Line scanning point cloud target recognition method based on GABP neural network

Abstract

The application discloses a line scanning point cloud target recognition method based on a GABP neural network, first normalizes a full backup target point cloud feature multidimensional prior knowledge base constructed; in view of a line array laser radar imaging mode, fully utilizes time difference of adjacent imaging of the line array laser radar, designs a feature extraction network in a targeted manner, completes feature dimensioning and feature extraction of line scanning data; arranges multiple line scanning data in a scanning order to form a time sequence, inputs the time sequence into the feature extraction network in turn, and finally performs feature splicing and classification prediction; the designed GABP network is divided into two parts of a genetic algorithm parameter optimization and a BP neural network target recognition, through complementary advantages of the improved GA and the BP network, the GABP network has better generalization ability, so that the target recognition algorithm is optimized; the trained classifier model is loaded in an MCU of a flight vehicle, and rapid intelligent recognition and classification of battlefield targets can be completed.

Description

Technical Field

[0001] This invention pertains to lidar target recognition technology, specifically relating to a method for rapid target recognition of line-scan sparse point clouds based on a GABP neural network. Background Technology

[0002] The three-dimensional range image acquired by lidar imaging radar reflects the three-dimensional scale information of the target in the real world. Compared with traditional two-dimensional information, it contains more information, can describe the target more comprehensively and realistically, and is not affected by lighting conditions and target background contrast characteristics. Using three-dimensional point cloud information can also make full use of prior knowledge about the target in the real world, and can reliably achieve target identification in complex terrain environments.

[0003] Laser line pushbroom imaging is a one-dimensional scanning imaging method based on linear array detectors. It uses an array detector to receive echo signals, acquiring a row or column of target information in a single scan, thus obtaining three-dimensional target information through one-dimensional scanning. This method enables fast, high-resolution 3D imaging and is well-suited for applications involving moving platforms and large fields of view. Currently, almost all mainstream point cloud target recognition networks, such as PointPillars, VoxNet, PointNet++, and FastPointR-CNN, are based on acquiring complete 3D point cloud information of the target before data processing and target recognition. However, these networks have a large number of parameters, placing high demands on the computing and storage capabilities of the system. The imaging time interval between adjacent timestamps in laser line scanning can be fully utilized for data processing, enabling partial neural network feature extraction. Designing a specifically tailored neural network computation model can reduce the number of network parameters and significantly improve the speed of point cloud target recognition, shortening the target detection and recognition time.

[0004] Current research on point clouds mainly focuses on areas with low intersection speeds, and point clouds obtained by high-speed aircraft are typically sparse, necessitating improvements to existing processing algorithms. Backpropagation (BP) neural networks have slow convergence speeds and are prone to getting trapped in local minima; their weight parameters do not minimize error and cannot guarantee convergence to the global optimum. Therefore, genetic algorithms are introduced for optimization. As a global optimization algorithm simulating biological evolution, genetic algorithms possess excellent global optimization capabilities, iteratively evolving from a population to obtain the optimal or near-optimal solution. Combining GA genetic algorithms with backpropagation (BP) improves the stability of the neural network and shortens training time. Summary of the Invention

[0005] The purpose of the present application is to provide a line scanning point cloud target recognition method based on GABP neural network, which can quickly and effectively complete the identification and detection of the intended target.

[0006] The technical solution of the present application is a line scanning point cloud target recognition method based on GABP neural network, the steps are as follows:

[0007] Step 1, the line array laser radar is carried on the bottom surface of the aircraft head, there is a forward inclination angle between the detector of the line array laser radar and the aircraft, combined with the motion trajectory of the aircraft, the line array laser radar is used to scan and detect the detection area, and the measured sparse point cloud image of each target section in the detection area is obtained in turn, so as to obtain the measured data, and step 2 is entered.

[0008] Step 2, use Blender three-dimensional modeling software to simulate the radar detection process and output simulation data; combine the measured data and the simulation data to construct a full target point cloud feature multi-dimensional prior knowledge base, and enter step 3.

[0009] Step 3, design MLP multilayer perceptron, determine the network topology structure and feature component parameters at the same time, complete the structure and final loss function design of GABP neural network:

[0010] Among them, the network topology structure and feature component parameters include feature extraction layer, number of neurons in each layer, learning rate and activation function.

[0011] Enter step 4.

[0012] Step 4, use GA genetic algorithm to optimize GABP neural network to obtain the optimal initial parameters of GABP neural network, and enter step 5.

[0013] Step 5, use the optimal initial parameters of GABP neural network as the initial training parameters, use the full target point cloud feature multi-dimensional prior knowledge base and BP back propagation algorithm to train GABP neural network, obtain the optimal parameters of GABP neural network, and enter step 6.

[0014] Step 6, deploy the obtained optimal neural network parameters to the above GABP neural network to obtain a neural network classifier model, and then load the neural network classifier model in the MCU of the aircraft to realize the identification and classification of the battlefield target.

[0015] The present application has the following advantages compared with the prior art:

[0016] 1) According to the imaging characteristics of the line array laser radar, the feature extraction layer is designed to fully utilize the detection time difference of the radar at adjacent time stamps, shorten the feature extraction time and improve the recognition speed.

[0017] 2) The data amount of line scanning sparse point cloud is small, feature extraction is carried out on single timestamp point cloud, and then the splicing mode can greatly reduce the parameters of the neural network and reduce the calculation amount.

[0018] 3) The neural network is optimized by using the GABP algorithm, the GA algorithm finds the optimal initial weight, bias and threshold of the neural network, avoids the local optimal solution when the BP back propagation algorithm is used to solve the neural network, improves the stability of the neural network, shortens the training time, and further improves the recognition accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is a schematic diagram of line array laser radar scanning.

[0020] Figure 2 It is a GABP algorithm optimization flow chart.

[0021] Figure 3 It is a schematic diagram of neural network structure. DETAILED DESCRIPTION

[0022] The application will be further described below in combination with the drawings.

[0023] In combination with Figure 1 , Figure 2 and Figure 3The application discloses a line scanning point cloud target recognition method based on a GABP neural network. In the training process, a learning rate attenuation method is used to promote network convergence and accelerate the training speed, a corresponding loss function is designed to measure the training error, and in order to prevent overfitting, a weight attenuation method is used to update the gradient of the weight, and improved GA genetic algorithm is combined to use selection operation, crossover operation and mutation operation to continuously iterate to generate new individuals and populations, and the individual with the highest fitness is obtained as the initial weight parameter of the neural network, so that the neural network optimization process is prevented from falling into a local minimum value. The BP back propagation algorithm and the improved Adam gradient optimizer are used to continuously optimize the parameters, and the optimal network weight, bias and threshold are obtained. The obtained optimal neural network parameters are deployed in the above GABP neural network to obtain a neural network classifier model, and the obtained classifier model is loaded in the MCU of the aircraft, so that the rapid intelligent recognition and classification of battlefield targets can be realized.

[0024] The specific steps are as follows:

[0025] Step 1: mounting a line array laser radar on the bottom surface of the head of the aircraft, there being a front rake angle between the detector of the line array laser radar and the aircraft, combining the motion trajectory of the aircraft, using the line array laser radar to sweep the detection area, and sequentially obtaining measured sparse point cloud images of each target section in the detection area to obtain measured data.

[0026] Sweeping the detection area by using the line array laser radar specifically refers to parallel array push sweeping of the detection area of the aircraft by using the line array laser radar, and obtaining line array point cloud data composed of λ points in a single scan.

[0027] Go to step 2.

[0028] Step 2: using Blender three-dimensional modeling software to simulate the radar detection process and output simulation data. Jointly constructing a full target point cloud feature multidimensional prior knowledge base by using measured data and simulation data, and the specific steps are as follows:

[0029] The full-availability target point cloud feature multi-dimensional prior knowledge base includes the following information: target point cloud data of different categories, target point cloud data with a missing degree of 0-30%, and target point cloud data under a meeting angle of 0-360°; the full-availability target point cloud feature multi-dimensional prior knowledge base is divided into a training set, a validation set and a test set according to proportions of 80%, 10% and 10% respectively, and the point cloud normalization method is used to normalize the full-availability target point cloud feature multi-dimensional prior knowledge base to ensure that the data of the full-availability target point cloud feature multi-dimensional prior knowledge base has the same scale:

[0030]

[0031] wherein X i , Y i , Z i are three-axis coordinate components of a point in the linear array point cloud data composed of λ points, is the mean value of the three-axis coordinate components of a point in the linear array point cloud data, max is a maximum value function, X i ', Y i ', Z i ' are the three-axis coordinate components of a point after normalization;

[0032] The normalized training set, the normalized validation set and the normalized test set are obtained.

[0033] Go to step 3.

[0034] Step 3, design MLP multilayer perceptron, determine network topology structure and feature component parameters at the same time, complete the structure and final loss function design of GABP neural network:

[0035] Specifically:

[0036] Referring to the PointNet point cloud target feature extraction network, a GABP neural network for line scanning point cloud is designed. The linear array laser radar obtains one row of cross-section point cloud data every time it scans, which contains λ points, represented as λ×3, so that it sequentially passes through 5 layers of MLP multilayer perceptron, i.e. sequentially passes through convolution layers containing 64, 64, 128, 256 and 1024 neurons with shared weights, so that the data is sequentially upgraded in feature dimension to λ×64, λ×64, λ×128, λ×256 and λ×1024.

[0037] Assuming that it takes t times of scanning to obtain all the point cloud data of the target area, each time of scanning generates a timestamp, and the timestamps are arranged in time sequence to form a time sequence containing t timestamps, the time sequence is sequentially passed through the MLP multilayer perceptron to generate t groups of feature data of λ1×1024, λ2×1024…λ t ×1024, λ1, λ2, …, λ tThese represent the point cloud counts of different timestamps, concatenated along the first dimension of the data to form unm×1024, with the concatenation feature unm=λ. 1+ λ² + … + λ t After passing through a global maximum pooling layer, 1×1024 dimensional data is obtained. After computation by each neuron in the above layer, an activation function f(h) is applied.

[0038]

[0039] Where ε is a hyperparameter controlling the slope of the negative region, and h is the value of the neuron; the initial learning rate parameter is set to 0.01, and when the final loss function value does not decrease for ten consecutive times, the learning rate is automatically reduced to one-tenth of its original value.

[0040] Finally, after passing through two fully connected layers, the probability value of each category is output as 1×k, where k is the number of categories of the target to be identified, thus completing the structural design of the GABP neural network.

[0041] Based on the final output probability values ​​of each category, the cross-entropy loss function L, suitable for multi-class classification tasks, is adopted. cls :

[0042]

[0043] Where k is the number of categories of the target to be identified, pre[j] is the predicted probability value of the j-th category, pre[m] is the predicted probability value of the m-th category, and y[m] is the true predicted value of the corresponding pre[m].

[0044] w p Let L be the weights of the p-th layer neurons. The cross-entropy loss function L is obtained using the chain rule. cls The gradient grad with respect to the weights of each layer of neurons p (L cls ,w p ):

[0045]

[0046] The GABP neural network has a total of 8 layers of neurons, defined as L. cls The gradient of the weights of each neuron in each layer, after being L2 regularized, becomes the gradient regularization loss function L. 2reg1 :

[0047]

[0048] The weight parameters of each neuron in a layer are defined as the weight regularization loss function L after L2 regularization. 2reg2 :

[0049]

[0050] The final loss function includes a cross-entropy loss function, a gradient regularization loss function and a weight regularization loss function, and the total loss function is obtained after integration as follows:

[0051]

[0052] r1 and r2 are the proportion coefficients of the gradient regularization loss function and the weight regularization loss function, respectively.

[0053] Step 4 is entered.

[0054] Step 4, the GA genetic algorithm is used to optimize the GABP neural network to obtain the optimal initial parameters of the GABP neural network, and the specific method is as follows:

[0055] First, all the weights and bias parameters of the GABP neural network are encoded as the genes of the individual, and the gene range is set to , the initial population number is N, N should satisfy an integer multiple of t, and N individuals are randomly generated, the normalized training set data and the coded individuals are brought into the GABP neural network for training, and the prediction value pre of k classifications can be obtained, combined with the corresponding true value y, the fitness of the original population N individuals is calculated, and the fitness value fitness of the nth individual is obtained n :

[0056]

[0057] C is the fitness parameter, and for the population, the higher the fitness, the more the individual meets the conditions; the selection operation adopts the roulette method to "survival of the fittest" for individuals, so that the probability of each individual being selected is proportional to its fitness, and the selected probability prob is obtained n :

[0058]

[0059] The selection operation obtains two individuals A and B for pairing, the arithmetic crossover method is adopted for the crossover operation, and the genes of individual A and individual B in the crossover part and are operated as:

[0060]

[0061] wherein, is the new gene after gene crossover operation, and cross is the crossover parameter, which can control the proportion of genes obtained by the offspring from the two parents, thereby obtaining two new individuals, and a random parametercross Improvement, introduce cross parameter control coefficient improvement arithmetic crossover operation:

[0062]

[0063] Wherein, l c , r c All are natural numbers, r c For control degree of change, r c >0, p rand For random number, the improved parameter α cross By changing the coefficient adjustment in the parent class to get the proportion of the gene, mutation operation adopts the method of uniform mutation, for the individual gene mutation part X mut , after mutation, the new individual X′ mut :

[0064]

[0065] The upper and lower limits of individual gene coding are And γ mut Mutation parameter; The original population after completing the crossover mutation operation produces new individuals, and a new generation population is produced after multiple operations; After a finite number of iteration cycles, the individual with the highest fitness is obtained, the individual gene is decoded, and the optimal initial weights, bias and threshold of the neural network are obtained to avoid the local minimum value when solving the neural network by BP back propagation algorithm.

[0066] Go to step 5.

[0067] Step 5, using the optimal initial parameters of GABP neural network as the initial training parameters, using the full backup target point cloud feature multi-dimensional prior knowledge base and BP back propagation algorithm to train GABP neural network, and obtaining the optimal parameters of GABP neural network, the specific method is as follows:

[0068] The normalized training set data is input into the GABP neural network, and the optimal initial parameters of the GABP neural network obtained by the GA algorithm are used as the initial training parameters of the neural network, when the data is forward propagated, the value h p+1 Of the next layer of neurons is calculated in turn f p (w p h p +b p ) and the value L of the final loss function is calculated, the BP back propagation algorithm mainly depends on the negative direction of the gradient, which is the fastest direction of function value decrease, through the continuous training of the model, the weights, bias and threshold of the loss function relative to each layer of neurons are updated, to obtain the optimal network parameters;

[0069] Let h pw p b p is the bias of the pth layer neuron, f p (x) is the corresponding activation function; when backpropagation, by chain rule, the gradients of the loss function with respect to the value, weight and bias of each layer neuron are obtained:

[0070]

[0071] where, is the gradient of the pth layer neuron value, is the gradient of the pth layer weight, is the gradient of the pth layer bias, h p+1 is the value of the (p+1)th layer neuron.

[0072] When performing gradient update, the gradient optimization updates the weight of the neuron in this layer to obtain the new weight w p ′, referring to the Adam (Adaptive Moment Estimation) gradient optimizer and the SGD (Stochastic Gradient Descent) gradient optimizer, a new gradient optimizer is designed, which introduces historical information, i.e. momentum m p and velocity v p , respectively representing the exponential moving average of the gradient and the square of the gradient, while combining the stochastic gradient optimization value of SGD to complete the update of the weight of the neuron in this layer;

[0073] Momentum m p contains historical gradient information:

[0074]

[0075] Velocity v p contains information of historical gradient square:

[0076]

[0077] Combining momentum m p , velocity v p , weight decay coefficient τ and the stochastic gradient optimization value of SGD , the weight w p of the pth layer neuron is updated to obtain the new weight w p ′:

[0078]

[0079] Wherein β1, β2 is the historical information attenuation coefficient, θ is the SGD gradient optimizer weight coefficient, and the parameters of each layer of neurons can be updated in the same way. Each training will complete a parameter update, and the updated parameters will be used for testing the normalized validation set. The parameters with the best test effect are retained as the optimal parameters of the GABP neural network. The optimal parameters are used for the normalized test set verification, and the training effect of the neural network model can be obtained. The training frequency, learning rate, and attenuation coefficient parameters are adjusted in time.

[0080] Go to step 6.

[0081] Step 6, deploy the obtained optimal neural network parameters to the above-mentioned GABP neural network to obtain a neural network classifier model. Then, load the neural network classifier model in the MCU of the aircraft to realize the identification and classification of battlefield targets. The specific method is as follows:

[0082] Deploy the obtained optimal neural network parameters to the above-mentioned GABP neural network to obtain a neural network classifier model. When the aircraft carrying a linear array laser radar scans, it will trigger a collection signal every time it encounters a suspicious target. A time sequence corresponding to the GABP neural network is generated along the flight direction, and each time stamp contains the cross-sectional contour point cloud information of the target. Make full use of the detection time difference of the radar at adjacent time stamps to complete the feature extraction of a cross-sectional point cloud data within that time. That is, after completing a time stamp scanning, the cross-sectional point cloud is input into the neural network, and the feature data generated by the MLP multilayer perceptron is saved. Use the parallel computing capability of the MCU to realize scanning and computing at the same time, and process multiple cross-sectional point cloud data feature extraction at the same time. After completing the entire time sequence, all the feature data generated by the time stamp is spliced for subsequent feature fusion and classification network calculation, and finally the probability value of the target category is output. When the probability value is greater than 90%, it is considered that there is a classification target corresponding to the probability value in the scanning area.

[0083] Example 1

[0084] The line scanning point cloud target recognition method based on the GABP neural network uses the improved GA genetic algorithm combined with the optimized BP algorithm, so that the neural network has better generalization ability, the designed line scanning sparse point cloud imaging and time sequence splicing method is more in line with the detection process of the missile-borne laser radar. In order to verify the effectiveness of the algorithm design, the recognition accuracy and recognition time are used as evaluation indexes, and comparative experiments of different models are carried out. The neural network model without GA genetic algorithm is adopted, and the neural network model without GA genetic algorithm is adopted, and the PointNet classification recognition network model without time sequence is designed as a comparative experiment. The recognition effect of different models is measured on the test machine with GPU 1050Ti (4G) and CPU i7-7700HQ (2.8GHz), and the recognition effect of different models is measured on the test machine with GPU 1050Ti (4G) and CPU i7-7700HQ (2.8GHz).

[0085] Table 1 recognition effect of model

[0086]

[0087] Through the comparative experiment in Table 1, it can be known that the line scanning point cloud target recognition method based on the GABP neural network can effectively reduce the parameters of the neural network, reduce the calculation amount, greatly shorten the recognition time and improve the recognition accuracy. It is further demonstrated that the sparsity of the line scanning point cloud can greatly shorten the recognition time of the network model, and the improvement of the GABP algorithm and the fusion use of the time sequence feature can further improve the recognition accuracy. The recognition effect of the network model can be applied to the rapid intelligent recognition and classification of battlefield targets.

Claims

1. A line-scan point cloud target recognition method based on a GABP neural network, characterized in that, The steps are as follows: Step 1: Mount a linear array laser radar on the bottom surface of the aircraft head, and there is a forward inclination angle between the detector of the linear array laser radar and the aircraft. Combine the motion trajectory of the aircraft and use the linear array laser radar to sweep the detection area to obtain the measured sparse point cloud image of each target cross section in the detection area in turn to obtain the measured data, and proceed to step 2; Step 2: Use Blender three-dimensional modeling software to simulate the radar detection process and output simulation data; combine the measured data and simulation data to build a full target point cloud feature multi-dimensional prior knowledge base, and proceed to step 3; Step 3: Design a MLP multi-layer perceptron, determine the network topology structure and feature component parameters, and complete the structure and final loss function design of the GABP neural network: The network topology structure and feature component parameters include feature extraction layers, the number of neurons in each layer, learning rate, and activation function; Referring to the PointNet point cloud target feature extraction network, a GABP neural network for linear scanning point cloud is designed. The linear array laser radar obtains a row of cross section point cloud data containing λ points each time, which is represented as λ x 3, and is sequentially passed through 5 layers of MLP multi-layer perceptron, i.e., sequentially passed through convolution layers containing 64, 64, 128, 256, and 1024 neurons with shared weights, so that the data is sequentially feature dimensioned to λ x 64, λ x 64, λ x 128, λ x 256, and λ x 1024; Assuming that it takes t times of scanning to obtain the entire point cloud data of the target area, each time of scanning generates a timestamp, and the scanning time is arranged to form a time sequence containing t timestamps, the time sequence is sequentially passed through the MLP multi-layer perceptron to generate λ1×1024, λ2×1024…λ t t groups of feature data, λ1, λ2, …, λ t t represent the number of section point clouds corresponding to different timestamps, and are spliced along the first dimension of the data to form num×1024, and the spliced features num=λ 1+ λ2+…+λ t t, and an activation function f(h) is used after each layer of neuron calculation. Wherein, ε is a hyperparameter that controls the slope of the negative region, and h is the value of the neuron; The initial learning rate parameter is set to 0.01, and when the value of the final loss function does not decrease for ten consecutive times, the learning rate is automatically set to one tenth of the original value; Finally, the final output of each classification probability value 1 x k is obtained through two fully connected layers, k is the number of recognized target categories, and the structure design of the GABP neural network is completed; Proceed to step 4; Step 4: Optimize the GABP neural network using the GA genetic algorithm to obtain the optimal initial parameters of the GABP neural network, and proceed to step 5; Step 5: Use the optimal initial parameters of the GABP neural network as the initial training parameters, use the full target point cloud feature multi-dimensional prior knowledge base and the BP back propagation algorithm to train the GABP neural network in a loop to obtain the optimal parameters of the GABP neural network, and proceed to step 6; Step 6: Deploy the obtained optimal neural network parameters to the above GABP neural network to obtain a neural network classifier model, and load the neural network classifier model in the MCU of the aircraft to realize the recognition and classification of battlefield targets.

2. The GABP neural network-based line-scan point cloud target recognition method according to claim 1, characterized in that: In step 1, the linear array laser radar sweeps the detection area, which specifically refers to parallel array push scanning of the detection area of the aircraft by the linear array laser radar, and a single scan obtains linear array point cloud data composed of λ points.

3. The GABP neural network-based line-scan point cloud target recognition method according to claim 2, characterized in that: In step 2, the full target point cloud feature multi-dimensional prior knowledge base contains the following information: target point cloud data of different categories, target point cloud data with a missing degree of 0-30%, and target point cloud data under a 0-360° intersection angle; The full backup target point cloud feature multi-dimensional prior knowledge base is divided into a training set, a validation set and a test set according to a proportion of 80%, 10% and 10%, respectively, and the point cloud normalization method is used to normalize the full backup target point cloud feature multi-dimensional prior knowledge base to ensure that the full backup target point cloud feature multi-dimensional prior knowledge base data have the same scale. wherein X i , Y i , Z i are three-axis coordinate components of a point in the line array point cloud data composed of λ points, is the mean value of the three-axis coordinate components of a point in the line array point cloud data, max is a maximum value function, X i ', Y i ', Z i ' are the three-axis coordinate components of a point after normalization; The normalized training set, the normalized validation set and the normalized test set are obtained.

4. The GABP neural network-based line-scan point cloud target recognition method according to claim 3, characterized in that, In step 3: According to the probability values of each classification of the final output, a cross-entropy loss function L suitable for a multi-classification task is adopted cls : wherein k is the number of categories of the identified target, pre[j] is the jth category prediction probability value, pre[m] is the mth category prediction probability value, and y[m] is the real value corresponding to the mth category prediction probability value; w p w cls p (L cls ,w p )​ The structure of the GABP neural network has 8 layers of neurons, defined as L cls The gradient with respect to each layer of neuron weights is regularized with L2 regularization to form the gradient regularized loss function L 2reg1 : The weight regularization loss function L after defining the L2 regularization of each layer of neuron weight parameters is 2reg2 : The final loss function includes a cross-entropy loss function, a gradient regularization loss function and a weight regularization loss function, and the total loss function obtained after comprehensive is as follows: r1 and r2 are the proportion coefficients of the gradient regularization loss function and the weight regularization loss function, respectively.

5. The GABP neural network-based line-scan point cloud target recognition method according to claim 4, characterized in that: In step 4, the GA genetic algorithm is used to optimize the GABP neural network to obtain the optimal initial parameters of the GABP neural network, and the specific method is as follows: First, all weights and bias parameters of the GABP neural network are encoded as genes of an individual, with the gene range set within... The initial population size is N, where N must be an integer multiple of t. N individuals are randomly generated. The normalized training set data and the encoded individuals are fed into the GABP neural network for training, yielding predicted values ​​`pre` for k categories. These predicted values ​​`y` are then combined with the corresponding true values ​​`y` to calculate the fitness of the original N individuals, resulting in the fitness value of the nth individual. n : C is the fitness parameter, the higher the fitness of the population, the more the individual meets the conditions; the selection operation adopts the method of roulette to "survival of the fittest" for individuals, so that the probability of each individual being selected is proportional to its fitness, and the selected probability prob is obtained n : The selection operation obtains two individuals A and B to be paired, the crossover operation adopts an arithmetic crossover method, and the genes of the individual A and the individual B in a crossover part and are operated as follows: wherein, is a first gene after the gene crossover operation, is a second gene after the gene crossover operation, α cross is a crossover parameter, which controls the proportion of genes obtained from two parents to obtain two new individuals, and a random parameter α cross Improvement, the introduction of crossover parameter control coefficient improves the arithmetic crossover operation: wherein, l c , r c are natural numbers, r c is used to control the degree of change, r c > 0, p rand is a random number, the improved parameter a cross adjusts the proportion of genes obtained in the parent class by changing the coefficient; the mutation operation adopts a uniform mutation method, and for a mutation part X mut of the individual gene, a new individual X' mut is obtained after mutation. The upper and lower limits of the individual gene coding are and γ mut is a mutation parameter; the original population generates new individuals after completing the crossover mutation operation, and a new generation population is generated after multiple operations; after a limited number of iteration cycles, the individual with the highest fitness is obtained, the individual gene is decoded, and the optimal initial weights, biases and thresholds of the neural network are obtained to avoid the local minimum value when the BP back propagation algorithm is used to solve the neural network.

6. The GABP neural network-based line-scan point cloud target recognition method according to claim 5, characterized in that: In step 5, the optimal initial parameters of the GABP neural network are used as the initial training parameters, the full backup target point cloud feature multi-dimensional prior knowledge base and the BP back propagation algorithm are used to cyclically train the GABP neural network, and the optimal parameters of the GABP neural network are obtained, and the specific method is as follows: The normalized training set data is input into the GABP neural network, and the optimal initial parameters of the GABP neural network obtained by the GA algorithm are used as the initial training parameters of the neural network. p+1 = f p (w p h p +b p ) and the value L of the final loss function, the BP back propagation algorithm mainly relies on the negative direction of the gradient, which is the fastest direction of the function value, through the continuous training of the model, the weights, the bias and the threshold of the loss function relative to each layer of neurons are updated to obtain the optimal network parameters. Let h p be the value of the neuron of the p-th layer, w p be the weight of the neuron of the p-th layer, b p be the bias of the neuron of the p-th layer, f p (*) be the corresponding activation function; during backpropagation, by the chain rule, the gradients of the loss function with respect to the value, the weight and the bias of each neuron are obtained: wherein, is a gradient of the value of a neuron of the pth layer, is a gradient of the weight of the pth layer, is a gradient of the bias of the pth layer, h p+1 is a value of a neuron of the p+1th layer; In the gradient update, the gradient optimization is used to update the weights of the neurons in the layer to obtain new weights w p , referring to the Adam gradient optimizer and the SGD gradient optimizer, a new gradient optimizer is designed, which introduces historical information, i.e. momentum m p and velocity v p , respectively, representing the exponential moving average of the gradient and the square of the gradient, while combining the random gradient optimization value of SGD to complete the update of the weights of the neurons in the layer; Momentum m p Contains historical gradient information: Speed v p Information containing historical gradient squares: combining momentum m p , velocity v p , weight decay coefficient τ, and the stochastic gradient descent optimization value the weights w of the p-layer neurons p updating to obtain new weights w p ′: wherein β1 and β2 are the historical information decay coefficients, and θ is the SGD gradient optimizer weight coefficient, and the parameters of each layer of neurons can be updated in the same way. Each training will complete a parameter update, and the updated parameters are used for the test of the normalized validation set, the parameters with the best test effect are retained as the optimal parameters of the GABP neural network, the optimal parameters are used for the test of the normalized test set, the training effect of the neural network model is obtained, and the number of cyclic training times, the learning rate and the decay coefficient parameters are adjusted in time.

7. The GABP neural network-based line-scan point cloud target recognition method according to claim 6, characterized in that: In step 6, the obtained optimal neural network parameters are deployed into the above GABP neural network to obtain a neural network classifier model, the neural network classifier model is loaded in the MCU of the aircraft, and the identification and classification of the battlefield target are realized, and the specific method is as follows: The obtained optimal neural network parameters are deployed into the GABP neural network to obtain a neural network classifier model. When a suspicious target is encountered during scanning by the aircraft-mounted linear laser radar, a collection signal is triggered, a time sequence of t time stamps corresponding to the GABP neural network is generated along the flight direction, and each time stamp contains the cross-sectional profile point cloud information of the target. The detection time difference of the radar at adjacent time stamps is fully utilized to complete feature extraction of a cross-sectional point cloud data within the time, that is, after scanning of a time stamp is completed, the cross-sectional point cloud is input into the neural network, and the feature data generated by the MLP multilayer perceptron is saved. The parallel computing capability of the MCU is used to realize scanning and computing simultaneously, to simultaneously process feature extraction of multiple cross-sectional point cloud data, and after the entire time sequence is completed, the feature data generated by all the time stamps is spliced, subsequent feature fusion and classification network calculation are performed, and finally the probability value of the target category is output. When the probability value is greater than 90%, it is considered that the scanning area contains the classification target corresponding to the probability value.

Images