Medium-resolution millimeter wave traffic radar road target optimization classification method

Abstract

The invention discloses a medium-resolution millimeter wave traffic radar road target optimization classification method, which comprises the following steps of S1, screening and preprocessing data acquired by a millimeter wave radar with medium distance or angle resolution to obtain M2D-MRRP; s2, data set construction is carried out on the effective to-be-classified target data, the effective to-be-classified target data are divided into a training set and a test set according to a certain proportion, and the training set contains training data and verification data; s3, building a target classification module based on the convolutional neural network, and adjusting the structure and parameters of the target classification module; and S4, performing target classification training by using the data set obtained in the above step and a target classification module, adjusting the model according to a test result, optimizing parameters, and calculating the accuracy. According to the method, the inaccurate classification effect caused by improper selection of the feature matrix can be effectively avoided; and meanwhile, the classification accuracy of the road moving targets is greatly improved, and particularly, the classification capability of weak targets is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of millimeter wave radars, and in particular to a method for optimizing the classification of road targets by a medium-resolution millimeter wave traffic radar. Background Art

[0002] Millimeter-wave radars are gradually being widely used in the field of smart transportation. Among them, a large proportion of millimeter-wave radars use multiple transmit (TX) and multiple receive (RX) antennas to form a multiple input multiple output (MIMO) array. By transmitting frequency modulated continuous wave (FMCW) signals, receiving and processing the target's reflected echo signals, the distance, speed, and angle information of the target are obtained. On this basis, the target type is identified and its trajectory is tracked at the same time, which effectively improves the operating efficiency and safety of the smart transportation system.

[0003] Target recognition and classification technology is a very important technology in the field of millimeter wave radar. Its purpose is to recognize the type of target by using the characteristics of the target, such as distinguishing between large trucks, medium trucks, cars, non-motor vehicles, pedestrians, etc. At present, a method that is widely used is to use high-resolution range profiles (HRRP), a classification algorithm that uses high resolution in the range dimension to identify the range-dimensional appearance features of targets (mostly stationary targets in open scenes).

[0004] However, in actual use, traffic radar application scenarios are very complex and contain a lot of noise. At the same time, the diversity and multi-posture of target vehicles will also have a great impact on the accuracy of the acquired target features. When the range resolution and angular resolution are limited (that is, medium resolution), it will become more difficult to extract the moving target features using only the medium resolution range profile (MRRP), especially for weak target recognition.

[0005] The information disclosed in this background technology section is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art already known to those skilled in the art. Summary of the invention

[0006] In view of the problems existing in the prior art, the purpose of the present invention is to provide a medium-resolution millimeter-wave traffic radar road target optimization classification method, which is applied to a millimeter-wave radar with a multi-input and multi-output array consisting of multiple transmitters and multiple receivers. It can greatly improve the classification accuracy of road moving targets under limited hardware and system design constraints, especially improve the classification ability of weak targets.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions: A medium-resolution millimeter-wave traffic radar road target optimization classification method, the method comprising the following steps: S1: Filter and preprocess the data collected by the millimeter-wave radar with medium range or angular resolution to obtain M2D-MRRP, which is a simplified two-dimensional extension of MRRP; S2: Build a data set for the valid target data to be classified, and split it into a training set and a test set according to a certain ratio. The training set contains training data and verification data. S3: Build a target classification module based on convolutional neural network and adjust its structure and parameters; S4: Use the data set and target classification module obtained in the above steps to perform target classification training, adjust the model according to the test results, optimize the parameters, and calculate the accuracy.

[0008] Further, the specific implementation method of step S1 is as follows: Step S11: According to the type of road target to be classified, select the optimal range of its movement distance, and store the range-Doppler two-dimensional RD spectrum data of its frame respectively; Step S12: removing the zero-frequency component from the RD spectrum and deleting invalid frames or abnormal frame data to obtain a valid RD spectrum; Step S13: Find the distance and Doppler index of the target, extract the 128 / 2 length data before and after to form a 128*128 target distance-Doppler two-dimensional image, named RD-128 spectrum; if the difference between the target location index and the effective RD spectrum edge index obtained in the previous step is less than 128 / 2, extend in the opposite direction until the size is satisfied while ensuring that this direction is the longest; Step S14: For the data extracted in the previous step, extract the Doppler index of the target and expand it bidirectionally by a width of L to obtain M2D-MRRP data of a size of (2L+1)*128. If the difference between the Doppler index of the target location and the edge of the RD-128 spectrum is less than L, extend in the opposite direction until the size is met while ensuring that this direction is the longest.

[0009] Furthermore, the RD spectrum refers to the data collected by ADC of the target echo of the radar, and then the range-Doppler spectrum is obtained by performing fast Fourier transform-FFT processing in the range dimension and Doppler dimension respectively.

[0010] Furthermore, the principle for selecting the best range of motion distance in step S11 is: within the effective range of radar measurement, medium and short distances are given priority.

[0011] Further, the specific implementation method of step S12 is as follows: According to the RD spectrum area where the distance and Doppler index are close to zero frequency, the data is set to zero or replaced with the background noise value. The distance value and speed value represented by this area are the data samples where the target to be classified cannot exist or cannot be reached; The frame data that cannot be distinguished due to abnormal RD spectrum background noise is deleted.

[0012] Further, in step S14, the specific method of extracting the Doppler index of the target and bidirectionally extending the width L to obtain M2D-MRRP data of size (2L+1)*128 includes: The Doppler index of the target is extracted and bidirectionally expanded by L width to obtain M2D-MRRP data of size (2L+1)*128. Considering the balance between information capacity and storage capacity, the L value of [1,3] is the best. If the difference between the Doppler index at the target location and the edge of the RD-128 spectrum is less than L, it is extended in the opposite direction until the size is met while ensuring that this direction is the longest. If the target Doppler index is at the edge of the RD-128 spectrum, it is only 2L long in the other direction.

[0013] Further, the specific implementation method of step S2 is as follows: Step S21: extract samples from the M2D-MRRP data of all targets, and label each sample to ensure the accuracy and consistency of the labeling; Step S22: randomly split the data according to a preset ratio to ensure that the ratio of each category in the training set and the test set is basically the same to ensure the representativeness of the data.

[0014] Further, the specific implementation method of step S3 is as follows: Step S31: Building a CNN-based target classification module for automatically learning and extracting the range image features and Doppler proximity distribution features of the target in M2D-MRRP; Step S32: After the initial model is imported into the road object classification module, the loss function and optimization algorithm parameters are configured; Step S33: In the model design, the network hyperparameters are adjusted according to the characteristics of the data set to adapt to the M2D-MRRP data size.

[0015] Further, the specific implementation method of step S4 is as follows: Step S41: Use the training set data to train the classification module, calculate the predicted value through forward propagation, calculate the loss using the loss function, and update the model parameters through back propagation; Step S42: Evaluate the model performance on the validation set and adjust the learning rate and other hyperparameters to avoid overfitting and improve the training effect; Step S43: After the training is completed, the model is finally evaluated using the test set, the accuracy and other performance indicators are calculated, and the model is adjusted and optimized as necessary; Step S44: Analyze the performance of the model according to the test results, and perform model improvement and retraining to achieve the best classification effect.

[0016] By adopting the above technical solution, the present invention has the following beneficial effects: (1) Compared with the traditional feature input classification method, the present invention can effectively avoid inaccurate classification results caused by improper feature matrix selection; at the same time, it optimizes the parameter debugging efficiency and accuracy to a certain extent. The traditional feature input classification method is highly sensitive to the feature matrix, and the parameter debugging cycle is long due to the variability of scene targets and scene environments. The target classification method of the present invention belongs to unsupervised learning and has adaptive capabilities for scenes and parameters, so it can effectively overcome the above problems; (2) Compared with the target classification method based only on MRRP, the present invention greatly improves the classification accuracy of road moving targets, especially the classification ability of weak targets. Due to the limited system design of most radars, when the distance resolution and angle resolution cannot achieve the point cloud effect, the MRRP extracted by them is not ideal for resolving moving targets in multiple lanes in traffic road scenes, especially the differentiation of pedestrians and non-motor vehicles is very low. The automatic classification method based on M2D-MRRP proposed by the present invention can improve the above problems and significantly improve the ability to distinguish weak target categories; at the same time, it does not need to change the hardware and system design, effectively saving manpower and time costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 The present invention is a flow chart of a method for optimizing the classification of road targets using a medium-resolution millimeter-wave traffic radar.

[0019] Figure 2 This is a comparison diagram of the road moving target classification effects of a medium-resolution millimeter-wave traffic radar road target optimization classification method described in the present invention and a target classification method based only on MRRP, that is, a comparison of confusion matrix results. DETAILED DESCRIPTION

[0020] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] The specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described here is only used to illustrate and explain the present invention, and is not used to limit the present invention.

[0022] Combination Figure 1 As shown in FIG. 1 , a flow chart of a method for optimizing the classification of road targets by a medium-resolution millimeter-wave traffic radar proposed in the present invention is shown, which is implemented in conjunction with a road target classification case and specifically includes the following steps: Step S1: Filter and preprocess the data collected by the millimeter-wave radar with medium distance or angle resolution to obtain M2D-MRRP (i.e., the simplified two-dimensional extended spectrum of MRRP); the specific implementation steps are as follows: Step S11: After confirming the type of road target to be classified (for example, cars, pedestrians, and bicycles), collect data respectively; according to the current hardware system with medium resolution (for example, the distance resolution is 0.4m), select the movement distance of [30m, 80m] as the optimal range, and after the radar performs ADC acquisition, distance dimension FFT, and Doppler dimension FFT processing on the target echo, store the optical image, road movement point data, and corresponding range-Doppler two-dimensional (RD) spectrum data of the frame respectively, so as to facilitate data alignment in subsequent processing; variables such as target movement direction, movement speed, and lateral position can be set at the same time to increase the sample size of the data set (for example, the target movement direction is set to the incoming direction and the outgoing direction respectively, the movement speed is set to two levels respectively, and the lateral position is set to two levels respectively, that is, different lanes); Step S12: remove the zero-frequency component of the RD spectrum, and delete invalid frames or abnormal frame data to obtain a valid RD spectrum; specifically, for the RD spectrum area where the distance and Doppler index are close to zero frequency (for example, the distance sampling range [0, IDX_Rmax], the Doppler sampling range [0, IDX_Dmax], the area where the two indexes are close to 0, IDX_Rmax, IDX_Dmax is the zero-frequency area), the data is set to zero or replaced with the background noise value. The distance value and speed value represented by this area are impossible for the target to be classified to exist or cannot be reached; the frame data where the target MRRP cannot be distinguished due to the abnormal background noise of the RD spectrum is deleted; Step S13: Find the distance and Doppler index of the target, extract the 128 / 2 length data before and after to form a 128*128 target distance-Doppler two-dimensional image, and obtain the RD-128 spectrum; if the difference between the target location index and the effective RD spectrum edge index is less than 128 / 2, extend in the opposite direction until the size is satisfied while ensuring that this direction is the longest; Step S14: for the data extracted in the previous step, extract the Doppler index of the target and expand it bidirectionally by L width to obtain M2D-MRRP data of size (2L+1)*128. Considering the balance between information capacity and storage capacity, the L value belongs to [1,3] optimally. If the difference between the Doppler index of the target location and the edge of the RD-128 spectrum is less than L, then extend in the opposite direction to meet the size while ensuring that this direction is the longest. If the target Doppler index is at the edge of the RD spectrum, only take 2L length in the other direction.

[0023] Step S2: Build a data set for the valid target data to be classified, and split it into a training set and a test set according to a certain ratio; the specific implementation steps are as follows: Step S21: extract samples from the M2D-MRRP data of all targets, and label each sample to ensure the accuracy and consistency of the labeling; Step S22: randomly split the data according to a preset ratio (for example, 90% training set and 10% test set, where the training set is divided into 90% training and 10% validation), ensuring that the proportion of each category in the training set and the test set is basically the same to ensure the representativeness of the data.

[0024] Step S3: Build a target classification module based on CNN (convolutional neural network) and adjust its structure and parameters; the specific implementation steps are as follows: Step S31: Build a CNN-based target classification module to automatically learn and extract the range image features and Doppler proximity distribution features of the target in M2D-MRRP; select an appropriate VGG network structure (such as VGG-16 or VGG-19), initialize the network model, and adjust the structure and parameter configuration of the model convolution layer and fully connected layer; Step S32: After the initial model is imported into the road object classification module, the loss function (such as cross entropy loss) and optimization algorithm (such as Adam or SGD) are configured, and hyperparameters such as learning rate are set; in this case, the loss function is set to cross entropy loss, and the optimization algorithm is set to SGD; Step S33: In the model design, the network hyperparameters (such as convolution kernel size, step size, pooling strategy, etc.) are adjusted according to the characteristics of the data set to adapt to the M2D-MRRP data size; since its data format is two-dimensional, the kernel size needs to be set to 2*2 or 2*3, and the padding size is (1, 1); the single Batch Size is set to 16 / 32 / 128 respectively, and the accuracy results are slightly different, but the impact on the final classification effect is small; in this embodiment, the Batch Size is set to 16.

[0025] Step S4: Use the data set and target classification module obtained in the above steps to perform target classification training, adjust the model according to the test results, optimize the parameters, and calculate the accuracy. The specific implementation steps are as follows: Step S41: Use the training set data to train the classification module, calculate the predicted value through forward propagation, calculate the loss using the loss function, and update the model parameters through back propagation; Step S42: Evaluate the model performance on the validation set and adjust the learning rate and other hyperparameters to avoid overfitting and improve the training effect; Step S43: After the training is completed, the model is finally evaluated using the test set, the accuracy and other performance indicators (such as precision and recall) are calculated, and the model is adjusted and optimized as necessary; Step S44: Analyze the performance of the model according to the test results, and perform model improvement and retraining to achieve the best classification effect; If you want to compare the effect with other classification methods, you need to formulate similar steps based on the characteristics of another method to complete the training and comparative testing; for example, if the MRRP method is used as the control group, steps S1 and S3 need to be adjusted; wherein steps S13 and S14 are replaced by "extracting the full-length range image of the Doppler index where the target is located to obtain MRRP data"; wherein the kernel size in step S33 is set to 2*1 or 3*1 to adapt to the one-dimensional characteristics of the MRRP data, and similarly the padding size is set to (1, 0); the Batch Size does not need to be changed; then all steps are performed in sequence to complete the training and testing of road target classification using the MRRP method; by comparing the confusion matrix results of this control group with the experimental group of M2D-MRRP, the road motion target classification effect comparison between a medium-resolution millimeter-wave traffic radar road target optimization classification method described in the present invention and a target classification method based only on MRRP can be achieved.

[0026] Combination Figure 2 As shown in the figure, the road target classification effect of the dataset created by the same hardware system is shown in the figure. Figure 2(1) and (2) are the test results of training using the MRRP method and the M2D-MRRP method proposed in the present invention, respectively. It can be seen that the classification accuracy of the M2D-MRRP method proposed in the present invention for pedestrians and bicycles is significantly better than that of the MRRP method.

[0027] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medium-resolution millimeter-wave traffic radar road target optimization classification method, characterized in that: The method comprises the following steps: S1: Filter and preprocess the data collected by the millimeter-wave radar with medium range or angular resolution to obtain M2D-MRRP, which is a simplified two-dimensional extension of MRRP; S2: Build a data set for the valid target data to be classified, and split it into a training set and a test set according to a certain ratio. The training set contains training data and verification data. S3: Build a target classification module based on convolutional neural network and adjust its structure and parameters; S4: Use the data set and target classification module obtained in the above steps to perform target classification training, adjust the model according to the test results, optimize the parameters, and calculate the accuracy.

2. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 1 is characterized in that: The specific implementation method of step S1 is as follows: Step S11: According to the type of road target to be classified, select the optimal range of its movement distance, and store the range-Doppler two-dimensional RD spectrum data of its frame respectively; Step S12: removing the zero-frequency component from the RD spectrum and deleting invalid frames or abnormal frame data to obtain a valid RD spectrum; Step S13: Find the distance and Doppler index of the target, extract the 128 / 2 length data before and after to form a 128*128 target distance-Doppler two-dimensional image, named RD-128 spectrum; if the difference between the target location index and the effective RD spectrum edge index obtained in the previous step is less than 128 / 2, extend in the opposite direction until the size is satisfied while ensuring that this direction is the longest; Step S14: For the data extracted in the previous step, extract the Doppler index of the target and expand it bidirectionally by a width of L to obtain M2D-MRRP data of a size of (2L+1)*128. If the difference between the Doppler index of the target location and the edge of the RD-128 spectrum is less than L, extend in the opposite direction until the size is met while ensuring that this direction is the longest.

3. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 2 is characterized in that: The RD spectrum refers to the data collected by ADC of the target echo of the radar, and then the range-Doppler spectrum is obtained by performing fast Fourier transform-FFT processing in the range dimension and the Doppler dimension respectively.

4. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 2 is characterized in that: The principle for selecting the best range of motion distance in step S11 is: within the effective range of radar measurement, medium and short distances are given priority.

5. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 2 is characterized in that: The specific implementation method of step S12 is as follows: According to the RD spectrum area where the distance and Doppler index are close to zero frequency, the data is set to zero or replaced with the background noise value. The distance value and speed value represented by this area are the data samples where the target to be classified cannot exist or cannot be reached; The frame data that cannot be distinguished due to abnormal RD spectrum background noise is deleted.

6. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 2 is characterized in that: In step S14, the specific method of extracting the Doppler index of the target and bidirectionally extending the width L to obtain M2D-MRRP data of size (2L+1)*128 includes: The Doppler index of the target is extracted and bidirectionally expanded by L width to obtain M2D-MRRP data of size (2L+1)*128. Considering the balance between information capacity and storage capacity, the L value of [1,3] is the best. If the difference between the Doppler index at the target location and the edge of the RD-128 spectrum is less than L, it is extended in the opposite direction until the size is met while ensuring that this direction is the longest. If the target Doppler index is at the edge of the RD-128 spectrum, it is only 2L long in the other direction.

7. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 1 is characterized in that: The specific implementation method of step S2 is as follows: Step S21: extract samples from the M2D-MRRP data of all targets, and label each sample to ensure the accuracy and consistency of the labeling; Step S22: randomly split the data according to a preset ratio to ensure that the ratio of each category in the training set and the test set is basically the same to ensure the representativeness of the data.

8. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 1 is characterized in that: The specific implementation method of step S3 is as follows: Step S31: Building a CNN-based target classification module for automatically learning and extracting the range image features and Doppler proximity distribution features of the target in M2D-MRRP; Step S32: After the initial model is imported into the road object classification module, the loss function and optimization algorithm parameters are configured; Step S33: In the model design, the network hyperparameters are adjusted according to the characteristics of the data set to adapt to the M2D-MRRP data size.

9. The medium-resolution millimeter-wave traffic radar road target optimization classification method according to claim 1 is characterized in that: The specific implementation method of step S4 is as follows: Step S41: Use the training set data to train the classification module, calculate the predicted value through forward propagation, calculate the loss using the loss function, and update the model parameters through back propagation; Step S42: Evaluate the model performance on the validation set and adjust the learning rate and other hyperparameters to avoid overfitting and improve the training effect; Step S43: After the training is completed, the model is finally evaluated using the test set, the accuracy and other performance indicators are calculated, and the model is adjusted and optimized as necessary; Step S44: Analyze the performance of the model according to the test results, and perform model improvement and retraining to achieve the best classification effect.

Images