A laser ranging wall target automatic classification method and ranging correction method

Abstract

The present invention discloses an automatic classification method for wall and line targets in laser ranging and a ranging correction method; in the field of long-distance laser ranging, due to the different reflective surface characteristics of the measured target, there are obvious differences in the measurement results of typical targets such as walls and lines at the same distance. If the wall and line targets can be automatically distinguished and different data processing methods are used for different targets, the accuracy and universality of the long-distance target ranging system can be improved. The classification process of the present invention is: 1. Constructing a sample space T; 2. Clustering and filtering the noise of the sample space T using an adaptive K-Means clustering method to extract a new sample space T'. 3. According to the sample space T' obtained in step 2, calculate and extract 11-dimensional statistical feature variables therefrom; 4. Input the feature variable f into a trained classifier to perform binary classification on the measured target.

Description

Technical Field

[0001] The present invention belongs to the technical field of laser ranging, and in particular relates to a machine learning automatic classification method and a ranging correction method for large targets such as walls and small targets such as lines during laser ranging. Background Art

[0002] With the continuous development of my country's smart grid infrastructure, routine power inspection tasks are becoming more and more important. One of the important tasks is to detect the safety distance and sag curvature of overhead transmission lines. The current methods mainly include drone shooting and recognition and handheld laser ranging. The drone shooting and recognition method uses image processing technology to detect targets and identify defects, but the recognition of complex environments or natural defect samples is poor, and the drone operation requires professionals, which is difficult. Handheld laser ranging has the characteristics of high accuracy, fast measurement time, and intuitive measurement results, so it is more widely used in practical applications. At present, the research on laser ranging systems for close-range and large targets is relatively mature, but the field of measurement for long-distance small target overhead lines is still very rare, and there is no ranging system that can be compatible with the precise measurement of large and small targets. In complex environments such as long-distance power inspections, it is easy to be disturbed by the external environment, such as the background noise introduced by the optical lens, the influence of the weather environment, and the reflective surface of the measured target. These factors seriously affect the ranging accuracy.

[0003] In addition, due to the characteristics of the cylindrical shape of overhead power lines, the reflection angles corresponding to different positions on the surface of the power lines are different. Therefore, when the laser spot is hit at different positions on the reflection surface of the power lines, the echo signal strength and the steepness of the signal rising edge will be significantly different from those of large targets such as walls. Since the characteristics of the measurement targets of different sizes have a significant impact on the ranging results, they must be distinguished. However, relying on manual selection by the user, it is often impossible to reasonably and accurately select the target situation. Automatically selecting the target type according to the echo situation of the laser ranging and using different fitting methods for different types can improve the intelligence of the ranging system. At present, some studies have conducted preliminary research on the reflection characteristics of the laser ranging echo signal, but there is no research result on the automatic classification of targets such as large targets such as walls and small targets such as lines. Summary of the invention

[0004] The purpose of the present invention is to provide a method for automatic classification of wall line targets based on machine learning, so as to realize a ranging system with universality for different targets and good ranging accuracy.

[0005] In a first aspect, the present invention provides a method for automatic classification of wall line targets based on machine learning, which comprises the following steps:

[0006] Step 1: Use the laser ranging device to collect data of the target multiple times to obtain a sample space T = {x1, x2, ..., x N}; Each sample x i It is a four-tuple consisting of four variables: gain q, distance dist, threshold time difference Δt1 and Δt2, i = 1, 2, ..., N.

[0007] Step 2: Use the adaptive K-Means clustering method to cluster the sample space T to filter out noise and extract a new sample space T′={x′1,x′2,…,x′ N′}.

[0008] 2-1. Standardize all variables in the sample space T and unify the dimension of each variable.

[0009] 2-2. Define sample x i The local density function Density(x i ), whose value is the W nearest neighbor samples from sample x i The mean of the Euclidean distance, i = 1, 2, ..., N. W is the preset number of neighbors. The most dense point and the most sparse point in the sample space T are defined as the cluster center C k According to the K-Means clustering method, the sample space is iterated and each sample x is calculated. i With each cluster center C k The similarity of the sample x i Attributed to its closest cluster center C k , so that each cluster center forms a cluster. At the same time, all samples x in the same cluster are calculated separately i With cluster center C k The sum of squared errors (Sum of Squared Error, SSE); the sum of squared errors greater than the set threshold V SSE Sample x i Defined as the new cluster center C k , and update the neighboring points of the cluster center.

[0010] 2-3. After the clustering iteration is completed, the cluster with the largest number of samples is defined as the new sample space T′={x′1,x′2,…,x′ N′}.

[0011] Step 3: Based on the sample space T′ obtained in step 2, calculate and extract the 11-dimensional statistical feature variable f={N sample ,N′,N error ,Q,D,D jitter ,D var ,ET mid ,ETjitter ,ET std ,G}. Among them, N sample is the number of triggers of the hardware circuit for data acquisition in step 1; N′ is the number of samples in the sample space T′; N error is the number of samples removed in step 2; Q and D are the gain q and distance dist corresponding to the sample with the smallest threshold time difference Δt1 in the sample space T′; D jitter and D var are the range and variance of all distances dist in the sample space T′; ET mid , E.T. jitter , E.T. std are the median, range, and standard deviation of all threshold time differences Δt1 in the sample space T′; G is the proportion of the number of samples with waveform saturation in the total number of samples in the sample space T′; the threshold time difference Δt2 is greater than the preset saturation threshold V growth The sample x′ i is the sample at which the waveform is saturated.

[0012] Step 4: Input the feature variable f into the trained classifier to perform binary classification on the target to determine whether the target belongs to a wall target or a line target.

[0013] As a preferred embodiment, in step 1, the hardware circuit is triggered a number of times N sample When the number of times exceeds 100 or the amount of data collected N is greater than 50, data collection ends.

[0014] Preferably, in step 2-1, the calculation formula for normalizing the variable v is as follows:

[0015]

[0016] Among them, v and v' are the variables before and after standardization respectively; Q(50) is the median of all variables of the same type as variable v, and Q(75)-Q(25) is the interquartile range of all variables of the same type as variable v.

[0017] As a preference, in step 2-2, the preset value of the number of neighbors W is 5, and the threshold V is set SSE The value of is 1.

[0018] Preferably, the classifier in step 4 adopts an XGB+LR classifier model. The XGB+LR classifier model includes a series of extreme gradient boosting (eXtreme Gradient Boosting, XGB) feature combination layer, random forests (RandomForests, RF) feature selection layer and logistic regression (Logistic Regression, LR) classification layer.

[0019] Preferably, the specific process of step 4 is as follows:

[0020] 4-1. The XGB feature combination layer model automatically combines different features in the feature vector f to construct a feature combination vector l. Set the parameter complexity complexity and parameter depth depth. Use the labeled training set to train the classifier. After the training is completed, the number of leaf nodes is 134.

[0021] 4-2. Use the random forest RF method to evaluate the importance of the feature combination vector l, filter out some features with higher importance, and obtain the updated vector l′; filter the threshold V RF Defined as:

[0022]

[0023] 4-3. Send the vector l′ into the trained logistic regression LR binary classification model to obtain the final output classification result y.

[0024] Preferably, in step 1, the threshold time difference Δt1 is the time difference between the echo signal amplitude reaching the threshold voltage V1 and reaching the threshold voltage V2; the threshold time difference Δt2 is the time difference between the echo signal amplitude reaching the threshold voltage V2 and reaching the threshold voltage V3; V1<V2<V3.

[0025] Preferably, in step 4-1, the complexity parameter is set to 20, and the depth parameter is set to 3.

[0026] In a second aspect, the present invention provides a laser ranging correction method, the process of which is as follows:

[0027] According to the aforementioned classification method, determine whether the measured object belongs to a wall target or a line target. Select the corresponding fitting formula according to the type of the measured object. Substitute the gain q, distance dist, and threshold time difference Δt1 obtained by laser ranging for the measured object into the fitting formula to obtain the actual distance to the target object.

[0028] Preferably, the fitting formula is a polynomial correction formula for fitting gain q, distance dist, and threshold time difference Δt1 about the actual distance through least squares method; the data set used in the fitting process is test data for targets of the same type (wall or line).

[0029] The beneficial effects of the present invention are:

[0030] 1. The wall line automatic classification method proposed in the present invention solves the problems of environmental noise and difference in the reflective surface of the measured target in laser ranging, and realizes a ranging system with universality for different targets and good ranging accuracy.

[0031] 2. Traditional data denoising methods cannot filter out all noise based on simple conditional judgment. The adaptive K-Means clustering method adopted in the present invention adaptively adjusts the number of cluster centers according to the sample space distribution, and has a good processing effect on outlier noise.

[0032] 3. The present invention calculates and extracts feature vectors from the sample space and feeds them into the XGB+LR integrated classifier. Compared with the traditional weak learner, the classification accuracy is improved, and compared with other integrated learners, the calculation time is reduced, and finally the ranging accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the laser ranging hardware circuit structure of the present invention;

[0034] Figure 2 is a flow chart of automatic classification performed by the present invention;

[0035] Figure 3 It is a comparison diagram of the clustering noise removal effect of the present invention and the existing method;

[0036] Figure 4 It is a comparison chart of the classification accuracy of the present invention and the existing method;

[0037] Figure 5 It is a comparison chart of the classification time overhead of the present invention and the existing method;

[0038] Figure 6 This is a comparison chart of the accuracy of the distance measured by the present invention and the distance measured without wall line classification. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the accompanying drawings.

[0040] First, the process of laser ranging is described as follows:

[0041] The front-end hardware circuit is a pulse laser ranging system (such as Figure 1 As shown in Figure 2), the automatic gain control process and the dual threshold correction compensation process are introduced. The control circuit triggers the laser tube to emit pulses and calibrates the starting time t start The automatic gain control circuit automatically adjusts the gain q of the receiving circuit according to the amplitude of the echo signal to meet the condition that the amplitude of the echo signal is greater than the threshold voltage V1, and marks this time as the stop time t end , the time measurement module calculates the flight time t of the laser pulse end -t start Calculate the distance to the target: dist = c (t end -t start ) / 2, where c is the speed of light in vacuum 3×10 8m / s. The threshold voltages V2 and V3 are set to be used to correct and detect the saturation of the waveform, respectively. Assume that the corresponding receiving times are t2 and t3, respectively. The two threshold time difference formulas are defined as follows:

[0042] Δt1=t2-t end (1)

[0043] Δt2=t3-t2 (2)

[0044] V1<V2<V3; the values ​​of the threshold voltages V2 and V3 are determined according to the prior art and will not be elaborated here.

[0045] According to the dual threshold correction theory, two thresholds V1 and V2 are fixedly selected, and the time difference between the two thresholds is calculated to correct the error caused by the amplitude of the echo signal. The true value of the flight time t r The expression is shown in formula (3):

[0046] t r =t end -e t (3)

[0047] Among them, e t is the flight time error.

[0048] Using the Gaussian model to model the laser echo signal, we can get e t Regarding the threshold time difference Δt1 error correction expression:

[0049]

[0050] Among them, τ s is a fixed constant, representing the echo pulse width. Finally, by measuring the threshold time difference Δt1 and substituting it into formula (4), we can obtain e t , and then e t Substitute into formula (3) to obtain the corrected real flight time.

[0051] like Figure 2 As shown, a ranging correction method based on automatic classification of wall line targets in laser ranging includes the following steps:

[0052] Step 1: Collect data through a handheld laser rangefinder. When the measurement switch is pressed, the ranging system starts to trigger and collect data. The hardware circuit is triggered N times. sample When the number of collected data exceeds 100 or the amount of data N is greater than 50, the processor extracts the N data containing noise measured by the ranging system to form the following sample space: T = {x1, x2, ..., x N}, each sample x iis a quaternion, including four variables: gain q, distance dist, threshold time difference Δt1 and Δt2. Among them, variable q is the current gain value of the automatic gain control circuit; variable dist is the flight distance converted from the laser pulse flight time; variable Δt1 is used to detect waveform quality and compensation correction, the smaller the value, the better the waveform quality; variable Δt2 is used to determine the saturation degree of the waveform; i=1,2,...,N.

[0053] Step 2: Use the adaptive K-Means clustering method to cluster the sample space T to filter out noise and extract useful data clusters as the new sample space T′={x′1,x′2,…,x′ N′}, remove the influence of useless noise signals on the ranging system.

[0054] 2-1. Perform RobustScaler standardization on the sample space T to unify the dimension of each variable. The standardization calculation formula is as follows:

[0055]

[0056] Among them, v and v' are variables before and after standardization, respectively; Q(index) is the quantile of index index in variable v, Q(50) is the median of all variables of the same type as variable v (for example, when standardizing a gain q, Q(50) is the median of all gains q), and Q(75)-Q(25) is the interquartile range of all variables of the same type as variable v. Compared with the Min-Max normalization method, the RobustScaler method reduces the impact of outliers on standardization, and compared with the Z-Score centering method, it retains the outlier characteristics to the maximum extent, and has good adaptability to noisy laser ranging systems.

[0057] 2-2. Define sample x i The local density function Density(x i ), whose value is the W nearest neighbor samples from sample x i The mean of the Euclidean distance, i = 1, 2, ..., N. According to the maximum and minimum idea, the densest point in the sample space (the sample point with the smallest density function) is defined as the first cluster center C1, and the sparsest point (the sample point with the largest density function) is defined as the second cluster center C2. According to the K-Means clustering method, the sample space is iterated to calculate each sample x i The similarity (Euclidean distance) with each cluster center is calculated, and the sample is assigned to its closest cluster center C. k , so that each cluster center forms a cluster. In addition, the sample that causes the largest error square sum SSE is selected in the calculation. If the error square sum of the sample is greater than the set threshold V SSE, then define the sample as the new cluster center and update the neighboring points of the cluster center to quickly reduce SSE and reduce the impact of outliers on the clustering effect. W is set to 5, V SSE Set to 1.

[0058] 2-3. After the clustering iteration is completed, the cluster with the largest number of samples is defined as the new sample space T′={x′1,x′2,…,x′ N′}.

[0059] Step 3: Based on the sample space T′ obtained in step 2, calculate and extract 11-dimensional statistical feature variables.

[0060] 3-1. Count the number of times N the ranging system hardware circuit triggers sampling required to obtain a sample space T with N samples sample , the number of samples N' in the sample space T', the number of noises N removed by K-Means clustering in step 2 error , calculated as shown in formula (6):

[0061] N error =NN′ (6)

[0062] 3-2. Statistically calculate the gain Q and distance D corresponding to the sample with the smallest threshold time difference Δt1 in the sample space T′.

[0063] 3-3. Calculate the maximum value D of all distances dist in the sample space T′ max , minimum value D min , variance D var , the maximum value ET of the threshold time difference Δt1 max , minimum ET min , median ET mid , Standard Deviation ET std , and convert the range D of all distances dist jitter The extreme difference ET of the time difference Δt1 from the threshold jitter as follows:

[0064] D jitter =D max -D min (7)

[0065] ET jitter =ET max -ET min (8)

[0066] 3-4. If the sample x′ i The threshold time difference Δt2 is greater than the preset saturation threshold V growth , then the echo signal amplitude of this sample is judged to be saturated, and the saturation of this sample is defined as g iis 1; otherwise, it is judged that the echo signal amplitude of this sample is not saturated, and the saturation g i is 0, i=1,2,...,N′. Calculate the waveform saturation (saturation g i The number of 1) accounts for G, as shown in formula (9):

[0067]

[0068] 3-5. Construct 11-dimensional feature vector f = {N sample ,N′,N error ,Q,D,D jitter ,D var ,ET mid ,ET jitter ,ET std ,G}.

[0069] Step 4: Input the feature vector f into the trained XGB+LR classifier model, perform binary classification on the measured target, and obtain the target value y (y∈{line, wall}) output by the classifier, so as to determine whether the measured target belongs to a wall or a line. The classifier is composed of an XGB feature combination layer, an RF feature selection layer, and an LR classification layer in series. The XGB model performs feature combination to make the data high-dimensional and linearly separable. The LR linear model has good processing capabilities for high-dimensional sparse matrices. At the same time, since the feature combination vector output by XGB is highly sparse, a RF feature selection layer is added before the LR model to reduce the complexity of the model.

[0070] 4-1. The XGB feature combination layer model automatically combines different features in the feature vector f to construct a feature combination vector l. The parameter complexity (number of decision trees) complexity is set to 20. The parameter depth (size of decision trees) depth is set to 3. For the training set Train{(f1,y1),(f2,y2),…,(f 2016 ,y 2016 )} for training. After the training is completed, the number of leaf nodes is 134, that is, the dimension of the feature combination vector l is 134, of which the number marked as 1 is 20, and the rest are 0.

[0071] 4-2. Use the random forest RF method to evaluate the importance of the feature combination vector l and filter out some features with higher importance to prevent the learner from falling into overfitting. Screening threshold V RF Defined as:

[0072]

[0073] In this embodiment, the screening threshold V RFThe value of is 1 / 160; after screening, the dimension of the feature combination vector l is reduced, and a new feature combination vector l′ is obtained, whose dimension is 35.

[0074] 4-3. Send the vector l′ into the trained logistic regression LR binary classification model to obtain the final output classification result y, where y∈{line, wall}.

[0075] Step 5: According to the target classification value y obtained in step 4, the measured gain q, distance dist, and threshold time difference Δt1 are substituted into the corresponding fitting formula for correction to obtain the actual distance from the ranging system to the measured target. There are two fitting formulas. The two fitting formulas correspond to wall targets and line targets respectively.

[0076] The fitting formula is a polynomial correction formula for the output variable (true distance) by fitting three input variables (gain q, distance dist, threshold time difference Δt1) through the least squares method. The data set used in the fitting process is the test data of the same type (wall or line) target.

[0077] In order to verify the denoising effect of the adaptive K-Means clustering proposed in the present invention, a comparative test on the operation time and accuracy of the adaptive K-Means method provided by the present invention, the existing K-Means method and the K-Means++ method was conducted; the results are compared as follows: Figure 3 shown; from Figure 3 It can be seen that the K-Means method has the lowest computing time and accuracy. K-Means++ has a faster computing speed, but its accuracy is affected by the K value. Adaptive K-Means achieves the best clustering effect at the expense of a little time overhead.

[0078] In order to verify the performance of the wall line classification model used in the present invention, experiments were conducted on using the LR classifier alone, the RF classifier alone, the GBDT classifier alone, the GBDT+LR classifier, and the XGB+LR method provided by the present invention. The results of the classification accuracy are compared. Figure 4 As shown; the results of the classification time cost comparison are as follows Figure 5 shown; from Figure 4 It can be seen that the accuracy of GBDT+LR and XGB+LR classifiers is significantly higher than that of the other three classifiers. The accuracy of XGB+LR classifier is 1.3% higher than that of GBDT+LR. Figure 5 It can be seen that the operation time of GBDT+LR and XGB+LR is more expensive than the other three classifiers, while the operation speed of XGB+LR is better than that of GBDT+LR method.

[0079] In order to verify the improvement of the laser ranging performance of the present invention, the automatic classification method of the wall line is not used, and the method in the specific example is tested; the results of the distance measurement standard deviation are compared. Figure 6 shown; from Figure 6 It can be seen that the method proposed in the present invention can better remove environmental noise and distinguish the target object with a higher accuracy. The standard deviation ranges of the targets classified as wall targets and line targets are 0-0.2 and 0.05-0.25 respectively, which are lower than the automatic wall and line classification method (standard deviation range is 0-0.4), thereby improving the ranging accuracy.

[0080] In the process of actual measurement of a line target in the present invention, after 100 times of hardware circuit triggering, the processor collects 32 sample data, which form the sample space T as follows:

[0081] T={x1(1100,43.31,3230,4434),x2(1500,83.01,1666,5434),…,x 32 (1500,42.76,970,3834)} (11)

[0082] After adaptive K-Means clustering, 7 noise data are filtered out, and the remaining 25 sample data are used as the new sample space T′={x′1,x′2,…,x′ 25}, from which the 11-dimensional feature vector f is calculated and extracted as follows:

[0083] f={100,25,7,1500,42.76,3.21,0.45,4270,2560,864,0.51} (12)

[0084] The feature vector f is input into the pre-trained XGB+LR wall line classification model, and the classification target result y is 0 (line target). Then it is substituted into the fitting formula of the line target to obtain the corrected true distance of 38.45 meters. The actual measurement result (calibrated with the TruPulse200 high-precision laser rangefinder) is 38.3 meters, with an error of 0.15 meters.

Claims

1. A laser ranging wall line target automatic classification method, characterized by: The following steps are involved: Step 1: Use the laser ranging device to collect data of the target multiple times to obtain a sample space consisting of N samples: ; Each sample is a four-tuple containing the gain ,distance , threshold time difference and These four variables, i=1,2,...,N; threshold time difference When the amplitude of the echo signal reaches the threshold voltage With reaching the threshold voltage time difference; threshold time difference When the amplitude of the echo signal reaches the threshold voltage With reaching the threshold voltage time difference; < < ; Step 2: Use the adaptive K-Means clustering algorithm to cluster the sample space Clustering is performed to filter out noise and extract new sample space ; 2-1. Sample space All variables in the are standardized to unify the dimensions of each variable; 2-2. Defining the sample The local density function , whose value is the W nearest neighbor sample distance sample The mean of the Euclidean distance, i=1,2,...,N; W is the preset number of neighbors; the densest and sparsest points in the sample space T are defined as cluster centers ; Iterate the sample space according to the K-Means clustering algorithm and calculate each sample With each cluster center The similarity of Assign to its closest cluster center , so that each cluster center forms a cluster; at the same time, all samples in the same cluster are calculated separately With cluster center The error sum of squares is greater than the set threshold Sample Defined as the new cluster center , and update the neighboring points of the cluster center; 2-3. After the clustering iteration is completed, the cluster with the largest number of samples is defined as the new sample space ; Step 3: Sample space obtained from step 2 , from which 11-dimensional statistical feature variables are calculated and extracted ;in, is the number of times the hardware circuit for data acquisition in step 1 is triggered; Sample Space The number of samples in ; is the number of samples removed in step 2; and The sample space Middle Threshold Time Difference The gain q and distance corresponding to the smallest sample ; and The sample space All distances The range and variance of , , The sample space All threshold time differences median, range, and standard deviation; The sample space The proportion of the number of samples with waveform saturation in the total number of samples; the threshold time difference Greater than the preset saturation threshold Sample is the sample of waveform saturation; Step 4: Input the feature variable f into the trained classifier to perform binary classification on the target to determine whether the target belongs to a wall target or a line target.

2. The method for automatic classification of wall line targets by laser ranging according to claim 1, characterized in that: In step 1, the hardware circuit is triggered several times When the number of times exceeds 100 or the amount of data collected N is greater than 50, data collection ends.

3. The method for automatic classification of wall targets by laser ranging according to claim 1, characterized in that: In step 2-1, the calculation formula for standardizing the variable v is as follows: ； Among them, v and v' are the variables before and after standardization respectively; is the median of all variables of the same type as variable v, is the interquartile range of all variables of the same type as variable v.

4. The method for automatic classification of wall line targets by laser ranging according to claim 1, characterized in that: In step 2-2, the preset value of the number of neighbors W is 5, and the threshold is set The value of is 1.

5. The method for automatic classification of wall line targets by laser ranging according to claim 1 is characterized by: The classifier described in step 4 adopts an XGB+LR classifier model; the XGB+LR classifier model includes a serially connected XGB feature combination layer, an RF feature selection layer and an LR classification layer.

6. The method for automatic classification of wall line targets by laser ranging according to claim 1, characterized in that: The specific process of step 4 is as follows: 4-1. XGB feature combination layer model transforms feature vector Automatically combine different features in the ; Set the parameter complexity complexity and parameter depth depth; Use the labeled training set to train the classifier. After the training is completed, the number of leaf nodes is 134; 4-2. Use the random forest RF algorithm to combine feature vectors Perform importance evaluation, filter out some features with higher importance, and obtain the updated vector ; Screening threshold Defined as: ； 4-3. Vector Send the trained logistic regression LR binary classification model to obtain the final output classification result .

7. The method for automatic classification of wall line targets by laser ranging according to claim 6 is characterized by: In step 4-1, the complexity parameter is set to 20, and the depth parameter is set to 3.

8. A laser ranging correction method, characterized in that: According to the method for automatically classifying wall and line targets in laser ranging as claimed in any one of claims 1 to 7, it is determined that the measured object belongs to a wall target or a line target; a corresponding fitting formula is selected according to the type of the measured object; a gain obtained by performing laser ranging on the measured object is converted into a linear target; a linear target is obtained by performing laser ranging on the measured object ... ,distance , threshold time difference Substitute it into the fitting formula to get the actual distance to the target object.

9. A laser ranging correction method according to claim 8, characterized in that: The fitting formula is to fit the gain by the least squares method ,distance , threshold time difference Polynomial correction formula for true distance; The data set used in the fitting process is the test data for the same type of target.

Images