Millimeter wave radar plot fusion method based on deep learning, electronic equipment and storage medium

Abstract

The invention discloses a millimeter wave radar plot fusion method based on deep learning, an electronic device and a storage medium, frequency domain prior model features are added in a training process to assist training, and time-frequency feature combination is used as a classification basis. In order to screen out parameterization features suitable for auxiliary network training, a parameterization feature selection method is designed before auxiliary training, relatively optimal model parameterization features are screened out, and then target classification detection fusing time-frequency features is completed. And designing a target parameterized feature importance calculation method by using a back propagation algorithm and feature splicing operation. According to the method, on the premise that main indexes such as radar power, precision and resolution are guaranteed, algorithm complexity is reduced, algorithm efficiency is greatly improved, and application in practical engineering is facilitated.

Description

Technical Field

[0001] The present invention relates to a millimeter-wave radar point trace fusion method, electronic equipment, and storage medium based on deep learning, and belongs to the field of radar measurement technology. Background Art

[0002] Millimeter-wave radar has demonstrated significant value in many fields, capable of acquiring key information such as target distance, speed, and angle. However, the data from a single millimeter-wave radar often has limitations. To obtain a more comprehensive and accurate picture of target status, it is crucial to fuse the data from multiple millimeter-wave radars to obtain comprehensive situational information.

[0003] Traditional point trace fusion methods include weighted average fusion and Kalman filter-based fusion.

[0004] Weighted average method: The points from different sources are assigned corresponding weights and then averaged to fuse them. It is simple and easy to operate, but it is difficult to accurately characterize nonlinear relationships in complex environments.

[0005] Fusion based on Kalman filtering: It is often used to process state estimation and fusion in dynamic systems. However, it relies on accurate system model assumptions. When the actual situation deviates greatly from the model, the fusion effect will be greatly reduced.

[0006] In addition, the integration of deep learning technology brings new development opportunities for millimeter-wave radar point trace fusion, overcoming the errors and uncertainties in radar measurement using traditional point trace fusion methods.

[0007] However, when faced with complex and changeable actual scenarios, high noise interference, and multi-source heterogeneous data, traditional methods often extract model parameters by modeling frequency domain echoes as the basis for classification. The artificially added prior information often cannot optimize the final classification problem, and it is difficult to adaptively extract effective fusion features, resulting in unsatisfactory fusion results.

[0008] Deep learning technology only automatically extracts the features of the target's time-domain high-resolution range image based on deep CNN and then classifies it. However, due to the random initialization of network weights, the network often needs to conduct trial and error to find the optimal solution, and the training process consumes a lot of time and computing resources.

[0009] Therefore, those skilled in the art urgently need to improve the existing wave radar point trace fusion method. Summary of the Invention

[0010] Objective: To overcome the deficiencies in the prior art, the present invention provides a millimeter-wave radar point trace fusion method, electronic device, and storage medium based on deep learning.

[0011] Technical solution: To solve the above technical problems, the technical solution adopted by the present invention is:

[0012] First, a millimeter-wave radar trace fusion method based on deep learning, specifically including:

[0013] Step 1: Preprocess the received distributed radar trace data and the main radar trace data to obtain the preprocessed trace data as a training set.

[0014] Step 2: Build a deep learning model for point-trace fusion.

[0015] Step 3: Use the training set for forward propagation training of the deep learning model of point trace fusion to obtain a trained deep learning model of point trace fusion.

[0016] Step 4: Use the back propagation algorithm to update the model parameters of the trained point-trace fusion deep learning model to obtain the optimized point-trace fusion deep learning model.

[0017] Step 5: Input the point trace data to be fused into the optimized deep learning model of point trace fusion to obtain the fused point trace data.

[0018] As a preferred solution, the step 1 specifically includes:

[0019] Step 1.1: Filter the received distributed radar trace data and the main radar trace data, remove abnormal traces, and obtain cleaned distributed radar trace data and cleaned main radar trace data.

[0020] Step 1.2: The cleaned distributed radar point trace data is subjected to coordinate transformation using a transformation formula to obtain transformed distributed radar point trace data.

[0021] Step 1.3: Synchronize the converted distributed radar trace data with the system time to obtain the converted and synchronized distributed radar trace data.

[0022] Step 1.4: Extract the eigenvalues of the converted and synchronized distributed radar trace data and the cleaned master radar trace data, normalize the eigenvalues, and obtain the normalized data as the training set.

[0023] As a preferred solution, the conversion formula is as follows:

[0024]

[0025] in, Indicates the coordinate of the x-axis in the reference coordinate system, Indicates the coordinate of the y-axis in the reference coordinate system, Indicates the coordinate of the z-axis in the reference coordinate system, Indicates the longitude in the spatial geodetic coordinate system, Indicates the latitude in the spatial geodetic coordinate system, It represents the altitude in the spatial geodetic coordinate system, e is the eccentricity of the ellipse, .

[0026] Where, , Represents the Earth's semi-major axis.

[0027] Where, , Represents the Earth's semi-minor axis.

[0028] As a preferred solution, the conversion formula is as follows:

[0029]

[0030] in, Indicates the coordinate of the x-axis in the reference coordinate system, Indicates the coordinate of the y-axis in the reference coordinate system, Indicates the coordinate of the z-axis in the reference coordinate system, Indicates the longitude in the spatial geodetic coordinate system, Indicates the latitude in the spatial geodetic coordinate system, Indicates the altitude in the spatial geodetic coordinate system, Represents the radius of the Earth.

[0031] As a preferred solution, the deep learning model of point trace fusion includes: a CNN layer, an LSTM unit and a fully connected unit, wherein the output end of the CNN layer is connected to the input end of the LSTM unit, the output end of the CNN layer is also connected to the first input end of the fully connected unit, and the output end of the LSTM unit is also connected to the second input end of the fully connected unit.

[0032] The CNN layer includes: 4 CP blocks and a regularization layer (dropout).

[0033] The LSTM unit includes: 6 LSTM layers connected in sequence.

[0034] The fully connected unit includes: two fully connected layers connected in sequence.

[0035] As a preferred solution, the four CP blocks are CP 1, CP 2, CP 3, and CP 4, each of which includes a first convolutional layer, a second convolutional layer, and a pooling layer connected in sequence. The convolution kernels of the first and second convolutional layers are set to a size of 3x3 and a stride of 1.

[0036] The output feature map sizes of CP 1, CP 2, CP 3 and CP4 are 32, 32, 64, 64 respectively.

[0037] The two fully connected layers are set with 512 and 1024 output neurons.

[0038] As a preferred solution, the step 4 specifically includes:

[0039] Step 4.1: Set the feature stitching layer between CP 3 and CP4.

[0040] Step 4.2: Extract the frequency domain features of the target echo respectively.

[0041] Step 4.3: Splice the frequency domain features at the end of the hidden layer feature map of each channel output by CP3 to form a feature map of concatenated features.

[0042] Step 4.3: Input the feature map of the concatenated features into CP4, complete the training of the backpropagation algorithm of the trained point-trace fusion deep learning model, and obtain the cross-entropy loss value.

[0043] Step 4.5: According to the cross-entropy loss value, obtain the updated model parameters and obtain the optimized point-trace fusion deep learning model.

[0044] As a preferred solution, the cross-entropy loss value expression is as follows:

[0045] loss =− [ y ln( y ^ ) + ( 1 − y )ln( 1 − y ^ )]

[0046] in: is the true label (0, 1), is the probability of predicting 1.

[0047] In a second aspect, a computer-readable storage medium stores a computer program thereon. When the computer program is executed by a processor, it implements a millimeter-wave radar point fusion method based on deep learning as described in any one of the first aspects.

[0048] According to a third aspect, a computer device includes:

[0049] a memory for storing instructions;

[0050] A processor is configured to execute the instructions so that the computer device performs the operations of the millimeter-wave radar point trace fusion method based on deep learning as described in any one of the first aspects.

[0051] Beneficial effects: The present invention provides a millimeter-wave radar trace fusion method, electronic device and storage medium based on deep learning. The present invention adds frequency domain prior model features to assist training during the training process, and realizes the combination of time-frequency features as the basis for classification. In order to screen out parameterized features suitable for auxiliary network training, a parameterized feature selection method is designed before auxiliary training to screen out the relatively optimal model parameterized features, and then complete the target classification detection that integrates time-frequency features. A target parameterized feature importance calculation method is designed using the back propagation algorithm and feature splicing operations. Under the premise of ensuring the main indicators such as radar power, accuracy, and resolution, the present invention reduces the algorithm complexity and greatly improves the algorithm efficiency, which is conducive to application in actual engineering. Compared with the existing technology, its advantages are as follows:

[0052] 1. The present invention improves the positioning and direction-finding accuracy of the millimeter-wave radar system.

[0053] 2. The present invention expands the spatial and temporal coverage of the millimeter-wave radar system.

[0054] 3. The present invention increases the accuracy of point trace correlation of the millimeter wave radar system and reduces ambiguity.

[0055] 4. The present invention improves the detection probability of the millimeter wave radar system for target detection.

[0056] 5. The present invention enhances the millimeter wave radar system's ability to continuously track targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of the millimeter-wave radar point trace fusion method based on deep learning.

[0058] Figure 2 A diagram of millimeter-wave radar data collection.

[0059] Figure 3 This is a structural diagram of the millimeter-wave radar point trace fusion feature extraction model based on deep learning.

[0060] Figure 4 This is the training flowchart of the millimeter-wave radar point trace fusion feature extraction model based on deep learning.

[0061] Figure 5 Schematic diagram of recall rate verification for the millimeter-wave radar point trace fusion feature extraction model training based on deep learning.

[0062] Figure 6 This is a graph verifying the correlation accuracy of the deep learning-based millimeter-wave radar point trace fusion feature extraction model training.

[0063] Figure 7Schematic diagram of the loss function for training the millimeter-wave radar point trace fusion feature extraction model based on deep learning.

[0064] Figure 8 Schematic diagram of the mean square error between the fused traces and the true traces.

[0065] Figure 9 Schematic diagram of the mean square error of the fusion trace and the real trace velocity.

[0066] Figure 10 Schematic diagram of the target trajectory detected by the main radar before point-track fusion.

[0067] Figure 11 This is a schematic diagram of the target trajectory detected by radar before point fusion.

[0068] Figure 12 Click the trajectory diagram after fusion of millimeter wave radar DETAILED DESCRIPTION

[0069] The following is a clear and complete description of the technical solutions in the examples of the present invention, in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative work are within the scope of protection of the present invention.

[0070] The present invention will be further described below with reference to specific embodiments.

[0071] Example 1:

[0072] This embodiment introduces a millimeter-wave radar point trace fusion method based on deep learning. The millimeter-wave radar network is composed of multiple millimeter-wave radars, among which one millimeter-wave radar is arbitrarily set as the main radar according to the actual scenario, and the remaining millimeter-wave radars are distributed millimeter-wave radars. The distributed millimeter-wave radars transmit their respective detection data to the data fusion center of the main radar. The data processor of the data fusion center fuses and outputs the point traces through the deep learning model of CNN-LSTM, and displays it through the control terminal. Figure 1 The laws include:

[0073] Step 1: Preprocess the received distributed radar trace data and the main radar trace data to obtain preprocessed trace data, and divide the preprocessed trace data into training set, validation set and test set. Specifically including:

[0074] Step 1.1: The data fusion center filters the collected trace data of the distributed millimeter-wave radar and the main radar to remove abnormal traces caused by noise and interference, and obtains cleaned trace data to ensure data quality.

[0075] like Figure 2 As shown in the figure, step 1.2: Set the spatial rectangular coordinate system as the reference coordinate system with the main radar as the coordinate origin. Use the conversion formula from the spatial geodetic coordinate system to the reference coordinate system to convert the distributed millimeter-wave radar's track data to the reference coordinate system. Set the system time to synchronize the clock of the distributed millimeter-wave radar's track data with the clock of the main radar's track data. Obtain the converted and synchronized distributed millimeter-wave radar's track data to ensure that the tracks used in each track fusion and track processing belong to the same period.

[0076] The conversion formula from the spatial geodetic coordinate system to the reference coordinate system is as follows:

[0077]

[0078] in, Indicates the coordinate of the x-axis in the reference coordinate system, Indicates the coordinate of the y-axis in the reference coordinate system, Indicates the coordinate of the z-axis in the reference coordinate system, Indicates the longitude in the spatial geodetic coordinate system, Indicates the latitude in the spatial geodetic coordinate system, It represents the altitude in the spatial geodetic coordinate system, e is the eccentricity of the ellipse, .

[0079] Where, , represents the Earth's semi-major axis, Represents the Earth's semi-minor axis.

[0080] Where, .

[0081] If the eccentricity is not considered, the conversion formula is simplified to:

[0082]

[0083] in, Represents the radius of the Earth.

[0084] Step 1.3: Obtain the eigenvalues of the converted and synchronized distributed millimeter-wave radar trace data and the master radar trace data. Normalize the eigenvalues (such as distance and amplitude) to obtain normalized data. This ensures that the data distribution is more conducive to deep learning model training and processing. Divide the normalized data into training, validation, and test sets to prepare for subsequent model training and evaluation.

[0085] Step 2: Build a deep learning model for point trace fusion, using an architecture that combines a convolutional neural network (CNN) with a long short-term memory (LSTM) network. The CNN layer extracts multi-dimensional features of the point trace data, including time, frequency, and spatial domains. The LSTM unit learns the correlation between the features of the point trace data across these domains. The model also includes fully connected units to integrate and map the extracted features, ultimately outputting the fused point trace information. Specifically, this includes:

[0086] like Figure 3 As shown, the deep learning model of point trace fusion includes: a CNN layer, an LSTM unit and a fully connected unit, wherein the output end of the CNN layer is connected to the input end of the LSTM unit, the output end of the CNN layer is also connected to the first input end of the fully connected unit, and the output end of the LSTM unit is also connected to the second input end of the fully connected unit.

[0087] The CNN layer includes: 4 CP blocks and a regularization layer (dropout).

[0088] The four CP blocks are CP 1, CP 2, CP 3, and CP 4. CP 1, CP 2, CP 3, and CP 4 each include a first convolutional layer, a second convolutional layer, and a pooling layer connected in sequence. The convolution kernels of the first and second convolutional layers are set to 3x3 in size and have a stride of 1.

[0089] The output feature map sizes of CP 1, CP 2, CP 3 and CP4 are 32, 32, 64, 64 respectively.

[0090] The LSTM unit includes: 6 LSTM layers connected in sequence.

[0091] The fully connected unit includes two fully connected layers connected in sequence, wherein the two fully connected layers are provided with 512 and 1024 output neurons respectively.

[0092] Step 3: Use the training set for forward propagation training of the deep learning model of point trace fusion to obtain a trained deep learning model of point trace fusion.

[0093] Step 4: Use the back propagation algorithm to update the model parameters of the trained point-trace fusion deep learning model to obtain the optimized point-trace fusion deep learning model.

[0094] Furthermore, the step 4 specifically includes:

[0095] Step 4.1: Set the feature stitching layer between CP 3 and CP4.

[0096] Step 4.2: Extract the frequency domain features of the target echo respectively.

[0097] Step 4.3: Splice the frequency domain features at the end of the hidden layer feature map of each channel output by CP3 to form a feature map of concatenated features.

[0098] Step 4.3: Input the feature map of the concatenated features into CP4, complete the training of the backpropagation algorithm of the trained point-trace fusion deep learning model, and obtain the cross-entropy loss value.

[0099] Step 4.5: According to the cross-entropy loss value, obtain the updated model parameters and obtain the optimized point-trace fusion deep learning model.

[0100] The calculation formula of the convolutional layer is as follows:

[0101]

[0102]

[0103] The calculation formula of the pooling layer is as follows:

[0104]

[0105] Among them: l represents the layer number, z represents the neurons of the current layer, δ represents the feature map error matrix of the current layer, represents the activation function derivative. The upsample function represents the upsampling operation, and ⊙ represents the Hadamard product.

[0106] When the pooling layer adopts maximum pooling, the current feature map neuron error corresponding to the original feature map area is set to 0 except for the current neuron error.

[0107] When average pooling is used, the error of each neuron position in the original feature map area is the current error divided by the area size.

[0108] Step 5: Input the point trace data to be fused into the optimized deep learning model of point trace fusion to obtain the fused point trace data.

[0109] Example 2:

[0110] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a millimeter-wave radar point trace fusion method based on deep learning as described in any one of Embodiment 1.

[0111] Example 3:

[0112] A computer device comprising:

[0113] Memory, used to store instructions.

[0114] A processor is used to execute the instructions so that the computer device performs the operations of a millimeter-wave radar point fusion method based on deep learning as described in any one of Example 1.

[0115] Example 4:

[0116] like Figure 4 As shown in the figure, the dataset contains 500 measured traces, 60% of which are used for training, 20% for validation, and 20% for testing. The initial learning rate is 0.001, and after every 10 training rounds, the learning rate is multiplied by a decay factor of 0.9. To avoid overfitting, the loss value is set to around 0.03.

[0117] After each training cycle, the validation set data is used to evaluate the performance of the deep learning model using point trace fusion. The recall rate and association accuracy are calculated. If the validation set performance of the deep learning model using point trace fusion does not improve over several consecutive cycles, training is terminated early and the optimal model parameters are saved.

[0118] Among them, Figure 5 As shown in Figure 2, recall rate is the ratio of the number of targets that can be successfully detected by the deep learning model of statistical point trace fusion to the actual number of targets.

[0119] like Figure 6 As shown in Figure 2, the association accuracy rate is the ratio of correctly associated points of the same target from different millimeter-wave radars.

[0120] Use the classic cross entropy loss function. Figure 7 As shown, the loss function is calculated as follows: loss =− [ y ln( y ^ ) + ( 1 − y )ln( 1 − y ^ )] .

[0121] in is the true label (0, 1), is the probability of predicting 1.

[0122] If a point is actually a target point ( ), the model predicts the probability of it being the target point , then the loss is − [ 1 × ln ( 0 . 8 ) + ( 1 − 1 ) × ln ( 1 − 0 . 8 ) ] =− ln ( 0 . 8 ) .

[0123] After multiple iterations, such as Figure 8-9 As shown in the figure, the mean square error of the distance and velocity is about 1m, and the mean square error of the velocity is 0.05m / s. The recall rate and association accuracy also reach about 98%.

[0124] The trained model was batch processed using test data, and the fusion results output by the model were compared with the actual target information corresponding to the test data. The point track distance, direction, speed, recall rate, and association accuracy all met the required indicators. The deep learning-based millimeter-wave radar point track fusion model has a significant effect on the point track fusion of millimeter-wave radar networks, and the continuous tracking capability of the target has been enhanced.

[0125] Two millimeter-wave radars were deployed on the east and west sides of a lake. The deep learning model was deployed in the point-track fusion system of the main millimeter-wave radar. The radars were simultaneously turned on to detect and match the target's motion trajectory. During the system operation, data was collected and fused in real time to verify the model's real-time performance, stability, and adaptability to dynamic environmental changes. Figure 10-12 As shown, the point traces before fusion are compared with the true value:

[0126] Main radar:

[0127] Document number Distance error Angular error Speed error Tracking points Extrapolated traces Probability of discovery 1 2.53m 0.36° 0.273m / s 126 50 71.6% 2 2.46m 0.43° 0.352m / s 118 57 67.4%

[0128] Radar:

[0129] Document number Distance error Angular error Speed error Tracking points Extrapolated traces Probability of discovery 1 1.57m 0.25° 0.162m / s 87 9 90.6% 2 1.74m 0.23° 0.248m / s 84 8 91.3%

[0130] Compare the point fusion with the true value:

[0131] Document number Distance error Angular error Speed error Tracking points Extrapolated traces Probability of discovery 1 0.92m 0.15° 0.041m / s 178 3 98.3% 2 0.89m 0.13° 0.037m / s 173 2 98.8%

[0132] Deploying a deep learning model to a millimeter-wave radar track fusion system can significantly improve the system's positioning and direction-finding accuracy, producing complete and stable track outputs. This overcomes the errors and uncertainties inherent in radar measurements using traditional track fusion methods, enabling it to better handle a variety of tasks in complex scenarios.

[0133] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A millimeter-wave radar point trace fusion method based on deep learning, characterized by: Specifically include: Step 1: Preprocess the received distributed radar trace data and the main radar trace data to obtain the preprocessed trace data as a training set; Step 2: Build a deep learning model for point-trace fusion; Step 3: Use the training set for forward propagation training of the deep learning model of point trace fusion to obtain a trained deep learning model of point trace fusion; Step 4: Use the back propagation algorithm to update the model parameters of the trained point trace fusion deep learning model to obtain the optimized point trace fusion deep learning model; Step 5: Input the point trace data to be fused into the optimized deep learning model of point trace fusion to obtain the fused point trace data.

2. The millimeter-wave radar point trace fusion method based on deep learning according to claim 1, characterized in that: The step 1 specifically includes: Step 1.1: Filter the received distributed radar trace data and the main radar trace data, remove abnormal traces, and obtain cleaned distributed radar trace data and cleaned main radar trace data; Step 1.2: Use the conversion formula to perform coordinate conversion on the cleaned distributed radar point trace data to obtain the converted distributed radar point trace data; Step 1.3: Synchronize the converted distributed radar trace data with the system time to obtain the converted and synchronized distributed radar trace data; Step 1.4: Extract the eigenvalues of the converted and synchronized distributed radar trace data and the cleaned master radar trace data, normalize the eigenvalues, and obtain the normalized data as the training set.

3. The millimeter-wave radar point trace fusion method based on deep learning according to claim 2, characterized in that: The conversion formula is as follows: ； in, Indicates the coordinate of the x-axis in the reference coordinate system, Indicates the coordinate of the y-axis in the reference coordinate system, Indicates the coordinate of the z-axis in the reference coordinate system, Indicates the longitude in the spatial geodetic coordinate system, Indicates the latitude in the spatial geodetic coordinate system, It represents the altitude in the spatial geodetic coordinate system, e is the eccentricity of the ellipse, ; Where, , represents the Earth's semi-major axis; Where, , Represents the Earth's semi-minor axis.

4. The millimeter-wave radar point trace fusion method based on deep learning according to claim 2, characterized in that: The conversion formula is as follows: ； in, Indicates the coordinate of the x-axis in the reference coordinate system, Indicates the coordinate of the y-axis in the reference coordinate system, Indicates the coordinate of the z-axis in the reference coordinate system, Indicates the longitude in the spatial geodetic coordinate system, Indicates the latitude in the spatial geodetic coordinate system, Indicates the altitude in the spatial geodetic coordinate system, Represents the radius of the Earth.

5. The millimeter-wave radar point trace fusion method based on deep learning according to claim 1, characterized in that: The deep learning model for point trace fusion includes: a CNN layer, an LSTM unit and a fully connected unit, wherein the output end of the CNN layer is connected to the input end of the LSTM unit, the output end of the CNN layer is also connected to the first input end of the fully connected unit, and the output end of the LSTM unit is also connected to the second input end of the fully connected unit; The CNN layer includes: 4 CP blocks and a regularization layer; The LSTM unit includes: 6 LSTM layers connected in sequence; The fully connected unit includes: two fully connected layers connected in sequence.

6. The millimeter-wave radar point trace fusion method based on deep learning according to claim 5, characterized in that: The four CP blocks are CP 1, CP 2, CP 3, and CP4. CP 1, CP 2, CP 3, and CP4 all include: a first convolutional layer, a second convolutional layer, and a pooling layer connected in sequence; the convolution kernels of the first and second convolutional layers are set to 3x3 in size and the stride is set to 1; the output feature map sizes of CP 1, CP 2, CP 3, and CP4 are 32, 32, 64, and 64, respectively; and the two fully connected layers are set with 512 and 1024 output neurons.

7. The millimeter-wave radar point trace fusion method based on deep learning according to claim 6, characterized in that: The step 4 specifically includes: Step 4.1: Set the feature stitching layer between CP 3 and CP4; Step 4.2: Extract the frequency domain features of the target echo respectively; Step 4.3: Splice the frequency domain features at the end of the hidden layer feature map of each channel output by CP3 to form a feature map of concatenated features; Step 4.3: Input the feature map of the concatenated features into CP4, complete the backpropagation algorithm training of the trained point-trace fusion deep learning model, and obtain the cross-entropy loss value; Step 4.5: According to the cross-entropy loss value, obtain the updated model parameters and obtain the optimized point-trace fusion deep learning model.

8. The millimeter-wave radar point trace fusion method based on deep learning according to claim 7, characterized in that: The cross entropy loss value expression is as follows: ； in: is the true label (0, 1), is the probability of predicting 1.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by a processor, a millimeter-wave radar point trace fusion method based on deep learning as described in any one of claims 1 to 8 is implemented.

10. A computer device, characterized in that: include: a memory for storing instructions; A processor, configured to execute the instructions so that the computer device performs the operations of the millimeter-wave radar point trace fusion method based on deep learning as described in any one of claims 1 to 8.

Images