Intelligent driving obstacle target detection method based on millimeter wave radar

Abstract

The invention discloses an intelligent driving obstacle target detection method based on a millimeter-wave radar, and the method comprises the steps: carrying out the feature extraction of a convolutional neural network algorithm innovation based on twin grey wolf optimization, carrying out the feature dimension reduction of a self-encoding neural network algorithm based on manifold learning, and carrying out the classification algorithm of a support vector machine based on hierarchical reinforcement learning. A learning process classifier model training method is improved through a dynamic forgetting factor, the analysis capability of radar signals is improved, the adaptability and robustness of the model to a complex environment are enhanced, and more accurate and more reliable obstacle recognition can be realized, so that an intelligent driving system can recognize and respond to various traffic participants and obstacles in real time, and the intelligent driving system is more accurate and reliable. And powerful support is provided for ensuring driving safety and efficiency.

Description

Technical Field

[0001] The present invention relates to the field of intelligent driving detection, and in particular to an intelligent driving obstacle target detection method based on millimeter wave radar. Background Art

[0002] Intelligent driving systems need to be able to identify and respond to various traffic participants and obstacles in real time to ensure driving safety and efficiency. Traditional vision-based perception systems have limited performance in bad weather or low light conditions, while millimeter-wave radar has become one of the key technologies in intelligent driving perception systems due to its insensitivity to environmental conditions.

[0003] Existing technologies have some limitations when processing millimeter-wave radar data. For example, the analysis and feature extraction of radar signals may not be precise enough, resulting in low accuracy in obstacle recognition. In addition, traditional data processing methods may lack adaptability to complex environmental changes and cannot effectively handle uncertainty and noise in dynamic traffic scenarios. These problems limit the performance and reliability of intelligent driving systems. There is an urgent need for a method that can further data processing to deal with the poor performance of existing technologies in the face of noise and outliers, the lack of rapid adaptability of traditional classifiers to new data features, and the lack of depth in feature extraction, to provide strong support for the safe operation of intelligent driving systems. Summary of the invention

[0004] Purpose of the invention:

[0005] To solve the problems mentioned in the background technology, the present invention discloses an intelligent driving obstacle target detection method based on millimeter-wave radar. Through innovative feature extraction, feature dimension reduction and classifier model training methods, the analysis capability of radar signals is improved, the adaptability and robustness of the model to complex environments are enhanced, and more accurate and reliable obstacle identification can be achieved, thereby providing strong support for the safe operation of the intelligent driving system.

[0006] Technical solution:

[0007] The present invention discloses a method for detecting intelligent driving obstacle targets based on millimeter wave radar, the method comprising the following steps:

[0008] S1 collects millimeter-wave radar detection signal data as a training set and labels the data;

[0009] S2 builds a feature extraction model based on the twin gray wolf optimized convolutional neural network algorithm and inputs the training set into the feature extraction model for feature extraction training;

[0010] S3 uses the autoencoder neural network algorithm based on manifold learning to build a feature dimensionality reduction model, and inputs the training set after feature extraction into the feature dimensionality reduction model for feature dimensionality reduction training;

[0011] S4 builds a classifier model based on the support vector machine classification algorithm of hierarchical reinforcement learning, puts the reduced-dimensional training set into the classifier to train the classifier model, uses a dynamic forgetting factor to improve the learning process, and dynamically adjusts the learning rate according to the importance of historical training data;

[0012] S5 uses the trained composite model to process the newly acquired millimeter-wave radar detection signal and outputs the final obstacle recognition result.

[0013] Furthermore, the feature extraction model construction training process described in S2 is as follows:

[0014] S201 constructs two convolutional neural network branches, each of which has the same layer structure, including convolutional layer, pooling layer and fully connected layer, and initializes network parameters. Specifically, set W (0) Represents the weight set of the network in the initial state, and the initialization method is random initialization;

[0015] S202 divides the collected radar signal into two parts. The first part is the normal radar signal, and the second part is the radar signal after Fourier transformation. They are input into two branches of the twin network and preprocessed, which are expressed as:

[0016]

[0017] In the formula, X is the input data, μ and σ are the mean and standard deviation of the data respectively, and X norm is the normalized data;

[0018] Each branch in S203 processes its input data and extracts features through convolutional layers and pooling layers. The way in which the convolution operation extracts features is expressed as follows:

[0019] F=Re(W*X norm +b)

[0020] In the formula, * represents the convolution operation, W and b are the weight and bias of the convolution layer respectively, Re() is the ReLU activation function, and F represents the extracted features;

[0021] S204 fuses the features output by the two branches, and uses cosine similarity as the fused feature representation:

[0022]

[0023] Where F1 and F2 are the features extracted from the two branches of the two-branch network, ∥∥ is the L2 norm, and S is the cosine similarity between the features;

[0024] S205 uses the gray wolf algorithm to optimize the weight parameters of the convolutional neural network, simulates the social hierarchy and hunting strategy of the gray wolf, and updates the network weights through group wisdom to find the global minimum of the loss function. The update method of the weight parameters is expressed as:

[0025] W new '=W alpha -α dq |β dq W prey -W new |

[0026] α dq =2a s rand(0,1)-a s

[0027] β dq =2a s rand(0,1)

[0028] Where W new ' is the updated weight; W alpha is the current optimal solution, that is, the optimal parameter in the current iteration process; W prey is the target solution, i.e. the optimal parameter; W new is the current solution, α dq and β dq is the control parameter, a s is the attenuation parameter, rand(0,1) is a random number in the interval [0, 1];

[0029] S206 performs back propagation according to the loss function and updates the network parameters:

[0030]

[0031] Where W new ' is the weight after optimization by the gray wolf algorithm, λ xa is the learning rate, is the gradient of the loss function with respect to the weights, W next is the weight of the next iteration;

[0032] S207 repeats the above steps until a preset stop iteration condition is met.

[0033] Furthermore, the weight W and bias b of the convolutional layer are updated using a mini-batch gradient descent method:

[0034]

[0035] Where η cf is the learning rate, m is the batch size, x i and i is the i-th sample in the batch and its label, is the partial derivative of the loss function with respect to the weight, W new is the updated weight, L() is the loss function L(W old ,x i ,y i ), is the parameter W old ,x i ,y i The loss function when W old is the weight before updating;

[0036] The update method of bias b is expressed as:

[0037]

[0038] In the formula, is the partial derivative of the loss function with respect to the bias, b new is the updated bias, b old is the bias before updating;

[0039] Learning rate λ xa It is set for dynamic adjustment, expressed as:

[0040]

[0041] In the formula, is the learning rate of the tth iteration, exp() is an exponential function based on the natural base e, is the learning rate of the t-1th iteration, and κ is a preset adjustment sensitivity coefficient.

[0042] Furthermore, the construction and training process of the S3 feature dimensionality reduction model is as follows:

[0043] S301 initializes the weights of the encoder and decoder;

[0044] S302 performs manifold structure sampling from the high-dimensional data obtained after feature extraction to estimate the local manifold structure of the data;

[0045] S303 generates a random mask for each input sample. For each input sample, that is, the data x after feature extraction i , generate a random mask m i It is expressed as:

[0046] m i = Bernoulli(p)

[0047] In the formula, mi is applied to x i A mask of , where each element independently takes the value 0 or 1; Bernoulli(p) represents the Bernoulli distribution, and p is the probability that the mask element is 1;

[0048] Bernoulli random variable m ij The generation method is expressed as:

[0049]

[0050] Where U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating m ij =1, if U is less than p, then m ij is set to 1; if U is greater than or equal to p, then m ij is set to 0;

[0051] S304 After the input data is masked, it is compressed by the encoder to obtain a low-dimensional potential representation z i The dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the input data, and the skewness S(x i ) and kurtosis K(x i ), expressed as:

[0052]

[0053] In the formula, μ pd and σ pd x i The mean and standard deviation of

[0054] Calculate the adjustment coefficient β csd :

[0055]

[0056] Where tanh() is the hyperbolic tangent function;

[0057] S305 uses manifold constraints to adjust the low-dimensional representation in the latent space, reconstructs the input data from the low-dimensional representation through the decoder, calculates the reconstruction error and the manifold constraint error, and updates the network parameters through back propagation. The total loss function Lur consists of the reconstruction error and the manifold constraint error, which is expressed as:

[0058]

[0059] In the formula, is the reconstructed data point; kc is the regularization parameter of the manifold constraint;

[0060] S306 uses the gradient descent method to optimize the network parameters and continuously reduce the total loss;

[0061] The preset maximum number of iterations in S307 is set to 1000 times, and the above steps are repeated until the preset stop iteration condition is met, and the model training is completed.

[0062] Furthermore, to ensure that the geometric and topological characteristics of the data are preserved, the calculation method of the manifold constraint loss is expressed as:

[0063]

[0064] In the formula, z i and z j They are x i and x j Representation in low-dimensional space, L manifold is the manifold constraint loss.

[0065] Furthermore, the classifier model construction training method described in S4 is as follows:

[0066] S401 initializes the basic parameters of the support vector machine. The initialization setting method of the kernel function parameter σ and the penalty coefficient C is expressed as:

[0067]

[0068] In the formula, x i Represents the input feature vector, y i represents the corresponding label, μ q is the average value of the feature vector, N is the number of samples, and w q and b q are the weight and bias of the support vector machine respectively;

[0069] S402: Let the state space be the feature vector distribution space output from the data dimension reduction model training step, defined as:

[0070] s q,t ={x t ,a q,t-1 ,r q,t-1}

[0071] The action space is defined as the adjustment of the classification decision boundary; the reward function is based on the classification accuracy, that is, the proportion of samples correctly classified, and the reward function r q,t is defined as:

[0072]

[0073] In the formula, s q,t Represents the state at time t, including the current feature vector x t , the action a in the previous stepq,t-1 , and the reward r of the previous step q,t-1 , r q,t It is a reward function calculated based on the current weight and bias. sgn() is a sign function. If the parameter is greater than 0, the sgn function returns 1; if the parameter is equal to 0, it returns 0; if the parameter is less than 0, it returns -1;

[0074] S403 uses the feature vector obtained from the feature dimensionality reduction model as input to perform classification training for the initial batch, collects preliminary classification effect data, selects the optimal action through the reinforcement learning algorithm based on the preliminary training results, and adjusts the decision boundary of the support vector machine, which is expressed as:

[0075]

[0076] Δr q,t =r q,t -r q,t-1

[0077] In the formula, λ q,t is the learning rate at time t, λ q,0 is the initial learning rate, κ q is the adjustment sensitivity coefficient, Δr q,t is the reward change;

[0078] S404 repeats the above steps until the preset stop iteration condition is met, which means that the model training is completed.

[0079] Furthermore, the gradient of the loss function with respect to the weights The calculation method is expressed as:

[0080]

[0081] In the formula, max(0,1-y i (w q,t ·x i +b q,t )) is used to calculate the contribution of each data point to the total loss. If the sample is correctly classified, this part of the contribution is 0; otherwise, it is added to the total loss;

[0082] Sensitivity coefficient κ q Controls the sensitivity of learning rate adjustment, and the calculation method is expressed as:

[0083]

[0084] In the formula, is the reward r for the past few iterations q,t The standard deviation of q This is linked to the instability of reward changes, making the adjustment of learning rate more refined and improving responsiveness.

[0085] Beneficial effects:

[0086] 1. The present invention enhances the model's resistance to environmental noise and abnormal data through feature dimensionality reduction and mask transformation mechanisms, and enhances the model's adaptability to extreme environments and weather as well as congested road conditions during intelligent driving.

[0087] 2. Through hierarchical reinforcement learning and dynamic adjustment of learning rate, the present invention enables the model to adapt to new data features and environmental changes more quickly, and respond to and identify emergencies during driving more quickly.

[0088] 3. The present invention constructs a feature extraction model through a twin gray wolf optimized convolutional neural network algorithm, which improves the efficiency and convergence speed of model training, and further accelerates the efficiency of obstacle identification and feedback. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a specific flow chart of the method of the present invention;

[0090] Figure 2 The present invention is a training flow chart of the auto-encoding neural network algorithm based on manifold learning; DETAILED DESCRIPTION

[0091] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0092] like Figure 1-2 As shown, this embodiment discloses a method for intelligent driving obstacle target detection based on millimeter wave radar, and the steps are as follows:

[0093] S1. Data collection and annotation

[0094] The training data used for model training in the present invention is collected from the detection signals obtained by the multi-source millimeter-wave radar, which is derived from the vehicle-mounted radar system. The vehicle-mounted radar system continuously obtains the electromagnetic wave reflection signals of the surrounding environment during driving through the radar sensor deployed on the smart driving car. The millimeter-wave radar detection signal covers the electromagnetic wave signals reflected from different directions, and each signal vector represents the radar detection data within a certain period of time.

[0095] The detection signal data obtained by the multi-source millimeter wave radar is labeled. In one embodiment, the labeled categories include: vehicles, pedestrians, obstacles, a total of 3 categories.

[0096] S2. Build feature extraction model training

[0097] The collected data is input into the feature extraction model to train the feature extraction model. The present invention uses a convolutional neural network algorithm based on twin gray wolf optimization for feature extraction. The convolutional neural network algorithm based on twin gray wolf optimization consists of two parallel convolutional neural network branches that share the same weights and parameters to process two different but related input signals, thereby enhancing the model's ability to capture millimeter-wave radar data features. The present invention uses the gray wolf optimization algorithm to optimize network parameters and find the optimal solution through group cooperation and information sharing.

[0098] Specifically, the training process of the convolutional neural network algorithm based on twin gray wolf optimization is as follows:

[0099] S201, construct two convolutional neural network branches, each of which has the same layer structure, including convolutional layer, pooling layer and fully connected layer, and initialize network parameters. Specifically, let W (0) Represents the weight set of the network in the initial state, and the initialization method is random initialization.

[0100] S202, the collected radar signal is divided into two parts, the first part is the normal radar signal, and the second part is the radar signal after Fourier transformation, which are respectively input into two branches of the twin network and preprocessed, which can be expressed as:

[0101]

[0102] In the formula, X is the input data, μ and σ are the mean and standard deviation of the data respectively, and X norm is the normalized data.

[0103] S203, each branch processes its input data and extracts features through convolutional layers and pooling layers. The way in which the convolution operation extracts features can be expressed as:

[0104] F=Re(W*X norm +b)

[0105] Where * represents the convolution operation, W and b are the weight and bias of the convolution layer, Re() is the ReLU activation function, and F represents the extracted features.

[0106] The weights W and bias b of the convolutional layer are updated using the mini-batch gradient descent method, which can be expressed as:

[0107]

[0108] Where η cf is the learning rate, m is the batch size, x i and i is the i-th sample in the batch and its label, is the partial derivative of the loss function with respect to the weight, W new is the updated weight, L() is the loss function, L(W old , x i ,y i ) is the parameter W old , x i ,y i The loss function when W old is the weight before updating. Preferably, η cf Set to 0.01.

[0109] The update method of bias b can be expressed as:

[0110]

[0111] In the formula, is the partial derivative of the loss function with respect to the bias, b new is the updated bias, b old is the bias before updating.

[0112] S204, the features output by the two branches are fused, and the cosine similarity is used as the fused feature, which can be expressed as:

[0113]

[0114] Where F1 and F2 are the features extracted from the two branches of the dual-branch network, || || is the L2 norm, and S is the cosine similarity between features.

[0115] S205. Use the gray wolf algorithm to optimize the parameters of the convolutional neural network, simulate the social hierarchy and hunting strategy of the gray wolf, and update the network weights through group wisdom to find the global minimum of the loss function. Taking the weights as an example, the parameter update method can be expressed as:

[0116] W new ′=W alpha -α dq |β dq W prey -W new |

[0117] α dq =2a s rand(0,1)-a s

[0118] β dq =2a s rand(0,1)

[0119] Where W new ′ is the updated weight; W alpha is the current optimal solution, that is, the optimal parameter in the current iteration process; W prey is the target solution, i.e. the optimal parameter; W new is the current solution, α dq and β dq is the control parameter, a s is the attenuation parameter, and rand(0,1) is a random number in the interval [0,1]. In this embodiment, a s Set to 0.95.

[0120] S206, back propagation is performed according to the loss function to update the network parameters, which can be expressed as:

[0121]

[0122] Where W new ' is the weight after optimization by the gray wolf algorithm, λ xa is the learning rate, is the gradient of the loss function with respect to the weights, W next is the weight for the next iteration. Preferably, λ xa The initial value of is set to 0.01.

[0123] In this embodiment, the loss function is a cross entropy loss function.

[0124] Learning rate λ xa To set it in a dynamic adjustment mode, it can be expressed as:

[0125]

[0126] In the formula, is the learning rate of the tth iteration, exp() is an exponential function based on the natural base e, is the learning rate of the t-1th iteration, and κ is a preset adjustment sensitivity coefficient. In this embodiment, κ is preferably set to 0.3.

[0127] S207, repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. In this embodiment, the preset stop iteration condition is to reach a preset maximum number of iterations. In this embodiment, preferably, the preset maximum number of iterations is set to 1000 times.

[0128] S3. Build feature dimensionality reduction model training

[0129] The data after feature extraction is input into the feature dimensionality reduction model to train the feature dimensionality reduction model. The present invention adopts an autoencoder neural network algorithm based on manifold learning to perform feature dimensionality reduction. The autoencoder neural network based on manifold learning includes two parts: an encoder and a decoder. The function of the encoder is to map the high-dimensional input features to a low-dimensional latent space, and the decoder attempts to reconstruct the original input from this low-dimensional latent space. The present invention adopts the manifold learning principle to enable the network to better understand and retain the intrinsic structure of the data, and at the same time adopts a mask transformation mechanism to enhance the robustness of the model to noise and outliers in the input data.

[0130] Specifically, the training process of the autoencoder neural network algorithm based on manifold learning is as follows:

[0131] S301, initialize the weights of the encoder and decoder. In this embodiment, the initialization method is random initialization, which can be expressed as:

[0132]

[0133] Where W e and W d are the weight matrices of the encoder and decoder respectively; b e and b d is the corresponding bias vector; σ 2 Represents the initialization standard deviation. In this embodiment, preferably, σ 2 Set to 0.01.

[0134] S302: Sampling the manifold structure from the high-dimensional data obtained after feature extraction to estimate the local manifold structure of the data. In this embodiment, this is achieved by the k-nearest neighbor algorithm, that is, using the k-nearest neighbor method to sample each data point x. i Sampling is performed and the average distance of its k nearest neighbors is calculated, which can be expressed as:

[0135]

[0136] In the formula, x i Represents the current data point; NN k (x i ) is x i The k nearest neighbors of ; ∥∥ is the L2 norm, which follows the calculation method of Euclidean distance; ∥x i -x j ∥ 2 is x i and x j The square of the Euclidean distance between them.

[0137] S303: Generate a random mask for each input sample, where the mask is used to simulate the partial loss of input data. Specifically, for each input sample (data after feature extraction) x i , generate a random mask m i can be expressed as:

[0138] m i = Bernoulli(p)

[0139] In the formula, m i is applied to x i The mask of , each element independently takes the value 0 or 1; Bernoulli(p) represents the Bernoulli distribution, p is the probability that the mask element is 1.

[0140] The Bernoulli distribution is a discrete distribution used to model a random variable with one of two possible outcomes. In this example, the Bernoulli random variable m ij The generation method can be expressed as:

[0141]

[0142] Where U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating m ij = 1. If U is less than p, then m ij is set to 1; if U is greater than or equal to p, then m ij is set to 0. In this embodiment, preferably, pr is set to 0.45.

[0143] S304: After the input data is masked, it is compressed by the encoder to obtain a low-dimensional potential representation z i , calculated as:

[0144] z i =Sig(W e ·(1+γ csd β csd )·(x i ⊙m i )+b e )

[0145] Where ⊙ represents the corresponding multiplication of elements; Sig() is the Sigmoid activation function; β csd is the dynamic coding adjustment coefficient; γ csd is a preset proportional factor. In this embodiment, preferably, γ csd Set to 0.03.

[0146] In this embodiment, the dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the input data. Specifically, the skewness S(x i ) and kurtosis K(x i ), which can be expressed as:

[0147]

[0148]

[0149] In the formula, μ pd and σ pd x i The mean and standard deviation of .

[0150] Adjustment coefficient β csd The calculation method can be expressed as:

[0151]

[0152] Where tanh() is the hyperbolic tangent function.

[0153] S305. In the latent space, the manifold constraint is used to adjust the low-dimensional representation to ensure that the geometric and topological characteristics of the data are preserved. The calculation method of the manifold constraint loss can be expressed as:

[0154]

[0155] In the formula, z i and z j They are x i and x j Representation in low-dimensional space, L manifold is the manifold constraint loss.

[0156] S306, reconstruct the input data from the low-dimensional representation through the decoder, calculate the reconstruction error and the manifold constraint error, and update the network parameters through back propagation. The total loss function Lur is composed of the reconstruction error and the manifold constraint error, which can be expressed as:

[0157]

[0158] In the formula, is the reconstructed data point; kc is the regularization parameter of the manifold constraint. In this embodiment, preferably, λ kc Set to 0.2.

[0159] S307, using the gradient descent method to optimize the network parameters and continuously reduce the total loss, the method of updating the parameters can be expressed as:

[0160]

[0161]

[0162] In the formula, α gk is the learning rate. In this embodiment, preferably, α gk Set to 0.04.

[0163] S308, repeat the above steps until the preset stop iteration condition is met, which means the model training is completed. In one embodiment, the preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations in this embodiment is set to 1000 times.

[0164] S4. Build classifier model training

[0165] The reduced-dimensional data is input into a classifier to train the classifier model. The present invention adopts a support vector machine classification algorithm based on hierarchical reinforcement learning. Aiming at the limitations of traditional support vector machines in dealing with nonlinear separable problems, the present invention improves the learning process by adopting a dynamic forgetting factor and dynamically adjusts the learning rate according to the importance of historical training data, so that the model can adapt to new data features more quickly, thereby enhancing the model's adaptability to time-varying data.

[0166] Specifically, the training process of the support vector machine classification algorithm based on hierarchical reinforcement learning is as follows:

[0167] S401, initializing the basic parameters of the support vector machine. Specifically, the initialization setting method of the kernel function parameter σ and the penalty coefficient C can be expressed as:

[0168]

[0169] In the formula, x i Represents the input feature vector, y i represents the corresponding label, μ q is the average value of the feature vector, N is the number of samples, and w q and b q are the weight and bias of the support vector machine respectively.

[0170] S402, let the state space be the feature vector distribution space output from the data dimension reduction model training step, defined as:

[0171] s q,t ={x t ,a q,t-1 ,r q,t-1}

[0172] The action space is defined as the adjustment of the classification decision boundary; the reward function is based on the classification accuracy, that is, the proportion of samples correctly classified, and the reward function r q,t is defined as:

[0173]

[0174] In the formula, s q,t Represents the state at time t, including the current feature vector x t , the action a in the previous step q,t-1 , and the reward r of the previous step q,t-1 . r q,t It is a reward function calculated based on the current weights and biases. sgn() is a sign function. If the parameter is greater than 0, the sgn function returns 1; if the parameter is equal to 0, it returns 0; if the parameter is less than 0, it returns -1.

[0175] S403, using the feature vector obtained from the feature dimensionality reduction model as input, perform classification training for the initial batch, collect preliminary classification effect data, select the optimal action through the reinforcement learning algorithm according to the preliminary training results, and adjust the decision boundary of the support vector machine, which can be expressed as:

[0176]

[0177] λ q,t =λ q,0 exp(-κ q Δr q,t )

[0178] Δr q,t =r q,t -r q,t-1

[0179] In the formula, λ q,t is the learning rate at time t, λ q,0 is the initial learning rate, κ q is the adjustment sensitivity coefficient, Δr q,t is the reward change.

[0180] The gradient of the loss function with respect to the weights The calculation method can be expressed as:

[0181]

[0182] In the formula, max(0,1-y i (w q,t ·x i +b q,t )) is used to calculate the contribution of each data point to the total loss. If the sample is correctly classified, this partial contribution is 0; otherwise, it is added to the total loss.

[0183] Sensitivity coefficient κ q Control the sensitivity of learning rate adjustment. The calculation method can be expressed as:

[0184]

[0185] In the formula, is the reward r for the past few iterations q,t By substituting κ q This, coupled with the instability of reward changes, allows for more subtle and responsive adjustments of the learning rate.

[0186] S404, repeat the above steps until the preset stop iteration condition is met, which means that the model training is completed. In one embodiment, the preset stop iteration condition is reaching a preset maximum number of iterations. In this embodiment, the preset maximum number of iterations is preferably set to 800 times.

[0187] S5. Obstacle recognition

[0188] The trained feature extraction model, data dimension reduction model and classifier model are used to process the newly acquired millimeter-wave radar detection signal and output the final obstacle recognition result.

[0189] The specific process is that the newly collected millimeter-wave radar detection signal is first input into the trained feature extraction model for feature extraction, then input into the feature dimension reduction model for feature dimension reduction, and finally input into the classifier for classification. The classification result is the category of the obstacle.

[0190] The above description of the embodiments enables professionals and technicians in the field to implement or use the present invention. Various modifications to the embodiments will be apparent to professionals and technicians. The general principles of the present invention can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention should not be limited to the embodiments shown herein, but should cover the widest range consistent with the principles and novel features disclosed in the present invention.

Claims

1. A method for intelligent driving obstacle target detection based on millimeter wave radar, characterized in that: The method comprises the following steps: S1 collects millimeter-wave radar detection signal data as a training set and labels the data; S2 builds a feature extraction model based on the twin gray wolf optimized convolutional neural network algorithm and inputs the training set into the feature extraction model for feature extraction training; S3 uses the autoencoder neural network algorithm based on manifold learning to build a feature dimensionality reduction model, and inputs the training set after feature extraction into the feature dimensionality reduction model for feature dimensionality reduction training; S4 builds a classifier model based on the support vector machine classification algorithm of hierarchical reinforcement learning, puts the reduced-dimensional training set into the classifier to train the classifier model, uses a dynamic forgetting factor to improve the learning process, and dynamically adjusts the learning rate according to the importance of historical training data; S5 uses the trained composite model to process the newly acquired millimeter-wave radar detection signal and outputs the final obstacle recognition result.

2. The intelligent driving obstacle target detection method based on millimeter wave radar according to claim 1 is characterized in that: The feature extraction model construction training process described in S2 is as follows: S201 constructs two convolutional neural network branches, each of which has the same layer structure, including convolutional layer, pooling layer and fully connected layer, and initializes network parameters. Specifically, set W (0) Represents the weight set of the network in the initial state, and the initialization method is random initialization; S202 divides the collected radar signal into two parts. The first part is the normal radar signal, and the second part is the radar signal after Fourier transformation. They are input into two branches of the twin network and preprocessed, which are expressed as: In the formula, X is the input data, μ and σ are the mean and standard deviation of the data respectively, and X norm is the normalized data; Each branch in S203 processes its input data and extracts features through convolutional layers and pooling layers. The way in which the convolution operation extracts features is expressed as follows: F=Re(W*X norm +b) In the formula, * represents the convolution operation, W and b are the weight and bias of the convolution layer respectively, Re() is the ReLU activation function, and F represents the extracted features; S204 fuses the features output by the two branches, and uses cosine similarity as the fused feature representation: Where F1 and F2 are the features extracted from the two branches of the two-branch network, ∥∥ is the L2 norm, and S is the cosine similarity between the features; S205 uses the gray wolf algorithm to optimize the weight parameters of the convolutional neural network, simulates the social hierarchy and hunting strategy of the gray wolf, and updates the network weights through group wisdom to find the global minimum of the loss function. The update method of the weight parameters is expressed as: W new ’=W alpha -α dq |β dq W prey -W new | α dq =2a s ·rand(0,1)-a s β dq =2a s ·rand(0,1) Where W new ' is the updated weight; W alpha is the current optimal solution, that is, the optimal parameter in the current iteration process; W prey is the target solution, i.e. the optimal parameter; W new is the current solution, α dq and β dq is the control parameter, a s is the attenuation parameter, rand(0,1) is a random number in the interval [0, 1]; S206 performs back propagation according to the loss function and updates the network parameters: Where W new ' is the weight after optimization by the gray wolf algorithm, λ xa is the learning rate, is the gradient of the loss function with respect to the weights, W next is the weight of the next iteration; S207 repeats the above steps until a preset stop iteration condition is met.

3. The intelligent driving obstacle target detection method based on millimeter wave radar according to claim 2 is characterized in that: The weight W and bias b of the convolutional layer are updated using the mini-batch gradient descent method: Where η cf is the learning rate, m is the batch size, x i and i is the i-th sample in the batch and its label, is the partial derivative of the loss function with respect to the weight, W new is the updated weight, L() is the loss function L(W old ,x i ,y i ), is the parameter W old ,x i ,y i The loss function when W old is the weight before updating; The update method of bias b is expressed as: In the formula, is the partial derivative of the loss function with respect to the bias, b new is the updated bias, b old is the bias before updating; Learning rate λ xa It is set for dynamic adjustment, expressed as: In the formula, is the learning rate of the tth iteration, exp() is an exponential function based on the natural base e, is the learning rate of the t-1th iteration, and κ is a preset adjustment sensitivity coefficient.

4. The intelligent driving obstacle target detection method based on millimeter wave radar according to claim 1 is characterized in that: The construction and training process of the S3 feature dimensionality reduction model is as follows: S301 initializes the weights of the encoder and decoder; S302 performs manifold structure sampling from the high-dimensional data obtained after feature extraction to estimate the local manifold structure of the data; S303 generates a random mask for each input sample. For each input sample, that is, the data x after feature extraction i , generate a random mask m i It is expressed as: m i =Bernoulli(p) In the formula, m i is applied to x i The mask of , each element independently takes the value 0 or 1; Bernoulli(p) represents the Bernoulli distribution, p is the probability that the mask element is 1; Bernoulli random variable m ij The generation method is expressed as: Where U is a random number drawn from the uniform distribution Uniform(0,1), pr is the preset threshold parameter, indicating m ij =1, if U is less than p, then m ij is set to 1; if U is greater than or equal to p, then m ij is set to 0; S304 After the input data is masked, it is compressed by the encoder to obtain a low-dimensional potential representation z i The dynamic coding adjustment coefficient is dynamically adjusted based on the statistical characteristics of the input data, and the skewness S(x i ) and kurtosis K(x i ), expressed as: In the formula, μ pd and σ pd x i The mean and standard deviation of Calculate the adjustment coefficient β csd : Where tanh() is the hyperbolic tangent function; S305 uses manifold constraints to adjust the low-dimensional representation in the latent space, reconstructs the input data from the low-dimensional representation through the decoder, calculates the reconstruction error and the manifold constraint error, and updates the network parameters through back propagation. The total loss function Lur consists of the reconstruction error and the manifold constraint error, which is expressed as: In the formula, is the reconstructed data point; kc is the regularization parameter of the manifold constraint; S306 uses the gradient descent method to optimize the network parameters and continuously reduce the total loss; The preset maximum number of iterations in S307 is set to 1000 times, and the above steps are repeated until the preset stop iteration condition is met, and the model training is completed.

5. The intelligent driving obstacle target detection method based on millimeter wave radar according to claim 4 is characterized in that: To ensure that the geometric and topological properties of the data are preserved, the calculation method of the manifold constraint loss is expressed as: In the formula, z i and z j They are x i and x j Representation in low-dimensional space, L manifold is the manifold constraint loss.

6. The intelligent driving obstacle target detection method based on millimeter wave radar according to claim 1 is characterized in that: The training method for constructing the classifier model described in S4 is as follows: S401 initializes the basic parameters of the support vector machine. The initialization setting method of the kernel function parameter σ and the penalty coefficient C is expressed as: In the formula, x i Represents the input feature vector, y i represents the corresponding label, μ q is the average value of the feature vector, N is the number of samples, and w q and b q are the weight and bias of the support vector machine respectively; S402: Let the state space be the feature vector distribution space output from the data dimension reduction model training step, defined as: s q,t ={x t ,a q,t-1 ,r q,t-1 } The action space is defined as the adjustment of the classification decision boundary; the reward function is based on the classification accuracy, that is, the proportion of samples correctly classified, and the reward function r q,t is defined as: In the formula, s q,t Represents the state at time t, including the current feature vector x t , the action a in the previous step q,t-1 , and the reward r of the previous step q,t-1 , r q,t It is a reward function calculated based on the current weight and bias. sgn() is a sign function. If the parameter is greater than 0, the sgn function returns 1. If the parameter is equal to 0, return 0; If the argument is less than 0, it returns -1; S403 uses the feature vector obtained from the feature dimensionality reduction model as input to perform classification training for the initial batch, collects preliminary classification effect data, selects the optimal action through the reinforcement learning algorithm based on the preliminary training results, and adjusts the decision boundary of the support vector machine, which is expressed as: l q,t =λ q,0 exp(-k q Δr q,t ) Δr q,t =r q,t -r q,t-1 In the formula, λ q,t is the learning rate at time t, λ q,0 is the initial learning rate, κ q is the adjustment sensitivity coefficient, Δr q,t is the reward change; S404 repeats the above steps until the preset stop iteration condition is met, which means that the model training is completed.

7. The intelligent driving obstacle target detection method based on millimeter wave radar according to claim 6 is characterized in that: The gradient of the loss function with respect to the weights The calculation method is expressed as: In the formula, max(0,1-y i (w q,t ·x i +b q,t )) is used to calculate the contribution of each data point to the total loss. If the sample is correctly classified, this part of the contribution is 0; Otherwise, it is added to the total loss; Sensitivity coefficient κ q Controls the sensitivity of learning rate adjustment, and the calculation method is expressed as: In the formula, is the reward r for the past few iterations q,t The standard deviation of q This is linked to the instability of reward changes, making the adjustment of learning rate more refined and improving responsiveness.

Images