector With Category Center This formula indicates that we assign the Hessian normal vector to the reflection distribution category represented by the category center that is closest to it.

[0213] Furthermore, in the voting mechanism of step S4, each component (Markov matrix, 3D Hough transform, SVM regression model) will "vote" for a certain reflection distribution category. The "voting" here is actually a metaphor, which is equivalent to each component contributing its own strength to the final decision. The Markov matrix votes according to the state transition probability, the 3D Hough transform votes by detecting the shape and spatial structure in the image, and the SVM regression model votes based on its trained pattern recognition ability. The final reflection distribution prediction value can be determined by counting the number of votes for each category:

[0214]

[0215] Among them, c represents different reflection distribution categories, and Votes(c) represents the number of votes obtained by category c. This expression defines that the final predicted reflection distribution is the category with the most votes. argmax(Votes(c)) means selecting the category c that makes Votes(c) the largest, that is, the category with the most votes. Therefore, is the final predicted reflection distribution category.

[0216] 1) For the Hessian normal vector (Votes(c)) counting function:

[0217] Category function is used to assign the Hessian normal vector to a specific reflectance distribution category, so we can count directly based on this classification result.

[0218]

[0219] Where I is the indicator function. 1 when , otherwise 0. This formula counts the number of all Hessian normal vectors classified into category c.

[0220] 2) For probability p ij Votes(c) counting function:

[0221]

[0222] Among them, I is also an indicator function. When a state j is the largest and the probability p ij ==c is 1, otherwise it is 0. This formula counts the number of all state transition probabilities that are mapped to category c.

[0223] 3) For the reflection distribution prediction value y output by the SVM regression model i Votes(c) counting function: It can be mapped to categories based on quantitative methods (such as binning, threshold judgment):

[0224]

[0225] Among them, I is also an indicator function. When the predicted value y i When it is greater than the ideal threshold δ and == c, it is 1, otherwise it is 0. This formula counts the number of all SVM regression model outputs that are mapped to category c.

[0226] Ultimately, the total votes for each category will be the sum of the votes of these three components, TotalVotes(c):

[0227]

[0228] It can be understood that by utilizing the probability of the Markov matrix and the information of the Hessian normal vector, combined with a voting mechanism, we can comprehensively consider multiple information sources to make more accurate reflection distribution predictions. This comprehensive approach reduces the errors and uncertainties that may be caused by a single information source. Moreover, when a certain information source (such as the output of the SVM regression model) is interfered by noise or has outliers, other information sources (such as the probability of the Markov matrix and the Hessian normal vector) can still provide valuable information to help the system make correct predictions. This method of multi-information source fusion enhances the robustness and stability of the system.

[0229] Furthermore, the C++ execution program of the above step S4 is as follows:

[0230]

[0231]

[0232]

[0233]

[0234] In the above program, three maps are first initialized to store the voting results from the Hessian normal vector, Markov matrix, and SVM regression model. These maps map the reflection distribution categories (ReflectionCategory) to the number of votes they each receive. The program then iterates over the voting results of each component and merges them into a total voting result map. This is achieved by simply adding the number of votes each category receives from each component. The std::max_element function is then used to find the reflection distribution category with the most total votes. This function accepts a comparison function that determines which element (in this case, the pair of category and vote) is the "largest", that is, the one with the most votes. Finally, the program outputs the predicted reflection distribution category, the visualization of which is shown below. Figure 7 As shown. At the same time, for convenience, the above program converts the ReflectionCategory enumeration into an integer for output.

[0235] In this embodiment, regarding step S5, the photon echo power P is obtained. r :In the traditional Lambertian source model, the reflected intensity is assumed to be cosine distributed and independent of the observation angle. However, in reality, the distribution of reflected light is affected by factors such as surface roughness, incident angle, and material. Therefore, a correction factor is introduced This factor relies on the fitting function f to synthesize the predicted value. and the incident angle θ, thereby adjusting the distribution of the reflected intensity:

[0236]

[0237] in, is the correction factor, f is the fitting function; τ is the laser atmospheric transmittance; R is the distance value; ρ is the diffuse reflection coefficient; A r is the effective receiving area; τ r is the transmittance of the receiving system; τ t is the transmittance of the emitting optical system; P t is the transmit power.

[0238] Preferably, the correction factor The form is:

[0239]

[0240] Among them, a and b are parameters that need to be fitted by experimental data and can be determined by the least square method, θ0 is the reference incident angle (for example, the angle at vertical incidence), is the average prediction value or a reference value of the SVM regression model on the training data. Reference incident angle θ0 and reference prediction value It provides a reference point for the correction factor, making the correction process more targeted and explainable.

[0241] This correction factor represents the decrease in reflected intensity when the incident angle deviates from the reference angle or the predicted value deviates from the average value. The exponential function form ensures that the correction factor is always positive and decreases smoothly as the deviation increases.

[0242] In summary, by introducing the correction factor This step can simulate and calculate the photon echo power more accurately, thereby improving the performance of the laser detection and ranging system. This optimization method not only takes into account a variety of practical factors that affect the reflection intensity, but also ensures the accuracy and versatility of the model through flexible parameter adjustment and experimental data fitting.

[0243] Embodiment 2: In step S1 of embodiment 1, the Markov matrix sets a state X t+1 Depends only on the current state X t The state transition is represented by the Markov matrix, forming the state X t+1 The probability of generating ij The idea of ​​the first embodiment is that the Markov matrix can be used to predict the next state according to the current state. On this basis, this embodiment will further disclose the specific implementation of the Markov matrix.

[0244] In the Markov matrix of this embodiment, "state" refers to the specific condition or configuration of the system at a certain moment. In the concept of target tracking in embodiment 1, a "state" includes the dynamic characteristics of the position, velocity and acceleration of the target. For example, if a point on a two-dimensional plane is used to represent the position of the target, then the state can be a vector containing x and y coordinates. "Probability" refers to the possibility of transitioning from one state to another. In the Markov matrix, this transition probability is fixed and depends only on the current state, not on the past state or the future state.

[0245] In the Markov matrix of this embodiment, its form is a square matrix whose size is equal to the number of all possible states. Each element p in the matrix ij (i.e., probability) represents the probability of transitioning from state i to state j. The sum of each row of this matrix must be 1, because the sum of the probabilities of transitioning from a state to all possible states must be 1.

[0246] Specifically, for n possible states, the Markov matrix P of this embodiment is expressed as:

[0247]

[0248] Among them, p ij (element in row i and column j) represents the transition from state X i Transfer to state X j probability.

[0249] It is understandable that in practical applications, this matrix needs to be constructed through statistical or empirical data. For example, in target tracking applications, there will be a large amount of historical data showing the frequency of the target moving from one position (state) to another. From this data, the transition probability can be calculated and filled into the above Markov matrix.

[0250] Exemplarily, the above filling method is:

[0251] P1. Data collection: It can come from observations from various sensors (such as radar, camera, GPS, etc.), or target trajectories obtained through other tracking algorithms.

[0252] P2. State definition: Define the state space, which is the set of all possible locations where the target can appear. These locations can be discrete grid points or a set of representative points in a continuous space. For example, on a two-dimensional plane, the space is divided into a grid, and each grid point represents a state.

[0253] P3. Count the frequency of state transitions: traverse the historical data and count the frequency of the target transitioning from one state to another. Specifically, for each pair of observation data (or trajectory points) at adjacent time points, determine the state of the target at the previous time point and the state at the next time point, and then increase the transition count between the two states. For example, use a two-dimensional array to record the number of transitions between states, where the rows of the array represent the starting state and the columns represent the ending state. Each time a state transition is observed, add one to the corresponding array element.

[0254] P4. Calculate the state transition probability:

[0255] Among them, n ij is the number of observations from state i to state j, and ∑ k n ik is the total number of transitions from state i to all possible states. This formula ensures that the sum of each row of the Markov matrix is ​​equal to 1, that is, the sum of the probabilities of transitioning from a state to all possible states is 1.

[0256] P5. Fill in the Markov matrix: Fill the calculated state transition probability into the Markov matrix. This will give you a complete Markov matrix, where each element represents the probability of transitioning from one state to another.

[0257] Furthermore, the fuzzy tracking program (C language) of the above Markov matrix is ​​as follows:

[0258]

[0259]

[0260] In the above program, NUM_STATES defines the number of states in the Markov matrix. markov_matrix is ​​a two-dimensional array that stores the pre-filled state transition probabilities. The predict_next_state function accepts the current state as a parameter and prints out the probability distribution of transitioning from the current state to all possible states.

[0261] Example 3: In step S3 of Example 1, an SVM regression model is constructed, but the SVM regression model needs to be trained before it can be used in step S4. Figure 3 , this embodiment will disclose the training method of the SVM regression model:

[0262] The first step is initialization: set the initial values ​​of the parameters of the SVM regression model. Prepare training data (training set D train ), including the label of the feature vector FV and the label of the corresponding reflection distribution.

[0263] D train ={(FV1,y1),(FV2,y2),...,(FV m ,y m )};

[0264] Among them, FV i represents the feature vector of the i-th sample, y i represents the corresponding reflection distribution label, and m is the number of training samples.

[0265] The second step is to train the SVM regression model: the optimization problem is set to solve the regression function f(x), that is, the regression function f(x) is regarded as a training function, so that it can accurately predict the reflection distribution under a given feature vector. This process includes using the training set D train The model is trained and the penalty coefficient C and hyperparameter γ of the model are adjusted by cross-validation to achieve the best performance:

[0266]

[0267] Then, the model penalty coefficient C and hyperparameter γ are adjusted through cross-validation to achieve the best performance (i.e., obtain their optimal values). The operation steps of cross-validation technology include:

[0268] P1. Use k-fold cross validation to divide the data set into k subsets (k values ​​are 5, 10, etc.). Each time, select one subset as the validation set, and the remaining k-1 subsets are combined as the training set.

[0269] P2. Set parameter range: Set a reasonable range for C and γ. For example, C can take a range of values, such as [0.1, 1, 10, 100], and γ can take a similar range.

[0270] P3, training and validation: For each pair of C and γ, use the training set to train the SVM regression model. Then evaluate the performance of the model on the validation set and record the evaluation results. The evaluation indicators can be accuracy, F1 score, AUC, etc. In this embodiment, the loss rate and recall rate are selected; for each pair of C and γ values, calculate the performance indicators on the validation set.

[0271] P4. Select the best parameters: Compare the performance of all parameter combinations on the validation set. Select the pair of C and γ with the best performance as the best parameters.

[0272] In P1 to P4 above, the data set is divided multiple times and the model is trained to evaluate the model performance under different parameter combinations, thereby selecting the optimal parameter settings. This method helps prevent the model from overfitting and improves its generalization ability on unknown data.

[0273] The third step is to evaluate the validation: val The performance of the SVM regression model is evaluated on the y-axis. The optimal coefficient vector ω and bias term b are determined based on the evaluation results, and the slack variable ξ is automatically determined. i and The key formula in the evaluation step is the decision function F(x) of the SVM regression model, which is used to predict new data points:

[0274] F(x)=ω T φ(x)+b;

[0275] In the validation phase, the above decision function F(x) is used to predict each sample in the validation set and compare it with the true label to evaluate the performance of the model. The mean square error (MSE) is selected as the evaluation indicator:

[0276]

[0277] Among them, y i is the true value of the i-th sample in the validation set, is the model's prediction for the sample, and n is the number of samples in the validation set. The assignment strategy is to compare the MSE values ​​under different parameter settings and select the coefficient vector ω and bias term b that minimize the MSE as the optimal parameters. By evaluating the validation set, we can ensure that the model can maintain good performance on unseen data.

[0278] Step 4, iteration and optimization: Iterate and optimize the model according to the test results, including adjusting the model parameters (coefficient vector ω, bias term b, penalty coefficient C, hyperparameter γ and slack variables ξi and ), and improve the fusion strategy (recommended when multiple sub-models or feature fusion strategies are used in the model). Repeat the above steps until satisfactory prediction performance is achieved or the maximum number of training iterations is reached. When faced with complex and changing laser detection, ranging or photon echo power detection scenarios, the optimized model can provide more reliable prediction results.

[0279] Furthermore, the C++ program (based on the LibSVM library) for the above training steps is as follows:

[0280]

[0281]

[0282]

[0283]

[0284]

[0285] In the above program, the SVM regression model is obtained by reading the SVMModel class in Example 1. In order to complete the training process, the training data set (std::vector <sample>) and fill it into trainingData. Each Sample structure contains a feature vector and a corresponding label, which is used to supervise the input and expected output of the learning process. Then the program uses the LibSVM's svm_train function to train the model. This function trains an SVM regression model based on the set parameters and the provided training data, and calculates its loss rate and recall rate (the number of iterations in this embodiment is 780, and the effect of its loss rate and recall rate is as follows Figure 6 As shown). And finally automatically form the optimal coefficient vector ω, bias term b, penalty coefficient C, hyperparameter γ and two slack variables ξi and

[0286] In this embodiment, the above four steps train the SVM regression model, but step S4 needs to use the regression function f(x) to fuse with the Markov matrix and Hough transform. Therefore, on the basis of the trained SVM regression model and before executing step S4, it is also necessary to fuse the trained SVM regression model with the Markov matrix and Hough transform. test The performance of the comprehensive model is tested on the . The operation steps include:

[0287] P1. Data preprocessing: According to the method in step S1, the test set D test For each data point in , we first apply a 3D Hough transform to identify the shape and position of the target and output the Hessian normal vector Based on the historical state sequence, use the Markov matrix to predict the future state X of the target t+1 .

[0288] P2, feature fusion: Hessian normal vector Predicted future state X t+1 and other relevant features such as the position predicted by the Markov matrix (x t+1 ,y t+1 ,z t+1 ) and the speed predicted by the Markov matrix ) are combined into a new feature vector FV'. Ensure that this new feature vector FV' can capture the spatial shape, dynamic behavior of the target and the future state information predicted by the Markov matrix.

[0289] P3, SVM regression model prediction: Input the fused feature vector obtained in the previous step into the regression function f(x) in the trained SVM regression model. The SVM regression model will output a predicted value y i , this value is a measure of the reflectance distribution.

[0290] P4. Comprehensive evaluation: Compare the predicted value of the SVM regression model with the test set D test The actual reflectance distribution labels in the image are compared using an appropriate correlation coefficient R 2 To quantify the accuracy of the model prediction. 2 The closer the value is to 1, the higher the correlation between the model's predicted value and the true value, that is, the better the model's prediction performance. This quantitative evaluation method helps to objectively evaluate the performance of the SVM regression model in laser detection and ranging tasks. This evaluation method not only helps to understand the performance of the model in the current task, but also provides a clear direction for further model improvement. Of course, if R 2 If the value is not ideal, the above SVM regression model can be retrained and the coefficient vector ω, bias term b, penalty coefficient C, hyperparameter γ and slack variables ξi and Until R 2 The value meets the ideal state.

[0291] Embodiment 4: Based on Embodiment 1, this embodiment further provides an exemplary operation method of the least squares method in step S5:

[0292] The following experimental data is provided for demonstration purposes, and corresponding replacements can be made in actual applications:

[0293]

[0294] Least squares fitting: Our goal is to find parameters a and b such that the correction factor The sum of squared errors between the predicted power and the experimentally measured power is minimized. The error sum of squares S is expressed as:

[0295]

[0296] Among them, θ0 and is a known reference value. At the same time, in order to simplify the calculation, the constant term in the above formula will be combined into a new constant C; so the error square sum is simplified to:

[0297]

[0298] Find the values ​​of a and b that minimize S(a,b) by solving the following system of equations:

[0299]

[0300] To solve the above system of equations, we need to use MATLAB mathematical tools to perform the gradient descent method. The procedure is as follows:

[0301] % Experimental data

[0302] theta = [30, 40, 50]; % incident angle

[0303] y_hat = [0.6, 0.7, 0.8]; % predicted value

[0304] Pr_measured = [0.8, 0.7, 0.5]; % measured power

[0305] theta0=0; % reference incident angle (angle at vertical incidence)

[0306] y_hat0 = mean(y_hat); % Average prediction value of the SVM regression model on the training data

[0307] %Objective function definition

[0308] objectiveFunction=@(params)sum((Pr_measured-(exp(-params(1)*((theta-theta0).^2))-params(2)*((y_hat-y_hat0).^2))).^2);

[0309] % Initial parameter guess

[0310] initial_guess = [1; 1]; % Initial guess value for [a; b]

[0311] % Use MATLAB's optimization function fminunc to minimize

[0312] options = optimoptions('fminunc', 'Algorithm', 'quasi-newton'); %Use quasi-newton algorithm

[0313] [params_optimized,fval]=fminunc(objectiveFunction,initial_guess,options);

[0314] % Output optimized a and b parameter values

[0315] a_optimized=params_optimized(1);

[0316] b_optimized=params_optimized(2);

[0317] fprintf('Optimized values:a=%f,b=%f\n',a_optimized,b_optimized);

[0318] The above program uses the function fminunc in MATLAB for unconstrained nonlinear minimization. fminunc uses the quasi-newton gradient descent algorithm internally. Finally, by solving the above equations, the values ​​of a and b after fitting are output. Through this value, the correction factor in step S5 of embodiment 1 can be The photon echo power P is finally obtained. r .

[0319] Example 5: Please refer to Figure 5 Based on all the above embodiments, this embodiment further discloses a photon echo power detection system based on the cat's eye effect, the core of which is to use the cat's eye effect to perform accurate photon echo power detection. The system uses wireless technology to achieve efficient data transmission, and combines a powerful processor and customized program instructions to provide users with fast and accurate photon echo power measurement services:

[0320] (I) System composition:

[0321] (1) Processor: As the brain of the system, the processor is responsible for executing program instructions stored in registers, processing and analyzing data transmitted from the optical receiving system, and controlling the operation of the optical transmitting system. Its powerful computing power and efficient data processing capabilities are the key to system performance.

[0322] (2) Wireless receiving device and wireless transmitting device: These two parts constitute the bridge of system data transmission. The wireless transmitting device connects the optical transmitting system and the optical receiving system, and transmits the collected photon echo signal to the wireless receiving device wirelessly. This wireless connection method not only improves the flexibility and convenience of data transmission, but also reduces the complexity of wiring and the restrictions on the use environment.

[0323] (3) Optical transmitting system and optical receiving system: These two systems are responsible for transmitting and receiving optical signals. The optical transmitting system transmits optical signals of specific intensity and wavelength according to the instructions of the processor, while the optical receiving system is responsible for capturing the reflected photon echo and converting it into an electrical signal for the processor to analyze. Among them:

[0324] (3.1) The optical launch system includes:

[0325] (3.1.1) Laser transmitter: This is the light source of the entire system, responsible for generating laser pulses of the required wavelength. In satellite laser ranging, the laser transmitter needs to be able to generate pulsed lasers with high peak power and small beam divergence angle to ensure that the laser can be accurately and efficiently directed to the target satellite.

[0326] (3.1.2) Optical Transmission Module: This part mainly expands and collimates the laser emitted by the laser transmitter, so that the laser beam can be transmitted over a long distance without being too divergent. This helps to increase the energy density of the laser pulse when it reaches the satellite, thereby increasing the strength of the reflected signal.

[0327] (3.2) The optical receiving system includes:

[0328] (3.2.1) Optical receiving module: responsible for receiving the weak laser pulse signal reflected from the satellite under test. It needs to have a large enough aperture and high-precision optical design to effectively collect the reflected laser signal.

[0329] (3.2.2) Photodetection module: The photodetector is located at the focal plane of the optical receiving system and is used to convert the collected light signals into electrical signals. These electrical signals are then further amplified and processed.

[0330] (3.2.3) Signal amplifier and processing unit: This part is responsible for amplifying the weak electrical signal output by the photodetector and processing the signal through a specific algorithm to extract information about the satellite distance and time.

[0331] (4) Registers: Registers store specially designed program instructions, which are the basis for the normal operation of the system. When the processor executes these instructions, it will operate according to the preset photon echo power detection method to ensure the accuracy and reliability of the measurement results.

[0332] (II) Working principle:

[0333] The working principle of this system is based on the cat's eye effect and photon echo power detection technology. When the optical transmitting system sends out light signals, these signals will be reflected on the target object to form photon echoes. These echoes are captured by the optical receiving system and converted into electrical signals and sent back to the processor through the wireless transmitting device. The processor analyzes and processes these signals according to the program instructions in the register, and finally obtains the power value of the photon echo and displays it on the display.

[0334] All the above embodiments only express the implementation methods of the relevant practical applications of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

[0335] For those skilled in the art, it can be further appreciated that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.< / sample>

Claims

1. A photon echo power detection method based on cat's eye effect, characterized in that: This is achieved through the following steps: S1, establish a data set D, in which the image or point cloud data is applied with a 3D Hough transform to form the Hessian normal vector At the same time, the Markov matrix is ​​combined for fuzzy tracking to form the state X t+1 The probability of generating ij ; S2, the Hessian normal vector and the probability p ij , combined with different bands λ, different divergence angles α and different target surface roughness r, the feature vector FV is constructed; S3, inputting the SVM regression model: learning the relationship between the feature vector FV and the reflection distribution through the regression function f(x); and fusing the trained SVM regression model with the Markov matrix and the Hough transform; S4, the probability p ij , the Hessian normal vector The prediction results of the SVM regression model are combined using a voting mechanism to obtain a comprehensive prediction value of the reflection distribution. S5, based on the predicted value The distribution of the reflected intensity is adjusted by the incident angle θ to obtain the photon echo power P r ; The photon echo power P r The calculation method is: in, is the correction factor, f is the fitting function; τ is the laser atmospheric transmittance; R is the distance value; ρ is the diffuse reflection coefficient; A r is the effective receiving area; τ r is the transmittance of the receiving system; τ t is the transmittance of the emitting optical system; P t is the transmit power; The correction factor for: Where a and b are fitting parameters, θ0 is the reference incident angle, is the average prediction value of the SVM regression model on the training data.

2. The photon echo power detection method according to claim 1, characterized in that: The data set D is established by extracting different wavebands λ, different divergence angles α and different target surface roughness r according to historical data: D=[(λ1,α1,r1),(λ2,α2,r2),…,(λ n ,a n ,r n )]; Among them, (λ i , α i , r i ) is any sample in the data set D, and n is the number of samples.

3. The photon echo power detection method according to claim 1, characterized in that: In S1, the Hessian normal vector The method to obtain is: Among them, (n x ,n y ,n z ) are the x-axis, y-axis, and z-axis coordinates of the normal vector of a point in the image or point cloud data; In S1, the fuzzy tracking is to calculate the historical state sequence (X1, X2, ..., X t )'s next state X t+1 Implemented: P(X t+1 =j|X t =i)=p ij ; Among them, state X i Contains different position features and speed features; p ij From a state X i Transition to another state X j probability.

4. The photon echo power detection method according to claim 2, characterized in that: In S3, the regression function f(x) is constructed by the following steps: S300, for a given training set D train ={(x1,y1),(x2,y2),...,(x m ,y m )}, the regression function f(x) is fitted to it: Among them, x i is the input feature, that is, the feature vector FV; y i is the corresponding reflection distribution; m is the training set D train The number of samples; ω is the coefficient vector of the regression function; b is the bias term, is a function that maps input data to a high-dimensional space; ξ i and is a slack variable; C is a penalty coefficient; ε is a parameter of the insensitive loss function; T represents a transposition operation.

5. The photon echo power detection method according to claim 4, characterized in that: Mapping the input data to a high-dimensional space and the insensitive loss function are implemented by the following two steps: S301, introduce the radial basis kernel function K(x,x'): K(x,x′)=exp(-γ|x-x′| 2 ); Among them, γ is a positive hyperparameter; x and x' are two vector data points in the input space, which also correspond to the training set D train The samples in; exp is the exponential function of the base e of the natural logarithm; S302, the reflection distribution y i The predicted value and the reflected distribution y i There is a tolerance for the difference between the actual values: L(x,y,f(x))=max{0,|yf(x)|-ε}; Wherein, y is the actual value, the parameter ε is a pre-given positive number in this formula; L(x, y, f(x)) is an ε-insensitive loss function.

6. The photon echo power detection method according to claim 5, characterized in that: The training method of the SVM regression model constructed by S3 includes: The first step is initialization: for the training set D train , setting the label of the feature vector FV and the corresponding label of the reflection distribution; The second step is to set the optimization problem to solve the regression function f(x) using the training set D train The SVM regression model is trained, and the optimal penalty coefficient C and the hyperparameter γ are obtained using a cross-validation method; The third step is to distribute the reflection i The predicted value is compared with the actual value of the reflection distribution, the coefficient vector ω and the bias term b that minimize the mean square error are selected as the optimal parameters, and the relaxation variable ξ is automatically determined. i and Step 4: Repeat the above training steps until the maximum number of training iterations is reached.

7. The photon echo power detection method according to claim 1, characterized in that: In S3, the fusion method is: The Hessian normal vector The state X t+1 , the position predicted by the Markov matrix (x t+1 ,y t+1 ,z t+1 ) and its speed Combine into a new feature vector FV', and input it into the trained regression function f(x), and output a reflection distribution y i The measure of the predicted value.

8. The photon echo power detection method according to claim 7, characterized in that: In S4, the prediction result is the output F(x) of the regression function f(x), and the voting mechanism is implemented by the following fusion function G: Among them, (w1,w2,w3) are weight coefficients, function of (p ij )and is the probability p ij and the Hessian normal vector Transformed into the reflection distribution y i Predict related mapping functions; The Markov matrix, the three-dimensional Hough transform and the SVM regression model are all for a certain reflection distribution y i The predicted value It is determined by counting the votes in each category: Here, c represents different reflection distribution categories, and Votes(c) represents the number of votes obtained by category c.

9. A photon echo power detection system based on the cat's eye effect, comprising an optical transmitting system and an optical receiving system, characterized in that: The system includes a processor, a wireless receiving device connected to the processor, a wireless transmitting device connected to the optical transmitting system and the optical receiving system, and a register connected to the processor, wherein program instructions are stored in the register; The optical transmitting system and the optical receiving system realize data transmission with the processor based on the wireless connection between the wireless transmitting device and the wireless receiving device; When the program instructions are executed by the processor, the processor is caused to execute the photon echo power detection method as described in any one of claims 1-8.

Images