Deep learning-based millimeter wave radar state detection method, device, and medium

Abstract

The application discloses a kind of millimeter wave radar state detection method, equipment and medium based on deep learning, it is related to artificial intelligence field, detection method includes: radar original detection data is preprocessed to obtain time series dataset;The time series dataset is input to the double-branch deep learning model comprising self-encoder branch and long short-term memory network branch, respectively reconstructs current frame trajectory to obtain current frame restoration data, and predicts next frame trajectory to obtain next frame prediction data;Real data of radar measured current frame and next frame are obtained, determine restoration loss based on the difference between restoration data and real data, determine prediction loss based on the difference between prediction data and real data, and generate offset confidence;When offset confidence is greater than preset abnormal threshold, it is determined that radar installation state is abnormal.This scheme uses single radar data and double error joint determination, without external auxiliary facilities or artificial inspection, can realize roadside radar installation state automatic high-precision monitoring.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a millimeter-wave radar state detection method, device, and medium based on deep learning. Background Technology

[0002] As a core sensing device in intelligent transportation systems and vehicle-road cooperative systems, the accuracy of millimeter-wave radar's installation orientation directly affects the accuracy of subsequent traffic flow statistics and vehicle trajectory tracking. When deployed in complex scenarios such as roads and intersections, this radar equipment not only has to withstand harsh natural weather conditions, but also faces the possibility of slight deflection of the mounting bracket due to external impacts or long-term wind load vibrations.

[0003] In existing technologies, conventional monitoring methods for addressing the risk of offset in this installation state mainly rely on periodic manual inspections or preliminary judgments using simple angle sensors integrated within the radar. However, manual inspections are time-consuming and labor-intensive, and struggle to detect subtle displacements. Furthermore, internal sensors often fail to accurately reflect the radar's actual operating position when it undergoes slow deformation. Some solutions attempt to use error calculation methods based on traditional statistics, trying to assess the radar's operating status by calculating translation and rotation matrices. However, such statistical models typically struggle to maintain accuracy when faced with diverse intersection structures and varying traffic flow characteristics, resulting in poor versatility. In recent years, optimized solutions have emerged that introduce fixed reference points for calibration or employ multi-sensor fusion for auxiliary verification. While these solutions can improve detection accuracy to some extent, the additional sensors or physical markers significantly increase the complexity of system deployment and hardware costs. They also add to the cumbersome process of synchronizing and cross-verifying multi-source data. If the auxiliary equipment itself malfunctions, it can lead to misjudgments.

[0004] It is evident that existing detection methods face problems of high cost or insufficient reliability in practical applications, and the overall inspection process is highly dependent on personnel experience and external conditions, making it difficult to meet the needs of remote, low-maintenance automatic monitoring for large-scale deployment of radar units on the roadside. Summary of the Invention

[0005] This invention provides a method, device, and medium for detecting the status of millimeter-wave radar based on deep learning. The technical problem it aims to solve is: how to provide an effective solution for automatically monitoring the installation status of roadside millimeter-wave radar without relying on external auxiliary facilities and frequent human intervention.

[0006] In a first aspect, the present invention provides a millimeter-wave radar state detection method based on deep learning, comprising:

[0007] In response to receiving raw detection data output from the millimeter-wave radar, the raw detection data is preprocessed to obtain a time-series dataset;

[0008] The time series dataset is input into a two-branch deep learning model that includes an autoencoder branch and a long short-term memory network branch. The vehicle motion trajectory of the current frame is reconstructed and restored through the autoencoder branch to obtain the restored data of the current frame. The vehicle motion trajectory of the next frame is temporally predicted through the long short-term memory network branch to obtain the predicted data of the next frame.

[0009] The current frame real data and the next frame real data output by the millimeter-wave radar are obtained. Based on the difference between the current frame restored data and the current frame real data, the restoration loss is determined. Based on the difference between the next frame predicted data and the next frame real data, the prediction loss is determined. The offset confidence is determined based on the restoration loss and the prediction loss.

[0010] Determine whether the offset confidence level is greater than a preset anomaly threshold; if it is greater than the preset anomaly threshold, then determine that the installation status of the millimeter-wave radar is abnormal.

[0011] Optionally, the preprocessing of the original detection data to obtain a time-series dataset includes:

[0012] The original detection data includes the target vehicle's horizontal and vertical coordinates relative to the radar, its speed, and its heading angle; based on the speed and heading angle, the speed is decomposed into a horizontal axis velocity component and a vertical axis velocity component, and the horizontal and vertical coordinates, the horizontal axis velocity component, and the vertical axis velocity component are used as a feature set;

[0013] Calculate the standard score for each feature dimension in the feature set, and remove outlier data points whose absolute value of the standard score is greater than a preset filtering threshold;

[0014] Based on the mean and standard deviation of the training dataset, the remaining feature set after removing the outlier data points is standardized, and the standardized data is reshaped into a time series data format according to a preset time step to obtain the time series dataset.

[0015] Optionally, the step of reconstructing and restoring the vehicle motion trajectory of the current frame through the autoencoder branch to obtain the restored data of the current frame includes:

[0016] The time series dataset is input into the encoder of the autoencoder branch, and the encoder extracts the data features of the time series dataset and compresses them to obtain a low-dimensional feature representation.

[0017] The low-dimensional feature representation is input into the decoder of the autoencoder branch, and the decoder performs data reconstruction and restoration based on the low-dimensional feature representation to obtain the restored data of the current frame.

[0018] Optionally, the Long Short-Term Memory (LSTM) network branch includes multiple LTM units and multiple fully connected layers connected sequentially; the step of performing temporal prediction of the vehicle trajectory for the next frame through the LTM network branch to obtain the next frame prediction data includes:

[0019] The temporal dataset is input into the Long Short-Term Memory (LSTM) network branch, and the multiple LSM units in the LSM network branch sequentially and incrementally extract temporal correlation features. The extracted temporal correlation features are then input into the multiple fully connected layers for dimensionality mapping in an incremental manner to obtain the next frame prediction data.

[0020] Optionally, determining the offset confidence based on the restoration loss and the prediction loss includes:

[0021] The restoration loss and prediction loss of the current detection period are compared with historical statistical period data to determine the average period difference percentage.

[0022] A dynamic adjustment slope that is positively correlated with the period mean difference percentage is determined. Based on the dynamic adjustment slope and a preset error threshold, the period mean difference percentage is mapped to an offset confidence level in the range of 0 to 100 using a hyperbolic tangent function.

[0023] Optionally, after determining that the installation status of the millimeter-wave radar is abnormal, the method further includes:

[0024] The offset of the millimeter-wave radar is calculated based on the predicted data of the next frame and the actual data of the next frame, and a correction prompt message containing the offset is output.

[0025] Optionally, calculating the offset of the millimeter-wave radar based on the predicted data of the next frame and the actual data of the next frame includes:

[0026] Based on the single-point error between the predicted data of the next frame and the actual data of the next frame, the horizontal and vertical coordinate translation offset is calculated.

[0027] Based on the predicted data of the next frame, a coordinate rotation transformation is performed, and based on the global cumulative error between the predicted data of the next frame after the coordinate rotation transformation and the actual data of the next frame, the angle rotation offset is calculated.

[0028] The horizontal and vertical coordinate translation offsets and the angular rotation offsets are used as the offsets.

[0029] Optionally, the step of calculating the angular rotation offset based on the global cumulative error between the predicted data of the next frame after coordinate rotation transformation and the actual data of the next frame includes:

[0030] The mean square error between the predicted data of the next frame after coordinate rotation transformation and the coordinates corresponding to the real data of the next frame is calculated, and the mean square errors corresponding to all trajectory points in the current detection period are accumulated and summed to obtain the global accumulated error.

[0031] The angular rotation offset is solved with the goal of minimizing the global cumulative error.

[0032] Secondly, the present invention also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0033] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described method.

[0034] This invention provides a deep learning-based millimeter-wave radar state detection method, device, and medium. The method involves a computing device that receives raw detection data from the radar itself. After preprocessing, a time-series dataset is obtained. A two-branch deep learning model, including an autoencoder branch and a long short-term memory network branch, is used to reconstruct the vehicle trajectory in the current frame and predict the vehicle trajectory in the next frame, outputting corresponding reconstructed data for the current frame and predicted data for the next frame. Further, after acquiring the actual data of the current and next frames measured by the radar, the difference between the reconstructed data and the actual data is used to determine the reconstruction loss, and the difference between the predicted data and the actual data is used to determine the prediction loss. Based on these, a comprehensive offset confidence level is determined. When the offset confidence level exceeds a preset anomaly threshold, an anomaly in the radar installation state can be accurately identified. This method relies solely on the output data of a single radar, without the need for any fixed reference points or the fusion of other sensors, and without frequent manual inspections. It achieves automated, high-precision monitoring of the installation state of roadside millimeter-wave radar, effectively solving the drawbacks of existing detection methods that rely on external facilities and manual intervention. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 A flowchart illustrating a deep learning-based millimeter-wave radar state detection method provided in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating a practical application scenario of a millimeter-wave radar installation status detection system provided in an embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of a deep learning model structure provided in an embodiment of the present invention;

[0039] Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0042] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0043] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0044] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0045] Please see Figure 1 This invention provides a deep learning-based millimeter-wave radar state detection method. This method is executed by a computing device (e.g., an edge computing device) deployed in a roadside or intersection environment, and the computing device maintains a data communication connection with the millimeter-wave radar to be detected. This embodiment constructs a two-branch deep learning model including an autoencoder branch and a long short-term memory network branch to perform current frame data reconstruction and next frame data prediction on the time-series detection data of the target vehicle acquired by the radar. By utilizing the joint fluctuation of reconstruction error and prediction error, it achieves automated and high-precision detection of minute radar installation offsets. Figure 2 As shown, in a practical deployment scenario, a millimeter-wave radar is mounted on a pole on the side of a road. When the millimeter-wave radar is in its normal installation state, its detection area can accurately cover the area where the target traffic flow is located (indicated by the yellow dashed line in the figure). When the millimeter-wave radar is affected by external forces during operation, causing its installation attitude to deviate in angle or position, its detection area will shift accordingly (indicated by the red dashed line in the figure). This installation deviation will directly cause a systematic deviation in the horizontal and vertical coordinates and heading angle of the vehicles sensed by the millimeter-wave radar relative to the radar, thereby causing the overall translation or rotation of the subsequent analysis data.

[0046] Specifically, the method includes the following steps:

[0047] S1, in response to receiving the raw detection data output by the millimeter-wave radar, preprocess the raw detection data to obtain a time-series dataset.

[0048] In practice, the millimeter-wave radar continuously emits electromagnetic waves towards the target vehicle within its radiation coverage area, and calculates the real-time motion state of the target vehicle based on the echo signal, outputting raw detection data. Upon receiving this raw detection data, the computing device first preprocesses it to eliminate data jitter caused by radar electromagnetic wave scattering and environmental noise interference, while also eliminating dimensional differences between different physical quantities, thus obtaining a time-series dataset.

[0049] In some preferred embodiments, the preprocessing of the original detection data to obtain a time-series dataset includes: the original detection data containing the target vehicle's horizontal and vertical coordinates relative to the radar, its driving speed, and its heading angle; decomposing the driving speed into horizontal and vertical velocity components based on the driving speed and the heading angle, using the horizontal and vertical coordinates, the horizontal and vertical velocity components, and the vertical velocity components as a feature set; calculating the standard score for each feature dimension in the feature set, and removing outlier data points whose absolute values ​​of the standard scores are greater than a preset filtering threshold; standardizing the remaining feature set after removing the outlier data points based on the mean and standard deviation of the training dataset, and reshaping the standardized data into a time-series data format according to a preset time step to obtain the time-series dataset.

[0050] In practical implementation, in edge computing scenarios, the data directly output by millimeter-wave radar often contains noise, outliers, and dimensional differences between different physical quantities. If used directly in deep learning models without processing, it can lead to convergence difficulties or inference biases. This embodiment utilizes vector decomposition and statistical anomaly filtering to transform the raw detection data into an effective temporal feature set. The specific implementation method is as follows:

[0051] The raw detection data output by the millimeter-wave radar records the spatial position and motion state of the target vehicle relative to the radar, specifically including the target vehicle's horizontal and vertical coordinates, speed, and heading angle relative to the radar. The speed is a scalar value reflecting the vehicle's instantaneous kinetic energy, and the heading angle is the deflection angle of the vehicle's current speed direction measured counterclockwise with due north as the reference. Further, after extracting the above four-dimensional physical quantities, the computing device first decomposes the vehicle's speed into horizontal and vertical velocity components based on the principle of velocity vector decomposition. This decomposition process follows the principle of vector projection, projecting the vehicle's current resultant velocity vector onto a two-dimensional Cartesian coordinate system with the radar's location as the origin. This allows the subsequent deep learning model to simultaneously capture the vehicle's lateral lane changes and longitudinal straight-line motion characteristics from the evolution of the vehicle's horizontal and vertical velocity components. Further, the velocity vector decomposition formula used is as follows:

[0052]

[0053] Where v represents the current speed (resultant speed) of the target vehicle; α represents the vehicle's heading angle measured counterclockwise with due north as the reference direction; v x The horizontal velocity component obtained after decomposition; v y The vertical axis velocity component is obtained after decomposition. After this operation, the target vehicle information at each time step is standardized into a four-dimensional feature set containing horizontal and vertical coordinates, horizontal axis velocity component, and vertical axis velocity component.

[0054] Furthermore, due to the potential for radar to be obstructed by rain and snow, experience sporadic multipath effects, or suffer from transient electromagnetic interference in complex outdoor environments, the original feature set may contain some outliers that deviate from physical laws. Therefore, the computing device calculates the standard score (Z-score) for each feature dimension in the feature set. The standard score is a statistical measure, using standard deviation as the unit, to indicate the degree to which a data point deviates from the mean. The outlier filtering formula is shown below:

[0055]

[0056] Where D represents the inclusion of x, y, and v x v y The complete set of original features in four dimensions; D filtered This represents the valid feature dataset retained after filtering; z x z y z vx z vy These represent the standard scores (Z-scores) calculated for the corresponding four dimensions, which are the quotients of the difference between the feature value and the sample mean divided by the sample standard deviation; max(...) represents taking the maximum value among the absolute values ​​of the standard scores for these four dimensions. According to statistical principles, under a normal distribution, the probability of the absolute value of the standard score being greater than 3 is extremely low. Therefore, using ≤3 as a filtering condition can effectively eliminate outlier noise caused by occasional interference.

[0057] Furthermore, after filtering, the computing device performs standardization on the retained feature set based on the mean and standard deviation of the pre-statistical training dataset. This eliminates the numerical magnitude differences between the horizontal and vertical coordinates (meters) and the speed (meters / second) caused by different units, ensuring that the update speed of each dimension during model training remains synchronized. Further, the standardized data is reshaped into a time-series data format according to a preset time step (for example, the time step can be set to 5, meaning 5 consecutive frames constitute a sample), thus obtaining the time-series dataset.

[0058] This embodiment effectively eliminates the nonlinear mapping between the original heading angle and rectangular coordinates by performing vector decomposition of the driving speed, providing the model with a coordinate system input with clear physical meaning. Furthermore, the Z-score filtering method can remove abnormal jump points caused by environmental factors in radar detection, reducing the interference of dirty data on model feature extraction; the standardization operation eliminates the difference in physical quantities, which helps to accelerate the gradient descent convergence of the deep learning model during training.

[0059] S2, the time series dataset is input into a two-branch deep learning model including an autoencoder branch and a long short-term memory network branch. The vehicle motion trajectory of the current frame is reconstructed and restored through the autoencoder branch to obtain the restored data of the current frame. The vehicle motion trajectory of the next frame is temporally predicted through the long short-term memory network branch to obtain the predicted data of the next frame.

[0060] In practice, the computing device inputs the time-series dataset into a pre-trained, offline-trained dual-branch deep learning model. This dual-branch deep learning model is logically divided into two independently processing sub-branches. The first branch is an autoencoder branch, which learns the feature distribution of the vehicle in the radar's normal coordinate system. It can reconstruct the vehicle's trajectory at the current moment based on the current time-series input and output the current frame's reconstructed data. The second branch is a long short-term memory network branch, which focuses on capturing the vehicle's motion patterns and dynamic characteristics in continuous time series. It can infer the future position based on current and historical data and output the next frame's predicted data. In practice, both the autoencoder and the long short-term memory network have the ability to process time-series data. The two branches can comprehensively understand the input sequence from two complementary dimensions: "data reproduction" and "trend prediction."

[0061] For example, in some preferred embodiments, the step of reconstructing and restoring the vehicle motion trajectory of the current frame through the autoencoder branch to obtain the restored data of the current frame includes: inputting the time-series dataset into the encoder of the autoencoder branch, whereby the encoder extracts the data features of the time-series dataset and compresses them to obtain a low-dimensional feature representation; inputting the low-dimensional feature representation into the decoder of the autoencoder branch, whereby the decoder performs data reconstruction and restoration based on the low-dimensional feature representation to obtain the restored data of the current frame.

[0062] In practice, an autoencoder is an unsupervised learning model. Its core design concept is to force the network to learn a compressed representation of the input data in a low-dimensional latent space by constructing a bottleneck structure of "encoder-decoder". Then, the decoding network restores it to an output that is as consistent as possible with the input, thereby achieving automatic extraction of the essential distribution features of the data. The specific implementation method is as follows:

[0063] The computing device inputs the time-series dataset obtained in the aforementioned steps into the encoder of the autoencoder branch. In this embodiment, the encoder is composed of multiple layers of long short-term memory units sequentially connected in series. Because vehicle trajectories exhibit significant temporal continuity, ordinary fully connected neural networks lack the ability to remember historical information. However, long short-term memory units, through their internal forget gate, input gate, and output gate structures, can dynamically control the retention and updating of historical information, effectively extracting the long-range dependencies of vehicle trajectories over time. Furthermore, after receiving the time-series input, as the data is passed through layer by layer of long short-term memory units, the encoder's output feature dimension is gradually compressed, ultimately extracting a low-dimensional feature representation. This low-dimensional feature representation is a compact mathematical code for the vehicle motion pattern.

[0064] Furthermore, after the encoder completes feature compression, it passes the low-dimensional feature representation to the decoder of the autoencoder branch. The decoder's network structure is basically symmetrical to the encoder in terms of the number of layers, also consisting of multiple layers of long short-term memory units. The decoder reconstructs the low-dimensional feature representation layer by layer, finally outputting the current frame reconstruction data with the same dimension as the input time-series dataset. This output reflects the autoencoder branch's accurate understanding of the current vehicle trajectory distribution.

[0065] In this embodiment, the autoencoder branch uses long short-term memory units as its basic module, enabling the network to effectively preserve the temporal contextual dependencies of the input trajectory. The encoder compresses high-dimensional temporal features into a low-dimensional representation, essentially searching for the latent space distribution pattern of the vehicle trajectory; the decoder then uses this distribution pattern to reconstruct the data. When radar offset causes translation or rotation of the coordinate system, the distribution features of the vehicle trajectory will systematically drift in the latent space. The reconstruction based on the old distribution pattern by the autoencoder will produce significant deviations, thus sensitively reflected in the reconstruction loss.

[0066] In some preferred embodiments, the Long Short-Term Memory (LSTM) network branch includes a plurality of LSM units and a plurality of fully connected layers connected sequentially; the step of performing temporal prediction of the vehicle trajectory in the next frame through the LSM network branch to obtain the next frame prediction data includes: inputting the temporal dataset into the LSM network branch, extracting temporal correlation features sequentially and incrementally from the plurality of LSM units in the LSM network branch, and inputting the extracted temporal correlation features into the plurality of fully connected layers for dimensionality mapping in a sequentially and incrementally manner to obtain the next frame prediction data.

[0067] In practice, Long Short-Term Memory (LSTM) networks overcome the gradient vanishing or gradient exploding problems that classic recurrent neural networks easily encounter when training long sequences by introducing a gating mechanism. This makes them suitable for processing data with strong temporal correlation, such as vehicle trajectories. The specific implementation is as follows:

[0068] In this embodiment, the Long Short-Term Memory (LSTM) network branch includes multiple LSM units and multiple fully connected layers connected sequentially. In actual deployment, the number of LSM units can be set to 2 or 3 layers depending on the length of the time series to be processed, and the number of fully connected layers can be set to 3 layers. After the computing device inputs the time series dataset obtained in the previous embodiment into the LSM network branch, the LSM units in the network branch first extract time-series correlation features in an incremental manner. The term "incrementally" means that as the data is passed from the first layer of the network to subsequent layers, the level and complexity of the abstract features extracted by the network gradually increase. This layer-by-layer incremental feature extraction design enables the LSM network branch to obtain structured, hierarchical time-series dynamic information from the original trajectory data.

[0069] Furthermore, after feature extraction is completed by the multi-layer long short-term memory units, the computing device passes the temporally correlated features to the fully connected layer. The role of the fully connected layer is to perform linear weighting and nonlinear transformations, mapping the high-dimensional temporal features to the system's desired target output space. In this embodiment, the target output space corresponds to the vehicle's predicted trajectory coordinates and velocity components in the next time frame, i.e., values ​​in four dimensions. The number of neurons in the fully connected layer is mapped in a progressively decreasing manner, eventually shrinking to four neurons. Through this feature dimensionality reduction operation, the network concretizes the abstract temporal motion pattern into quantifiable prediction data for the next frame, thereby achieving effective prediction of the vehicle's motion state in future moments.

[0070] In this embodiment, the Long Short-Term Memory (LSTM) network branch, through a progressively increasing network structure, can gradually learn from the short-term displacement of the vehicle trajectory to the long-term macroscopic motion pattern, overcoming the memory decay problem of single-layer networks in long sequence modeling. Simultaneously, the decreasing structure of the fully connected layers achieves a smooth mapping from high-dimensional temporal correlation features to low-dimensional physical prediction values, ensuring that the output prediction data is highly consistent with the actual vehicle trajectory in both numerical magnitude and physical meaning. This accurate prediction capability keeps the prediction loss low under normal radar conditions, while during radar offset, the prediction loss clearly reflects the distribution deviation.

[0071] For example, such as Figure 3 As shown, the dual-branch deep learning model is configured with an autoencoder branch on the left and a long short-term memory network branch on the right. The input end of the model is the input layer, which receives and processes data batches with a data tensor structure of (N, 5, 4). Here, N represents the number of samples contained in the processing batch; 5 represents the time step of the time series data, that is, each sample consists of 5 consecutive frames of vehicle trajectory observation data; 4 represents the dimension of the input feature vector, that is, the horizontal and vertical coordinates and the horizontal and vertical velocity components mentioned above.

[0072] Furthermore, the autoencoder branch on the left contains an encoder and a decoder. The encoder comprises two layers of Long Short-Term Memory (LSTM) units with decreasing node counts and a fully connected (Dense) layer. Dropout layers and batch normalization (BN) layers are sequentially configured between each layer. The dropout layers randomly discard some neuron outputs to prevent overfitting, and the batch normalization layers stabilize the input distribution of the network layers during training. The encoder compresses data layer by layer and outputs a low-dimensional feature representation. The decoder comprises two layers of LSTM units with increasing node counts and a fully connected layer, used to reconstruct the data distribution features based on the low-dimensional feature representation, ultimately generating the recovered output for the current frame.

[0073] Furthermore, in the Long Short-Term Memory (LSTM) branch on the right, it first consists of two layers of LSM units with progressively increasing numbers of nodes, used to gradually extract higher-order temporal correlation features. Following this are multiple fully connected layers with progressively decreasing numbers of nodes, used to map the extracted temporal correlation features to the target dimension space. Dropout layers are configured between the progressively decreasing fully connected layers to improve generalization ability. The LSM branch ultimately outputs the predicted data for the next frame.

[0074] It should be further noted that the dual-branch deep learning model needs to be trained offline before being deployed to the inference stage. Specifically, the training stage adopts a supervised learning paradigm. The training dataset comes from multiple vehicle trajectory data collected by actual roadside radar within a preset time period. After the aforementioned preprocessing steps, standardized time-series samples are formed, with a sample size of no less than 100,000 frames. During training, the reconstruction loss of the autoencoder branch adopts the mean squared error loss function, and the prediction loss of the long short-term memory network branch also adopts the mean squared error loss function. The entire dual-branch network is jointly trained end-to-end based on the Adam optimizer. The initial learning rate is set to 0.001, the batch size can be set to 32, and the number of training epochs can be set to 200. This invention does not specifically limit these settings.

[0075] S3, acquire the current frame real data and the next frame real data output by the millimeter-wave radar, determine the restoration loss based on the difference between the current frame restored data and the current frame real data, determine the prediction loss based on the difference between the next frame predicted data and the next frame real data, and determine the offset confidence level based on the restoration loss and the prediction loss.

[0076] In specific implementation, while the model performs inference, the computing device acquires real-time data from the millimeter-wave radar, including the actual data of the current frame and the next frame detected by the radar at the same moment. The computing device calculates the mean square error between the reconstructed data and the actual data of the current frame, defining this error as the reconstruction loss, which quantifies the degree of deviation of the autoencoder branch at the data reconstruction level. Simultaneously, it calculates the mean square error between the predicted data and the actual data of the next frame, defining this error as the prediction loss, which quantifies the degree of deviation of the long short-term memory network branch at the trend prediction level. Furthermore, when the radar installation is stable and without deviation, the vehicle's projection in the radar coordinate system conforms to the distribution pattern learned by the model during historical training, and both the reconstruction loss and prediction loss remain at historically low levels. However, when the radar is subjected to external impact, strong wind vibration, or aging of the bracket causing a physical shift in the installation angle or position, the vehicle trajectory in the radar coordinate system will undergo overall translation or rotation. This will cause a systematic shift in the distribution of the model's input data, resulting in a significant gap between the model's reconstruction and prediction based on old patterns and the actual collected data, thus causing a substantial increase in both the reconstruction loss and prediction loss.

[0077] Furthermore, based on the calculated restoration loss and prediction loss, the computing device comprehensively determines a nonlinear index, namely the offset confidence level, that can quantify the risk of radar current installation state deviation. The value range of this confidence level is mapped to an integer interval between 0 and 100.

[0078] In some preferred embodiments, determining the offset confidence level based on the restoration loss and the prediction loss includes: comparing the restoration loss and the prediction loss of the current detection period with historical statistical period data to determine the period average difference percentage; determining a dynamic adjustment slope positively correlated with the period average difference percentage; and mapping the period average difference percentage to an offset confidence level in the range of 0 to 100 using a hyperbolic tangent function based on the dynamic adjustment slope and a preset error threshold.

[0079] In practice, after the computing device obtains the restored loss and predicted loss for the current detection cycle, it first compares these losses longitudinally with historical statistical cycle data to determine the percentage of the cycle mean error. Historical statistical cycle data refers to the baseline mean and standard deviation of these two losses during a period when the radar was in a stable installation state and confirmed as "healthy." The percentage of the cycle mean error reflects the relative degree to which the error of the current cycle deviates from the historical normal fluctuation range. The formula for calculating the mean square error is as follows:

[0080]

[0081] Where RL is the Reconstruction Loss; PL is the Prediction Loss; N is the number of sample points participating in the current detection period; y r The data is reconstructed from the current frame output by the model's autoencoder branch; y p The next frame prediction data output by the Long Short-Term Memory (LSTM) branch of the model; y t The current frame's actual data is the radar's actual input at the current moment; y t+1 This represents the actual data of the next frame input by the radar at the next moment; ∑ represents the summation operation of the error of all sample points in the current period.

[0082] Further, the computing device determines a dynamically adjusted slope that is positively correlated with the stated percentage of the period mean difference. This dynamically adjusted slope is a hyperparameter, denoted as k, that automatically adjusts the rate of change based on the magnitude of the percentage of the period mean difference. A nonlinear confidence mapping for this percentage is calculated using the following formula:

[0083]

[0084] Where confidence represents the offset confidence level of the final output; c is the specific confidence score, limited to between 0 and 100; z is the percentage of the average period difference calculated above; k is the dynamic adjustment slope, which is positively correlated with z, that is, when z increases, k also increases, thus making the alarm more sensitive; threshold is the preset error baseline threshold; tanh is the hyperbolic tangent function.

[0085] Furthermore, by using the difference (k · (z - threshold)) as the input to the hyperbolic tangent function, a nonlinear mapping effect can be achieved, where small error fluctuations are suppressed under normal conditions and amplified sharply under abnormal critical conditions.

[0086] This embodiment maps the discrete fluctuations of error to a confidence level of 0 to 100 with a clear physical boundary by combining the hyperbolic tangent function with a dynamic slope, thus achieving a quantitative expression of abnormal risk. Furthermore, the adaptive adjustment strategy of the dynamic slope k ensures that the system maintains a certain tolerance to filter out environmental noise when error fluctuations are small, while rapidly increasing alarm sensitivity when errors increase. This nonlinear mapping mechanism effectively avoids the "false alarm" or "missed alarm" problems that easily occur in fixed threshold determination, ensuring the accuracy of subsequent threshold judgments.

[0087] S4, determine whether the offset confidence level is greater than a preset abnormal threshold; if it is greater than the preset abnormal threshold, determine that the installation status of the millimeter-wave radar is abnormal.

[0088] In practice, it is determined whether the aforementioned offset confidence level is greater than a preset anomaly threshold. The preset anomaly threshold is an empirical critical value calibrated through extensive experimental testing; for example, it can be set to 80. If the offset confidence level is greater than 80, the computing device determines that the current installation status of the radar has undergone a significant anomaly, i.e., it determines that the millimeter-wave radar installation status is abnormal, and the subsequent offset quantization and correction prompt process needs to be initiated. If the offset confidence level is not greater than 80, the computing device determines that the millimeter-wave radar installation status is normal, automatically returns to the aforementioned step of receiving the original detection data, and enters the next detection cycle to continue continuous monitoring.

[0089] This embodiment achieves automated detection of minute deviations in the installation status of roadside millimeter-wave radar by employing a complementary dual-path error judgment mechanism of reconstruction and prediction, without relying on any manual inspections or additional sensor equipment. Since reconstruction loss and prediction loss provide verification from two independent dimensions—"reconstruction of past distribution" and "prediction of future trends"—the joint judgment of dual errors effectively suppresses occasional false alarms that may arise from single-path evaluation, thereby significantly improving the robustness of the system's judgment while ensuring detection sensitivity.

[0090] In some preferred embodiments, after determining that the installation status of the millimeter-wave radar is abnormal, the method further includes: calculating the offset of the millimeter-wave radar based on the next frame predicted data and the next frame real data, and outputting a correction prompt message containing the offset.

[0091] In practice, when the system detects an installation offset of the radar, simply outputting a qualitative "abnormal" label is insufficient for actual operation and maintenance. Quantitative data on the specific offset direction, translation distance, and rotation angle are also required to guide precise physical calibration or software compensation. The specific implementation method is as follows:

[0092] Once the computing device determines that the millimeter-wave radar's installation status is abnormal, it continues to calculate the radar's offset based on the predicted data of the next frame and the actual data of the next frame. The underlying logic is that when the radar experiences a physical installation offset, the vehicle's predicted position calculated by the model based on historical patterns will inevitably have a systematic translational or rotational deviation from the vehicle's actual position measured by the radar in the physical coordinate system. Since the model has learned the absolute coordinate system distribution of the vehicle's trajectory under healthy conditions during training, the magnitude, direction, and angle of this deviation directly reflect the true state of the radar's offset in physical space. Therefore, by finding the geometric transformation parameters (i.e., translation and rotation angle) that minimize the overall deviation between the model's predicted value and the radar's measured value using optimization methods, the current physical offset state of the radar can be solved.

[0093] First, regarding the translation offset, the system employs a single-point error minimization method. If the radar only undergoes coordinate translation, the vector difference between the predicted and true vehicle coordinates in the current frame should equal the radar's translation offset. The formula for calculating the translation offset is as follows:

[0094]

[0095] Among them, (Δx) * , Δy * ) represents the minimum translation offset of the final solved x and y coordinates; argmin is the optimization operator for finding the minimum value of the expression within the parentheses; x p y p Δx and Δy are the predicted coordinates of the next frame output by the Long Short-Term Memory (LSTM) network branch; Δx and Δy are the translation compensation values ​​to be solved as independent variables during the optimization process; x t+1 y t+1 The actual coordinates of the next frame measured by the millimeter-wave radar.

[0096] Furthermore, after determining the translational offset, the computing device needs to solve for the radar's angular rotation offset. Before solving for the rotation offset, a spatial rotation transformation needs to be performed on the predicted coordinates, and the transformation relationship is shown in the following formula:

[0097]

[0098] Where x' and y' represent the original predicted coordinates (x, y, ..., y) respectively. p , y p The new predicted coordinates are obtained by rotating the radar origin by an angle Δα; Δα is the angle offset to be solved as the independent variable.

[0099] Furthermore, the calculation of a single point in a single frame is susceptible to occasional noise, resulting in poor stability of the calculation results. Therefore, the computing device performs global optimization by accumulating the overall error of all trajectory points within the current detection period. The optimal angle offset is then calculated by minimizing this global error, as shown in the following formula:

[0100]

[0101] Among them, (Δα) * ) represents the final optimal angular rotation offset; ∑ represents the summation of errors for all vehicle trajectory points within the current detection cycle; x' and y' are the theoretical predicted coordinates calculated using the aforementioned coordinate rotation transformation formula; x t+1 y t+1 The actual input coordinates are for the corresponding millimeter-wave radar.

[0102] Furthermore, after the computing device completes the calculation of the above translation and rotation, it uses them as the final offset to generate a correction prompt message containing information such as the horizontal and vertical coordinate translation offset, the angular rotation offset, and the radar device number. This message is then output to the system backend or maintenance terminal to provide accurate quantitative guidance for subsequent radar physical reset or algorithm software compensation.

[0103] This embodiment decouples radar offset into independent translation and rotation components, which are then solved using single-point least squares and global cumulative optimization methods, respectively, achieving accurate differentiation of the offset properties. Furthermore, compared to traditional comprehensive error solutions, this step-by-step calculation avoids the ambiguity caused by the coupling of translation and rotation errors, and the physical meaning of the output offset is clear, which can be directly used to guide actual operation and maintenance, improving the processing efficiency after anomaly detection.

[0104] In some preferred embodiments, calculating the offset of the millimeter-wave radar based on the next frame predicted data and the next frame actual data includes: solving for the horizontal and vertical coordinate translation offset based on the single-point error between the next frame predicted data and the next frame actual data; performing coordinate rotation transformation based on the next frame predicted data, and solving for the angle rotation offset based on the global cumulative error between the next frame predicted data and the next frame actual data after coordinate rotation transformation; and using the horizontal and vertical coordinate translation offset and the angle rotation offset as the offset.

[0105] In practical engineering applications, the installation offset of millimeter-wave radar often involves two independent physical forms: the first is a translation of the radar along the horizontal or vertical axis, but its orientation remains unchanged; the second is an angular rotation of the radar around its installation point, but its position coordinates remain unchanged. These two forms have drastically different mathematical representations in spatial geometry, and performing step-by-step calculations helps improve the accuracy of the calculations.

[0106] First, the horizontal and vertical coordinate translation offsets are calculated using the aforementioned translation offset calculation formula. The system calculates the single-point squared error by subtracting the actual trajectory coordinates from the predicted trajectory coordinates after adding an adjustable compensation value, and then uses the argmin operator to minimize this single-point error to solve for the horizontal and vertical coordinate offsets. Based on this, a coordinate rotation transformation with a parameter of Δα is applied to the entire predicted trajectory according to the aforementioned coordinate rotation transformation formula, simulating the overall rotation of the vehicle trajectory in the radar coordinate system when the radar undergoes an angular deflection. Further, the errors of all trajectory points within the current detection period are accumulated according to the aforementioned global optimization formula, and the optimal rotation angle that minimizes the total global error, i.e., the angular rotation offset, is locked using the argmin operator.

[0107] Furthermore, the computing device uses the horizontal and vertical coordinate translation offsets and the angular rotation offsets together as the final offset output, which fully describes the radar's deflection in physical space.

[0108] This embodiment separates the physical form of the radar installation state offset through a step-by-step solution. The translation is solved using least squares based on single-point residuals, while the rotation is solved using a search based on global accumulated error. This phased optimization strategy allows the two deformations with different physical meanings to be calculated independently and accurately, avoiding mutual interference between translation and rotation parameters caused by numerical coupling in the hybrid model.

[0109] In some preferred embodiments, the step of solving for the angle rotation offset based on the global cumulative error between the predicted data of the next frame after coordinate rotation transformation and the real data of the next frame includes: calculating the mean square error of the coordinates corresponding to the predicted data of the next frame after coordinate rotation transformation and the real data of the next frame, and summing the mean square errors corresponding to all trajectory points in the current detection period to obtain the global cumulative error; and solving for the angle rotation offset with the goal of minimizing the global cumulative error.

[0110] In practice, during actual road tests, the trajectory of a single vehicle may be affected by sudden changes in driving behavior, leading to unstable single-point errors. Therefore, this embodiment adopts a global accumulation optimization strategy to eliminate the influence of single-point noise.

[0111] The computing device first calculates the mean square error between the predicted data of the next frame after coordinate rotation transformation and the corresponding coordinates of the actual data of the next frame. Specifically, for each vehicle trajectory point in the current detection cycle, its predicted coordinates are rotated by an assumed angle Δα, and the difference between these coordinates and the actual coordinates of the radar at the corresponding time is squared to obtain the mean square error of a single trajectory point.

[0112] Furthermore, based on the aforementioned global optimization formula, the computing device sums the mean square errors corresponding to all trajectory points within the current detection period to obtain the global cumulative error. Here, "cumulative summation" (∑) means that the system integrates the contributions of all passing vehicle samples within a time window. This operation can significantly dilute the interference caused by a single abnormal trajectory point because the error information of the correct trajectory is accumulated over a large area, while excessively large single-point noise accounts for only a small proportion of the overall cumulative error, thus making the overall error trend smoother and more reliable.

[0113] Furthermore, the computing device, with the objective of minimizing the global accumulated error (i.e., the argmin operator), traverses all possible rotation angles within a preset reasonable angle search range according to a set step size, and locates the angle value that minimizes the accumulated error. The output result is (Δα)* ( ), which can accurately reflect the actual deflection angle of radar in the physical world.

[0114] This embodiment expands the error calculation object from a "single frame point" to the "entire trajectory," fully leveraging the smoothing effect of large number statistics and effectively overcoming occasional noise deviations caused by single vehicle trajectory jitter. Furthermore, the use of global cumulative error ensures that the output of the angle offset is based on the globally optimal fitting result of the overall geometric distribution, rather than a local extremum, thereby greatly improving the engineering robustness and accuracy of angle offset calculation in complex dynamic traffic scenarios.

[0115] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0116] The aforementioned deep learning-based millimeter-wave radar state detection device can be implemented as a computer program, which can be used in, for example... Figure 4 It runs on the computer device shown.

[0117] Please see Figure 4 , Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a terminal or a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0118] The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0119] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. When the computer program 5032 is executed, it causes the processor 502 to execute a deep learning-based millimeter-wave radar state detection method.

[0120] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0121] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a millimeter-wave radar state detection method based on deep learning.

[0122] The network interface 505 is used for network communication with other devices. Those skilled in the art will understand that the above structure is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. A specific computer device 500 may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements.

[0123] The processor 502 is used to run a computer program 5032 stored in a memory to implement the steps of a deep learning-based millimeter-wave radar state detection method provided in any of the above method embodiments.

[0124] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0125] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program may be stored in a storage medium, which is a computer-readable storage medium. The computer program is executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0126] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program. When executed by a processor, the computer program causes the processor to perform the steps of a deep learning-based millimeter-wave radar state detection method provided in any of the above method embodiments.

[0127] The storage medium is a physical, non-transient storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk, or any other physical storage medium capable of storing program code. The computer-readable storage medium can be non-volatile or volatile.

[0128] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0129] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0130] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0131] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0132] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0133] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, this invention is also intended to include these modifications and variations as long as they fall within the scope of the claims of this invention and their equivalents.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A millimeter-wave radar state detection method based on deep learning, characterized in that, include: In response to receiving the raw detection data output by the millimeter-wave radar, the raw detection data is preprocessed to obtain a time-series dataset; The time series dataset is input into a two-branch deep learning model that includes an autoencoder branch and a long short-term memory network branch. The vehicle motion trajectory of the current frame is reconstructed and restored through the autoencoder branch to obtain the restored data of the current frame. The vehicle motion trajectory of the next frame is temporally predicted through the long short-term memory network branch to obtain the predicted data of the next frame. The current frame real data and the next frame real data output by the millimeter-wave radar are obtained. Based on the difference between the current frame restored data and the current frame real data, the restoration loss is determined. Based on the difference between the next frame predicted data and the next frame real data, the prediction loss is determined. The offset confidence is determined based on the restoration loss and the prediction loss. Determine whether the offset confidence level is greater than a preset anomaly threshold; if it is greater than the preset anomaly threshold, then determine that the installation status of the millimeter-wave radar is abnormal.

2. The millimeter-wave radar state detection method based on deep learning according to claim 1, characterized in that, The preprocessing of the original detection data to obtain a time-series dataset includes: The original detection data includes the target vehicle's horizontal and vertical coordinates relative to the radar, its speed, and its heading angle; based on the speed and heading angle, the speed is decomposed into a horizontal axis velocity component and a vertical axis velocity component, and the horizontal and vertical coordinates, the horizontal axis velocity component, and the vertical axis velocity component are used as a feature set; Calculate the standard score for each feature dimension in the feature set, and remove outlier data points whose absolute value of the standard score is greater than a preset filtering threshold; Based on the mean and standard deviation of the training dataset, the remaining feature set after removing the outlier data points is standardized, and the standardized data is reshaped into a time series data format according to a preset time step to obtain the time series dataset.

3. The millimeter-wave radar state detection method based on deep learning according to claim 1, characterized in that, The process of reconstructing and restoring the vehicle motion trajectory of the current frame through the autoencoder branch to obtain the restored data of the current frame includes: The time series dataset is input into the encoder of the autoencoder branch, and the encoder extracts the data features of the time series dataset and compresses them to obtain a low-dimensional feature representation. The low-dimensional feature representation is input into the decoder of the autoencoder branch, and the decoder performs data reconstruction and restoration based on the low-dimensional feature representation to obtain the restored data of the current frame.

4. The millimeter-wave radar state detection method based on deep learning according to claim 1, characterized in that, The long short-term memory network branch includes multiple long short-term memory units and multiple fully connected layers connected in sequence; The step of predicting the vehicle trajectory for the next frame using the long short-term memory network branch to obtain the next frame prediction data includes: The temporal dataset is input into the Long Short-Term Memory (LSTM) network branch, and the multiple LSM units in the LSM network branch sequentially and incrementally extract temporal correlation features. The extracted temporal correlation features are then input into the multiple fully connected layers for dimensionality mapping in an incremental manner to obtain the next frame prediction data.

5. The millimeter-wave radar state detection method based on deep learning according to claim 1, characterized in that, Determining the offset confidence level based on the restoration loss and the prediction loss includes: The restoration loss and prediction loss of the current detection period are compared with historical statistical period data to determine the average period difference percentage. A dynamic adjustment slope that is positively correlated with the period mean difference percentage is determined. Based on the dynamic adjustment slope and a preset error threshold, the period mean difference percentage is mapped to an offset confidence level in the range of 0 to 100 using a hyperbolic tangent function.

6. The millimeter-wave radar state detection method based on deep learning according to claim 1, characterized in that, After determining that the installation status of the millimeter-wave radar is abnormal, the method further includes: The offset of the millimeter-wave radar is calculated based on the predicted data of the next frame and the actual data of the next frame, and a correction prompt message containing the offset is output.

7. The millimeter-wave radar state detection method based on deep learning according to claim 6, characterized in that, The calculation of the millimeter-wave radar offset based on the predicted data of the next frame and the actual data of the next frame includes: Based on the single-point error between the predicted data of the next frame and the actual data of the next frame, the horizontal and vertical coordinate translation offset is calculated. Based on the predicted data of the next frame, a coordinate rotation transformation is performed, and based on the global cumulative error between the predicted data of the next frame after the coordinate rotation transformation and the actual data of the next frame, the angle rotation offset is calculated. The horizontal and vertical coordinate translation offsets and the angular rotation offsets are used as the offsets.

8. The millimeter-wave radar state detection method based on deep learning according to claim 7, characterized in that, The global cumulative error between the predicted data of the next frame after coordinate rotation transformation and the actual data of the next frame, and the solution of the angle rotation offset, include: The mean square error between the coordinates of the predicted data of the next frame after coordinate rotation transformation and the coordinates of the actual data of the next frame is calculated, and the mean square errors corresponding to all trajectory points in the current detection period are accumulated and summed to obtain the global accumulated error. The angular rotation offset is solved with the goal of minimizing the global cumulative error.

9. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, can implement the method as described in any one of claims 1-8.

Images