Improved mrf-based gm-apd array lidar imaging reconstruction method, device, storage medium and program product

Abstract

The application discloses an improved MRF-based GM-APD array laser radar imaging reconstruction method, equipment, a storage medium and a program product, belongs to the technical field of laser imaging, and solves the problem that a peak intensity threshold cannot distinguish signal and background noise in a low signal-to-background ratio case.The method comprises the following steps: extracting a peak value feature and a curvature feature according to a trigger histogram Y; acquiring a feature fusion distance image and a feature fusion intensity image according to the peak value feature and the curvature feature; acquiring a target noise distribution initial value according to the feature fusion distance image and the feature fusion intensity image; uniformly processing an intensity image to acquire an observation image; and inputting the target noise distribution initial value and the observation image into a Markov random field to obtain a reconstructed image.The application is suitable for extraction and target reconstruction of low echo data of a laser imaging radar, and is specifically used for improving the image signal-to-noise ratio of a reconstructed three-dimensional image.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of laser imaging, in particular to GM-APD array laser radar imaging. BACKGROUND

[0002] The Gm-APD laser imaging radar is single-photon imaging, the detector has high sensitivity, and is greatly affected by background light. When the signal photon number to noise photon number ratio (signal to background ratio) in the gated window is lower than 0.01, the reconstruction is very difficult, and it is also very difficult to distinguish the signal and background noise. At present, the reconstruction of low signal to noise ratio echo data mainly uses data enhancement means, sparse reconstruction, spatial superposition, etc., and the peak value is extracted to obtain distance and intensity information. The signal and background pixel points are distinguished through the threshold of peak value intensity. It is considered that the target echo is stronger than the background noise. In the case of low signal to background ratio, the background noise level is high, and the intensity information has certain randomness, which causes the peak value intensity threshold to be unable to distinguish the signal and background noise. SUMMARY

[0003] The purpose of the present application is to solve the problem that the peak value intensity threshold cannot distinguish the signal and background noise in the case of low signal to background ratio, and provide a GM-APD array laser radar imaging reconstruction method based on improved MRF, equipment, storage medium and program product.

[0004] The present application is realized by the following technical scheme. In one aspect, the present application provides a GM-APD array laser radar imaging reconstruction method based on improved MRF, which comprises:

[0005] Obtaining a trigger histogram Y obtained by statistics of detector echo data;

[0006] According to the trigger histogram Y, extracting peak value features and curvature features;

[0007] According to the peak value features and curvature features, obtaining a feature fusion distance image and a feature fusion intensity image;

[0008] According to the feature fusion distance image and the feature fusion intensity image, obtaining a target noise distribution initial value;

[0009] Uniformly processing the intensity image to obtain an observation image;

[0010] Inputting the target noise distribution initial value and the observation image into a Markov random field to obtain a reconstruction image.

[0011] Further, the extracting peak value features and curvature features according to the trigger histogram Y specifically comprises:

[0012] According to the trigger histogram Y, and using the following feature extraction formula:

[0013]

[0014]

[0015]

[0016]

[0017] wherein, rng is the distance image obtained by peak feature, cnt is the intensity image obtained by peak feature, rng' is the distance image obtained by curvature feature, cnt' is the intensity image obtained by curvature feature, i p is the horizontal coordinate of array pixel, j p is the vertical coordinate of array pixel, Y is the trigger histogram obtained by statistics of detector echo data, Y" is the second derivative of Y;

[0018] extracting peak feature and curvature feature.

[0019] Further, the method for obtaining the feature fusion distance image, specifically comprises:

[0020] using point-to-point correlation, the distance values under two features, judging the size of distance difference obtained by respective features and the allowed error distance radius Eps, obtaining the feature fusion distance image rng new , specifically comprising:

[0021]

[0022] When the condition is met, the distance value of the feature fusion distance image can be obtained:

[0023]

[0024] When the distance difference is greater than the parameter Eps:

[0025]

[0026] using its neighborhood space points to calculate the distance radius R p :

[0027]

[0028] wherein, n p is the number of neighborhood distance value points, D pis an array of difference values of the neighborhood distance values and the center point distance values, arranged in ascending order of the difference values;

[0029] The distance value of the fusion distance image is obtained by judging the relationship between the neighborhood calculation distance radius and Eps.

[0030]

[0031] .

[0032] Further, the feature fusion intensity image acquisition method specifically comprises:

[0033] The intensity difference of each feature is judged by using the point-to-point correlation and the intensity values of two features, and the size of the allowed error intensity radius CEps is obtained. cnt new Specifically, it comprises:

[0034]

[0035] When the condition is met, the intensity value of the feature fusion intensity image can be obtained:

[0036]

[0037] When the intensity difference is greater than the parameter CEps:

[0038]

[0039] The intensity radius is calculated using the neighborhood space points C p :

[0040]

[0041] Wherein, cn p The number of neighborhood points with non-zero intensity values, DC p is an array of difference values of the neighborhood distance values and the center point distance values, arranged in ascending order of the difference values;

[0042] The intensity value of the fusion intensity image is obtained by judging the relationship between the neighborhood calculation intensity radius and CEps.

[0043]

[0044] .

[0045] Further, according to the feature fusion distance image and the feature fusion intensity image, the target noise distribution initial value is obtained, specifically comprising:

[0046] performing density clustering on the fusion distance image to obtain classes performing density clustering on the fusion distance image to obtain classes for each class a class center where x, y are the horizontal and vertical coordinate values of the pixels within the class, and the dispersion of each class is:

[0047]

[0048] The formula for converting the dispersion into a class probability is:

[0049]

[0050] combining the fusion intensity image into a 0-1 probability distribution, and multiplying the two to obtain the final probability map:

[0051]

[0052] setting the points with P = 0 as noise points and the other points as target points to obtain the initial target noise distribution I 0.

[0053] Further, the intensity image is uniformly processed to obtain an observation image, specifically including:

[0054] The intensity image is uniformly processed using data processing methods, specifically including:

[0055]

[0056] where Y' is the preprocessed trigger histogram, Y is the original histogram, w is a spatial kernel function with a size of a x a x b, and Y is convolved with w for preprocessing;

[0057] The preprocessed trigger histogram Y' is processed using a concave-convex search algorithm to extract and preprocess signals, obtaining an observation image X, and simultaneously obtaining a distance image Rng for related processing with the MRF output result final and an intensity image Cnt final .

[0058] Further, the initial target noise distribution and the observation image are input into a Markov random field to obtain a reconstructed image, specifically including:

[0059] Each pixel of the observation image X corresponds to a label of the scene , and the total number of pixels N = 64 x 64 = 4096, where M is the number of classes, is the probability distribution of I; the current image where X is the observed image; determine the maximum probability result of I obtained by final iteration, that is, calculate its conditional probability:

[0060]

[0061] According to the Bayesian theory:

[0062]

[0063] P I Use Gibbs random field probability, the probability value of which is determined by the initial value and the iteration process, and the initial value of the first step is brought into the following formula for calculation I 0 and :

[0064]

[0065] wherein Z is a normalization factor, T controls the distribution form of P, C is the set of all potential groups, and U is the potential energy;

[0066]

[0067] wherein β is the coupling coefficient; it is considered that each label is a Gaussian distribution, and the mean and variance are calculated by the corresponding value of the label in the objective function;

[0068]

[0069]

[0070] The joint maximum probability is

[0071]

[0072] Each time the joint maximum probability is calculated, the updated I is obtained, and after multiple iterations, the stable I is obtained, which is the result of the final I. The result of I is related to Y' after the pre-processing of Y, the distance image Rng final and the intensity image Cnt final obtained by the concave-convex search signal extraction, and the final reconstructed distance image and intensity image are obtained.

[0073] In a second aspect, the application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the steps of the improved MRF-based GM-APD array laser radar imaging reconstruction method described above are executed.​

[0074] In a third aspect, the present application provides a computer readable storage medium, wherein a plurality of computer instructions are stored in the computer readable storage medium, and the plurality of computer instructions are used to make a computer execute the improved MRF-based GM-APD array lidar imaging reconstruction method as described above.

[0075] In a fourth aspect, the present application provides a computer program product, wherein the computer program is executed by a processor to realize the improved MRF-based GM-APD array lidar imaging reconstruction method as described above.

[0076] Advantages of the present application:

[0077] Firstly, the present application realizes more accurate and faster class calculation based on the initial value and target image in the improved Markov random field.

[0078] Secondly, the signal-to-noise ratio of the reconstructed image is improved by distinguishing the classes of the target and the background.

[0079] The method of the present application can improve the signal-to-noise ratio of the GM-APD lidar reconstructed image in the case of low signal-to-background ratio.

[0080] 1. More accurate range image distribution is obtained by combining different distance features, secondly, target distribution probability is obtained by using distance value which is more stable than intensity value, finally, more reliable target distribution probability map is obtained by combining distance and intensity, thereby obtaining the initial value of the Markov random field.

[0081] 2. On the target image, the spatial correlation is used for pretreatment, and then the concave-convex search is used for signal extraction to obtain the range image and intensity image with good uniformity.

[0082] 3. The initial value is brought into the Markov random field to calculate the probability of the pixel points, and the output class map is used to obtain the range image and intensity image with high signal-to-noise ratio.

[0083] The present application is suitable for the extraction and target reconstruction of Geiger mode-avalanche photodiode (Gm-APD) laser imaging radar low SBR (Signal Background radio) echo data, and is specifically used for improving the image signal-to-noise ratio of the reconstructed three-dimensional image. BRIEF DESCRIPTION OF DRAWINGS

[0084] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0085] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.Figure 1 Flowchart of the method of the present application;

[0086] Figure 2 Results of the present application with different initial values through MRF;

[0087] Figure 3 Results of the present application with different target images through MRF;

[0088] Figure 4 Distance image and intensity image obtained using different reconstruction algorithms under different conditions of the present application. DETAILED DESCRIPTION

[0089] Embodiments of the present application are described in detail below with reference to the attached drawings, which show by way of example, embodiments in which the same or similar elements have the same or similar reference numbers and in which:

[0090] Embodiment 1. A GM-APD array laser radar imaging reconstruction method based on improved MRF, the method comprising:

[0091] Obtaining a trigger histogram Y from statistics of detector echo data;

[0092] According to the trigger histogram Y, extracting peak value features and curvature features;

[0093] According to the peak value features and curvature features, obtaining a feature fusion distance image and a feature fusion intensity image;

[0094] According to the feature fusion distance image and the feature fusion intensity image, obtaining a target noise distribution initial value;

[0095] Uniformly processing the intensity image to obtain an observation image;

[0096] Inputting the target noise distribution initial value and the observation image into a Markov random field to obtain a reconstruction image.

[0097] In this embodiment, 1. A more accurate distance image distribution is obtained by combining different distance features, the target distribution probability is obtained in a more stable manner using distance values than intensity values, and finally a more reliable target distribution probability map is obtained by combining distance and intensity, thereby obtaining the initial value of the Markov random field.

[0098] 2. On the target image, spatial correlation is used for preprocessing, and concave-convex search is used for signal extraction to obtain a distance image and an intensity image with good uniformity.

[0099] 3. The initial value is brought into the Markov random field for pixel probability calculation, and the output category map is used to obtain high signal-to-noise ratio range images and intensity images.

[0100] In the embodiment two, the embodiment is a further limitation of the GM-APD array laser radar imaging reconstruction method based on the improved MRF in the embodiment one. In the embodiment, the extraction of the peak value feature and the curvature feature according to the trigger histogram Y is further limited, and specifically includes:

[0101] According to the trigger histogram Y, and using the following feature extraction formula:

[0102]

[0103]

[0104]

[0105]

[0106] wherein, rng is a range image obtained by the peak value feature, cnt is an intensity image obtained by the peak value feature, rng' is a range image obtained by the curvature feature, cnt' is an intensity image obtained by the curvature feature, i p is the horizontal coordinate of the array pixel, j p is the vertical coordinate of the array pixel, Y is a trigger histogram obtained by the detector echo data statistics, and Y” is the second derivative of Y;

[0107] The peak value feature and the curvature feature are extracted.

[0108] In the embodiment, the extraction method of the peak value feature and the curvature feature is specifically given, which is used to obtain more accurate range image distribution by combining different distance features, to obtain target distribution probability by using distance value which is more stable than intensity value, and to obtain more reliable target distribution probability map by combining distance and intensity, so as to obtain the initial value of the Markov random field.

[0109] In the embodiment three, the embodiment is a further limitation of the GM-APD array laser radar imaging reconstruction method based on the improved MRF in the embodiment two. In the embodiment, the acquisition method of the feature fusion range image is further limited, and specifically includes:

[0110] The point-to-point correlation is used to judge the size of the distance difference obtained by each feature and the allowed error distance radius Eps by using the distance values under two features, so as to obtain the feature fusion range image rngnew , specifically comprising:

[0111]

[0112] When the condition is met, the distance value of the feature fusion distance image can be obtained:

[0113]

[0114] When the distance difference is greater than the parameter Eps:

[0115]

[0116] Calculate the distance radius using its neighborhood space points R p :

[0117]

[0118] wherein, n p is the number of neighborhood distance value points, D p is the difference array of neighborhood distance values and center point distance values, arranged in ascending order of difference value;

[0119] The distance value of the fusion distance image is obtained by judging the relationship between the neighborhood calculated distance radius and Eps:

[0120]

[0121] .

[0122] In this embodiment, the distance images rng and rng' extracted from two features are fused on the target image using the spatial correlation of distance, to obtain a more accurate feature fusion distance image.

[0123] Embodiment four, the present embodiment is a further limitation of the GM-APD array laser radar imaging reconstruction method based on improved MRF described in embodiment three, in the present embodiment, the acquisition method of the feature fusion intensity image is further limited, specifically comprising:

[0124] Using point-to-point correlation, the intensity values under two features are used to judge the size of the intensity difference obtained by each feature and the allowed error intensity radius CEps, to obtain a feature fusion distance image cnt new : Specifically comprising:

[0125]

[0126] When the condition is met, the intensity value of the feature fusion intensity image can be obtained:

[0127]

[0128] When the intensity difference is greater than the parameter CEps:

[0129]

[0130] Calculate the intensity radius using the neighborhood spatial points C p :

[0131]

[0132] Wherein, cn p is the number of neighborhood points with non-zero intensity values, DC p is the difference array of neighborhood intensity values and center point intensity values, arranged in ascending order of difference value;

[0133] The intensity value of the fused intensity image is obtained by judging the relationship between the neighborhood calculated intensity radius and CEps:

[0134]

[0135] .

[0136] In this embodiment, the distance images cnt and cnt' extracted from two features using the spatial correlation of intensity on the target image are fused to obtain a more accurate intensity image.

[0137] Embodiment five, this embodiment is a further limitation of the GM-APD array laser radar imaging reconstruction method based on improved MRF described in embodiment four, in this embodiment, the initial value of target noise distribution is further limited according to the feature fusion distance image and the feature fusion intensity image, specifically including:

[0138] The fused distance image is subjected to density clustering to obtain classes , for each class The class center , wherein x and y are the horizontal and vertical coordinate values of the pixels within the class, and the dispersion of each class is:

[0139]

[0140] The formula for converting the dispersion into class probability is:

[0141]

[0142] Combined with the fused intensity image Convert to 0-1 probability distribution, multiply both to get the final probability map:

[0143]

[0144] Set the label of the point P as 0 to the noise point, and set the other points to the target point to get the initial value of the target noise distribution I 0.

[0145] In this embodiment, a more accurate distance image distribution is obtained by combining different distance features, a target distribution probability is obtained by using distance values which are more stable than intensity values, and a more reliable target distribution probability map is obtained by combining distance and intensity, thereby obtaining the initial value of the Markov random field.

[0146] Embodiment six, this embodiment is a further limitation of the GM-APD array laser radar imaging reconstruction method based on improved MRF according to embodiment five, in this embodiment, the intensity image is uniformly processed to obtain an observation image, which is further limited, specifically including:

[0147] The intensity image is uniformly processed using a data processing method, specifically including:

[0148]

[0149] Where Y' is the preprocessed trigger histogram, Y is the original histogram, w is the spatial kernel function, and the size is a x a x b, Y is convolved with w for preprocessing;

[0150] The preprocessed trigger histogram Y' is processed using a concave-convex search algorithm to extract and preprocess the signal, to obtain an observation image X, and to obtain a distance image Rng for related processing with the MRF output result final And the intensity image Cnt final .

[0151] In this embodiment, the observation image X is used to bring it into the Markov random field for pixel point probability calculation, and the output category map is used to obtain a distance image and an intensity image with high signal-to-noise ratio.

[0152] It should be noted that the concave-convex search algorithm used in this embodiment is a method disclosed in a kind of Gm-APD laser radar low signal-to-noise ratio echo data signal extraction method based on concave-convex search, which is a prior art with publication number CN111060887A and publication date April 24, 2020.

[0153] Embodiment seven, the embodiment is one kind based on the further limitation of the GM-APD array laser radar imaging reconstruction method of improved MRF of embodiment six, in the embodiment, the further limitation is made to the reconstruction image obtained by inputting the target noise distribution initial value and the observation image into Markov random field, and specifically includes:

[0154] The label of each pixel of the observation image X corresponding to the scene is The total number of pixels N=64x64=4096, wherein M is the number of categories, is the probability distribution of I; the current image Wherein X is the observation image; determine the maximum probability result of I obtained by final iteration, that is, to find its conditional probability:

[0155]

[0156] According to the Bayesian theory:

[0157]

[0158] P ( I )Use Gibbs random field probability, the probability value is determined by the initial value and the iteration process, the initial value of the first step I 0 is brought into the following formula for calculation And :

[0159]

[0160] Wherein, Z is a normalization factor, T controls the distribution form of P, C is the set of all clusters, is the potential energy;

[0161]

[0162] Wherein, β is the coupling coefficient; The observation image is composed of multiple labels, based on the central limit theorem, it is considered that the intensity of each label is approximately Gaussian distribution, and the mean and variance are calculated using the label corresponding value in the objective function;

[0163]

[0164]

[0165] The joint maximum probability is

[0166]

[0167] Each time the joint maximum probability is calculated, the I is updated iteratively, and after multiple iterations, the stable I is obtained as the final result of I; the result of I is correlated with the distance image and intensity image obtained in the second step to obtain the final reconstructed distance image and intensity image.

[0168] In this embodiment, the Markov random field performs pixel point probability calculation, and the output category map is used to obtain a distance image and an intensity image with high signal-to-noise ratio.

[0169] Embodiment eight, this embodiment is a specific embodiment of a GM-APD array laser radar imaging reconstruction method based on improved MRF as described above, as shown in Figure 1 , comprising:

[0170] Step 1: Calculate the initial value of MRF (Markov random field)

[0171] Step one: extract peak value features and curvature features, the target distances extracted by the two features have similarities and the noise distances have differences, and the fusion can highlight the target distance, the feature extraction formula: specifically includes:

[0172]

[0173]

[0174]

[0175]

[0176] Among them rng is the distance image obtained by the peak value feature, cnt is the intensity image obtained by the peak value feature, rng' is the distance image obtained by the curvature feature, cnt' is the intensity image obtained by the curvature feature, i p is the horizontal coordinate of the array pixel, j p is the vertical coordinate of the array pixel, Y is the trigger histogram obtained by the detector echo data statistics, and Y” is the second derivative of Y.

[0177] Step two: using point-to-point correlation, the distance values under the two features are used to judge the size of the distance difference obtained by each feature and the allowed error distance radius Eps, to obtain the feature fusion distance image rng new : specifically includes:

[0178]

[0179] When the condition is met, the distance value of the feature fusion distance image can be obtained:

[0180]

[0181] When the distance difference is greater than the parameter Eps:

[0182]

[0183] The distance radius is calculated using its neighborhood space points R p :

[0184]

[0185] Wherein n p The number of neighborhood distance value points, D p The difference array of neighborhood distance values and center point distance values, arranged in ascending order of difference.

[0186] The distance value of the fusion distance image is obtained by judging the relationship between the neighborhood calculated distance radius and Eps:

[0187]

[0188]

[0189] Step three: using point-to-point correlation, the intensity values of two features, judging the size of the intensity difference obtained by each feature and the allowed error intensity radius CEps, obtaining the feature fusion intensity image cnt new : Specifically includes:

[0190]

[0191] When the condition is met, the intensity value of the feature fusion intensity image can be obtained:

[0192]

[0193] When the intensity difference is greater than the parameter CEps:

[0194]

[0195] The intensity radius is calculated using its neighborhood space points C p :

[0196]

[0197] Wherein cn pThe number of points with non-zero intensity values in the neighborhood, DC p The difference array of the intensity values of the neighborhood and the center point, arranged in ascending order of the difference value.

[0198] The intensity value of the fusion intensity image is obtained by judging the relationship between the neighborhood calculation intensity radius and CEps:

[0199]

[0200]

[0201] Step four: density clustering on the fusion distance image to obtain the class For each class The class center where x and y are the horizontal and vertical coordinate values of the pixels in the class, and the dispersion of each class is:

[0202]

[0203] The formula for converting the dispersion to the class probability is:

[0204]

[0205] Convert the fusion intensity image to a 0-1 probability distribution, and multiply the two to obtain the final probability map:

[0206]

[0207] Set the label of points P to noise points, and set other points to target points to obtain the initial value of the target noise distribution I 0.

[0208] Second step: target image preprocessing

[0209] The points in the same class are as uniform as possible, and the points in different classes have a gap to obtain an accurate Gaussian probability distribution. Unlike the median filter which only processes the intensity image, this paper uses data processing methods to uniformly process the intensity image. Specifically, it includes:

[0210]

[0211] where Y' is the preprocessed trigger histogram, Y is the original histogram, w is a spatial kernel function with size a x a x b, and Y is convolved with w for preprocessing,

[0212] The preprocessed trigger histogram Y' is used to extract and preprocess the signal using the concave-convex search algorithm to obtain the observation image X.

[0213] Third step: put I 0 and X input Markov random field to get the reconstructed image

[0214] Each pixel of X corresponds to the label of the scene, and the label is The total number of pixels is N = 64x64 = 4096, where M is the number of categories, is the probability distribution of I. The current image , where X is the observed image. Determine the maximum probability result of I obtained by final iteration, that is, find its conditional probability:

[0215]

[0216] According to the Bayesian theory:

[0217]

[0218] P ( I ) Use Gibbs random field probability, whose probability value is determined by initial value and iterative process, and put the initial value of the first step I 0 into the following formula for calculation and : (It should be noted that I 0 is put into the following formula, but I will change with iteration, so I 0 is not reflected in the formula)

[0219]

[0220] where Z is the normalization factor, T controls the distribution of P, C is the set of all potential groups, is the potential energy.

[0221]

[0222] where β is the coupling coefficient, usually 0.5-1. Multiple labels are composed in the observed image, based on the central limit theorem, it is considered that the intensity of each label is approximately Gaussian distribution, and the mean and variance are calculated using the corresponding value of the label in the objective function.

[0223]

[0224]

[0225] Joint maximum probability is

[0226]

[0227] Each time the joint maximum probability is calculated, an iteratively updated I is obtained. After multiple iterations, a stable I is obtained. The final I is the result. The result of I is correlated with the pre-processed Y', the distance image Rngfinal and the intensity image Cntfinal obtained by the concave-convex search signal extraction, to obtain the final reconstructed distance image and intensity image.

[0228] The classification results of different initial values are respectively random initial values, threshold segmented initial values, and the initial values calculated by us. The proportion of correct and incorrect categories obtained by MRF output is shown in Figure 2 It can be seen that the results of different initial values output by MRF iteration are quite different. Among them, the proportion of correct and incorrect calculated by us is the highest, and the iteration converges the fastest.

[0229] The classification results of different target images are respectively the peak intensity image, the peak intensity image after median filtering, and the intensity image calculated by us. The proportion of correct and incorrect categories obtained by MRF output is shown in Figure 3 It can be seen that the results of different targets output by MRF iteration are quite different. Our target image can reach iterative convergence faster, and has higher accuracy.

[0230] The reconstruction results of different methods are shown in Figure 4 The transverse comparison shows that under the same signal-to-background ratio, the distance image obtained by our reconstruction algorithm has less background noise, and the target on the intensity image is clearer. The longitudinal comparison shows that as the signal-to-background ratio decreases, the image reconstruction effect becomes worse and worse, and the reconstruction difficulty becomes higher and higher. However, our reconstruction algorithm can maintain a certain stability, and can still extract most of the target pixel points under low signal-to-background ratio.

Claims

1. A method for reconstructing images using a GM-APD array lidar based on an improved MRF, characterized in that, The method includes: The trigger histogram Y is obtained by statistically analyzing the detector echo data. Based on the trigger histogram Y, extract peak features and curvature features; Based on the peak and curvature features, feature fusion distance and feature fusion intensity profiles are obtained; Based on the feature fusion distance profile and feature fusion intensity profile, the initial value of the target noise distribution is obtained; The intensity image is homogenized to obtain the observed image; The initial value of the target noise distribution and the observed image are input into a Markov random field to obtain a reconstructed image.

2. The imaging reconstruction method for a GM-APD array lidar based on an improved MRF as described in claim 1, characterized in that, The step of extracting peak features and curvature features based on the trigger histogram Y specifically includes: Based on the trigger histogram Y, and using the following feature extraction formula: in, rng The distance profile obtained from the peak features cnt Intensity image obtained from peak features, rng' The distance image obtained from the curvature features. cnt' Intensity image obtained from curvature features. i p is the x-coordinate of the array pixels. j p Y is the vertical coordinate of the array pixels, Y is the trigger histogram obtained from the statistical analysis of the detector echo data, and Y” is the second derivative of Y. Extract peak and curvature features.

3. The GM-APD array lidar imaging reconstruction method based on improved MRF according to claim 2, characterized in that, The method for obtaining the feature fusion distance image specifically includes: Using point-to-point correlation, the distance values ​​under two features are compared to determine the magnitude of the distance difference between the respective features and the allowable error distance radius Eps, resulting in a feature fusion distance profile. rng new Specifically, it includes: When the conditions are met, the distance value of the feature fusion distance image can be obtained: When the distance difference is greater than the parameter Eps: Calculate the distance radius using its neighboring spatial points R p : in, n p The number of points with distance values ​​in the neighborhood. D p This is an array of differences between neighborhood distance values ​​and center point distance values, arranged in ascending order of differences. The distance value of the fused range image is obtained by determining the relationship between the distance radius of the neighborhood calculation and Eps: 。 4. The GM-APD array lidar imaging reconstruction method based on improved MRF according to claim 3, characterized in that, The method for obtaining the feature fusion intensity image specifically includes: Using point-to-point correlation, the intensity values ​​under two features are compared with the magnitude of the intensity difference obtained from each feature and the allowable error intensity radius (CEps) to obtain the feature fusion distance profile. cnt new Specifically, it includes: When the conditions are met, the intensity value of the feature fusion intensity image can be obtained: When the strength difference is greater than the parameter CEps: Calculate the intensity radius using its neighborhood spatial points C p : in, cn p This represents the number of neighborhood points with non-zero intensity values. DC p This is an array of differences between the intensity values ​​of the neighborhood and the intensity value of the center point, arranged in ascending order of the differences; The intensity value of the fused intensity image is obtained by determining the relationship between the intensity radius of the neighborhood calculation and CEps: 。 5. The GM-APD array lidar imaging reconstruction method based on improved MRF according to claim 4, characterized in that, Based on the feature fusion distance profile and feature fusion intensity profile, the initial value of the target noise distribution is obtained, specifically including: For fused distance images Perform density clustering to obtain categories For each category Category Center Where x and y are the horizontal and vertical coordinates of pixels within a category, respectively, the dispersion of each category is: The formula for converting dispersion into class probability is: Combined fusion intensity image Convert to a 0-1 probability distribution, and multiply the two to obtain the final probability map: The point labeled with P=0 is designated as a noise point, and the other points are designated as target points, thus obtaining the initial value of the target noise distribution. I 0.

6. The GM-APD array lidar imaging reconstruction method based on improved MRF according to claim 5, characterized in that, The process of uniformly processing the intensity image to obtain the observed image specifically includes: The intensity image is homogenized using data processing techniques, specifically including: Where Y' is the preprocessed trigger histogram, Y is the original histogram, w is the spatial kernel function with size a×a×b, and Y and w are convolved for preprocessing; The preprocessed trigger histogram Y' is subjected to signal extraction and preprocessing using a concavity-convexity search algorithm to obtain the observed image X. Simultaneously, the distance image Rng is obtained and correlated with the MRF output. final And intensity like Cnt final .

7. The GM-APD array lidar imaging reconstruction method based on improved MRF according to claim 6, characterized in that, The step of inputting the initial value of the target noise distribution and the observed image into a Markov random field to obtain the reconstructed image specifically includes: The label of the scene corresponding to each pixel in the observed image X is: The total number of pixels N = 64 × 64 = 4096, of which M is the number of categories. The probability distribution of I; the current image Where X is the observed image; determine the maximum probability result of I obtained in the final iteration, that is, calculate its conditional probability: Based on Bayesian theory: P ( I Using Gibs random field probabilities, the probability value is determined by the initial value and the iteration process. The initial value in the first step... I Substitute 0 into the formula below for calculation. and : Where Z is the normalization factor. T controls the distribution pattern of P. Let C be the set of all potential cliques. Potential energy; in, β The coupling coefficient is assumed to be ; each label is assumed to follow a Gaussian distribution, and the mean and variance are calculated using the corresponding values ​​of the labels in the objective function. The highest probability of a union is Each time, the joint maximum probability is calculated to obtain the iteratively updated I. After multiple iterations, a stable I is obtained, which is the final result of I. The result of I is then combined with Y' after preprocessing Y to extract the distance image Rng from the concavity / convexity search signal. final And intensity like Cnt final Relevant calculations are performed to obtain the final reconstructed distance and intensity images.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, The steps of the method according to any one of claims 1 to 7 are performed when the processor runs the computer program stored in the memory.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of computer instructions, which are used to cause a computer to perform the method of any one of claims 1 to 7.

10. A computer program product, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images