UWB kicking radar detection method and system based on GRU

Abstract

The invention relates to a UWB kicking radar detection method and system based on a GRU. The method comprises the following steps: acquiring training CIR data received by UWB radar in different scenes, and marking corresponding labels for the training CIR data; the UWB radar is installed at the bottom of an automobile trunk, and the scene comprises a non-interference scene and an interference scene; taking the training CIR data with the labels as training data, inputting the training data into a prediction model based on a GRU neural network, and training the prediction model; and predicting target CIR data acquired in real time by using the trained prediction model, and judging whether to respond to the target CIR data to control the automobile trunk to finish an opening action or not according to a prediction result. The CIR data acquired by the UWB radar is classified and identified through the prediction model based on the GRU neural network, and compared with an existing technical scheme for identifying the kicking action of the trunk, the cost is lower, and the stability is higher.

Description

Technical Field

[0001] The present application relates to the field of automobile technology, and in particular to a GRU-based UWB kick radar detection method and system. Background Art

[0002] When a person is holding something in both hands, he can easily open and close the trunk door by kicking the rear bumper with his foot or sweeping his leg. This has become a comfort configuration for many car models, and there are different ways to implement the trunk kicking technology.

[0003] Most current technologies for kicking the trunk use capacitive sensors, which means placing two capacitive antennas in the rear bumper of the car. The capacitance change can be used to identify whether a person is kicking. The capacitive kicking trunk solution is relatively cheap, but because it only detects changes in capacitance, it lacks accuracy, sensitivity, and anti-interference, and is prone to false triggering or failure. Millimeter-wave radar is an alternative solution. By analyzing the millimeter-wave radar echo, it can effectively detect the kicking action, and the millimeter wave is not affected by the surrounding electromagnetic environment, and is not greatly affected by rain. However, the current vehicle-mounted millimeter-wave kick trunk radar solution is relatively expensive. Summary of the invention

[0004] Based on this, it is necessary to provide a GRU-based UWB kick radar detection method and system with low cost and precise trunk control to address the above technical problems.

[0005] In a first aspect, the present application provides a GRU-based UWB kick radar detection method. The method comprises:

[0006] Obtain the feedback signal sequence received by the UWB radar in different scenarios and the training CIR data of the corresponding feedback signal, and mark the training CIR data with corresponding labels; the UWB radar is installed at the bottom of the car trunk, and the scenarios include non-interference scenarios and interference scenarios;

[0007] The labeled training CIR data is input as training data into the prediction model based on the GRU neural network to train the prediction model;

[0008] The trained prediction model is used to predict the target CIR data collected in real time, and it is determined according to the prediction result whether to control the car trunk to open in response to the target CIR data.

[0009] In one embodiment, obtaining feedback signal sequences received by a UWB radar in different scenarios and training CIR data corresponding to the feedback signals includes:

[0010] Acquire a short pulse signal sequence emitted by a transmitting antenna of the UWB radar and a corresponding feedback signal sequence received by a receiving antenna of the UWB radar;

[0011] Based on the correlation between the short pulse signal sequence and the corresponding feedback signal sequence, the training CIR data corresponding to the feedback signal sequence is determined.

[0012] In one embodiment, after obtaining the labeled training CIR data, the method further includes:

[0013] The labeled training CIR data are divided into several batches, so that the labeled training CIR data can be input into the prediction model for training in batches;

[0014] The steps for determining the batch size include:

[0015] Obtain sample CIR data in a non-interference scenario, and obtain average power, maximum power, and minimum power based on the sample CIR data;

[0016] Collect offline CIR data of UWB radar after several kicks, count the amount of data that meets the preset conditions in the offline CIR data, and determine the batch size based on the ratio of the amount of data to the number of kicks;

[0017] Among them, the preset condition is that the power value of the offline CIR data is greater than the first threshold value or less than the second threshold value; the first threshold value is the sum of the average power and the fluctuation value, the second threshold value is the difference between the average power and the fluctuation value, and the fluctuation value is the difference between the maximum power and the minimum power.

[0018] In one embodiment, the interference scenario includes: one or at least two behavior scenarios that trigger the opening action of the car trunk.

[0019] In one embodiment, the labeled training CIR data is input as training data into a prediction model based on a GRU neural network, and training the prediction model includes:

[0020] Input the training CIR data into the prediction model and use the forward propagation algorithm to obtain the output value corresponding to the CIR data;

[0021] The loss function is used to obtain the difference between the output value and the true value;

[0022] The directional propagation algorithm is used to update the parameters of the prediction model based on the loss function.

[0023] In one embodiment, the feature dimension of the training CIR data is determined according to the sampling rate and the height of the car trunk.

[0024] In the second aspect, the present application also provides a GRU-based UWB kick radar detection device. The device includes:

[0025] The data acquisition module is used to obtain the feedback signal sequence received by the UWB radar in different scenarios and the training CIR data corresponding to the feedback signal, and mark the training CIR data with corresponding labels; the UWB radar is installed at the bottom of the trunk of the car, and the scenarios include non-interference scenarios and interference scenarios;

[0026] A model training module is used to input the labeled training CIR data as training data into a prediction model based on a GRU neural network to train the prediction model;

[0027] The prediction output module is used to predict the target CIR data collected in real time by using the trained prediction model, and determine whether to control the car trunk to open in response to the target CIR data according to the prediction result.

[0028] In a third aspect, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps in the above-mentioned GRU-based UWB kick radar detection method are implemented.

[0029] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the steps in the above-mentioned GRU-based UWB kick radar detection method when executed by a processor.

[0030] In a fifth aspect, the present application further provides a computer program product, including a computer program, which implements the steps in the above-mentioned GRU-based UWB kick radar detection method when executed by a processor.

[0031] The above-mentioned GRU-based UWB kick radar detection method and system obtains the training CIR data received by the UWB radar in different scenarios, and marks the training CIR data with corresponding labels; the UWB radar is installed at the bottom of the car trunk, and the scenarios include non-interference scenarios and interference scenarios; the labeled training CIR data is input as training data into the prediction model based on the GRU neural network to train the prediction model; the trained prediction model is used to predict the target CIR data collected in real time, and it is determined whether to control the car trunk to complete the opening action in response to the target CIR data according to the prediction results. The CIR data obtained by the UWB radar is classified and identified by the prediction model based on the GRU neural network, which is lower in cost and more stable than the existing technical solution for identifying the trunk kick action. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic diagram of a flow chart of a GRU-based UWB kick radar detection method in one embodiment;

[0033] Figure 2 is a schematic diagram of the frequency domain amplitude change corresponding to CIR data when a kicking action occurs in one embodiment;

[0034] Figure 3 A schematic diagram of the structure of a GRU neural network in one embodiment;

[0035] Figure 4 4 is a structural block diagram of a GRU-based UWB kick radar detection device in one embodiment. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0037] UWB (Ultra Wide Band) radar is a wireless communication technology that uses nanosecond pulses for data transmission. It has high stability (compared to capacitor solutions), moderate cost (more expensive than capacitor solutions and cheaper than millimeter wave solutions), and low power consumption (lower than millimeter wave radar). The UWB radar system does not require the presence of UWB tags, and its working principle is similar to that of various radars used on airplanes and cars. The UWB radar transmits a UWB pulse signal and receives the echo of the pulse signal after it is reflected by an obstacle. It determines whether there is an object (or person) near the UWB radar by analyzing the echo disturbance. Specifically, the UWB radar detects surrounding objects and their movement through the received CIR (Channel Impulse Response). When the surrounding objects move, the corresponding CIR value will change due to the Doppler effect, and the movement state can be estimated by analyzing the CIR spectrum.

[0038] The UWB radar in the present invention is installed at the bottom of the car trunk and connected to the MCU (MicroController Unit) of the vehicle. The UWB radar sends a short pulse signal outward through a transmitting antenna, and receives a feedback signal after the short pulse signal is reflected through a receiving antenna. The MCU is used to receive the signal information of the short pulse signal of the UWB radar and the signal information of the feedback signal, and execute the kick detection method of this embodiment to detect whether there is a normal kicking action.

[0039] Furthermore, the system may also include: an infrared sensor and a controller, wherein the infrared sensor is used to detect whether there is a living body in the current range, and the controller is used to open the trunk door after determining that there is a kicking action.

[0040] The present application embodiment provides a GRU-based UWB kick radar detection method, such as Figure 1 As shown, the following steps are included:

[0041] Step 102, obtaining feedback signal sequences received by the UWB radar in different scenarios and training CIR data of corresponding feedback signals, and marking corresponding labels on the training CIR data; the scenarios include non-interference scenarios and interference scenarios.

[0042] A batch of training samples are obtained based on the sending and receiving signals of the UWB radar. Each sample includes both input features, that is, the UWB radar CIR data obtained after the sending and receiving signals are processed by the algorithm, and manually or automatically marked labels. In order to facilitate subsequent model training, specific values ​​are used as labels to mark different scenarios. For example, when the detection of the kicking action of the car trunk is triggered, the label is set to 1, and when the detection of the kicking action of the car trunk is not triggered, the label is set to 0.

[0043] The scenarios include non-interference scenarios and interference scenarios. The scenario here refers to the scenario at the bottom of the car trunk. The UWB radar detects objects near the bottom of the trunk. The non-interference scenario means that no objects pass under the bottom of the car trunk, and the interference scenario means that objects pass under the bottom of the car trunk.

[0044] Furthermore, when constructing training CIR data, interference scenarios may include one or at least two behavioral scenarios that trigger the trunk of a car to open. For example, some car manufacturers set that the trunk can be opened by kicking, and the trunk can also be opened by sweeping the leg. Some car manufacturers set that the trunk can only be opened by kicking, and the foot sweep will not open it. In similar situations, CIR data without interference can be collected and marked as 0, CIR data when kicking can be collected and marked as 1, and CIR data when sweeping the foot can be collected and marked as 2. In the final application, whether to open the trunk when it is judged as 2 can be determined according to the specific needs of the car manufacturer. With this as an extension, corresponding sample data can be collected for different triggering methods when constructing training data, and corresponding annotations can be performed. By expanding the training data in the above manner, the wide applicability of this embodiment can be achieved.

[0045] Furthermore, when constructing training CIR data, interference scenarios include not only action scenarios that trigger the opening of the car trunk, but also action scenarios that do not trigger the opening of the car trunk, such as a ball rolling, a broom sweeping, etc. The CIR data corresponding to these scenarios should also be collected as training data to enrich the training data and facilitate subsequent model training to obtain more accurate parameters, thereby improving the model performance.

[0046] Step 104: input the labeled training CIR data as training data into a prediction model based on a GRU neural network to train the prediction model.

[0047] Take the training CIR data as input x t , GRU also accepts a hidden state h t-1 As input, the initial value can be initialized to an all-zero vector, so the output h can be obtained through the forward calculation formula t After passing through the hidden_size layer, it is mapped to the final output prediction value through a fully connected neural network Through the training of the neural network, the final output prediction value is infinitely close to the corresponding true value res (i.e. label).

[0048] Through continuous iteration and GRU parameter updates, the prediction result of the final output of the GRU neural network will be infinitely close to the marked label. In this way, by inputting any CIR data with prediction, the GRU neural network can predict what kind of scene it belongs to.

[0049] The acquired UWB radar CIR data is a time series. This embodiment adopts a GRU neural network, which will not cause the problem of gradient disappearance or gradient explosion compared to RNN (Recurrent Neural Network). Compared with LSTM (Long Short-Term Memory), the network model is simpler, with fewer parameters, less memory usage, and similar effects, and is more suitable for embedded development.

[0050] Step 106, using the trained prediction model to predict the target CIR data collected in real time, and judging whether to control the car trunk to open in response to the target CIR data according to the prediction result.

[0051] The target CIR data collected in real time is input into the trained prediction model to predict the scene to which the target CIR data belongs. Assuming that the scene action that triggers the opening of the car trunk is a kicking action, when the target CIR data is predicted to be a kicking action, the car trunk is triggered to open, and when the target CIR data is predicted not to be a kicking action, the car trunk remains closed.

[0052] This embodiment relies on the powerful learning ability of the GRU neural network to accurately determine the kicking action and filter out the interference action to avoid false triggering. At the same time, thanks to the anti-interference ability of the UWB radar, the control accuracy and reliability of the car trunk are effectively improved. In addition, the UWB radar has the advantage of moderate cost, which is more conducive to promotion and application.

[0053] In one embodiment, obtaining training CIR data received by a UWB radar in different scenarios includes:

[0054] A short pulse signal sequence emitted by a transmitting antenna of the UWB radar and a corresponding feedback signal sequence received by a receiving antenna of the UWB radar are obtained; and based on the correlation between the short pulse signal sequence and the corresponding feedback signal sequence, training CIR data corresponding to the feedback signal sequence is determined.

[0055] CIR data refers to the channel impulse response, which is used to describe the time domain characteristics of the signal when it is transmitted in the channel. In UWB radar, CIR data refers to the time domain response of the signal received by the receiver after the short pulse signal sent by the radar is reflected by the target. Assuming that the UWB radar transmits a pulse signal x, after being reflected by the target, the feedback signal received is y, then The CIR data is h in the formula. By processing and analyzing the CIR data, the characteristic information of the target can be extracted, such as converting the time domain data into the frequency domain data, specifically by performing a fast Fourier transform (FFT) on the CIR data to obtain the amplitude information in the frequency domain.

[0056] like Figure 2 As shown in the figure, it is the frequency domain amplitude calculated from the UWB radar CIR data. The smooth waveform box area is the waveform detected without kicking action, and the violent fluctuation area is the amplitude fluctuation detected with kicking action. When the UWB radar detects a kicking action, there will be a pulse peak at the corresponding position of the CIR data.

[0057] The training CIR data input into the prediction model is a power sequence, and the training CIR data needs to be converted into power first. After the time domain training CIR data is converted into frequency domain data, its representation is a complex number I+jQ, and the corresponding power calculation formula is: pow=I^2+Q^2.

[0058] Thus, a feasible implementation is to collect the UWB radar CIR data without interference, and calculate the pow value through the above data processing flow as the GRU input x t_0 , the corresponding label is 0. Collect the UWB radar CIR data when kicking, and calculate the pow value through the above data processing flow as the GRU input x t_1 , the corresponding label is 1. Collect the UWB radar CIR data during foot scanning, and calculate the pow value through the above data processing flow as the GRU input x t_2 , the corresponding label is 2, etc. The GRU input data is the UWB radar training CIR data collected in different scenarios, which is used as x in the GRU forward calculation formula t , including the pow corresponding to the training CIR data of different scenes (for example, the pow of the training CIR data of 4 scenes is combined into a sample sequence x t [x t_0 , x t_1 , xt_2 , x t_3 ]), the corresponding labels are input into GRU as the target values ​​of GRU, and the input label sequence and the input sample sequence need to correspond one to one.

[0059] In one embodiment, after obtaining labeled training CIR data, the method further includes: dividing the labeled training CIR data into several batches so that the labeled training CIR data is input into the prediction model for training in batches; wherein, the step of determining the batch size includes: obtaining sample CIR data in an interference-free scenario, and obtaining average power, maximum power and minimum power according to the sample CIR data; collecting offline CIR data of the UWB radar that has been kicked several times, counting the amount of data in the offline CIR data that meets preset conditions, and determining the batch size according to the ratio of the amount of data to the number of kicks; wherein, the preset condition is that the power value of the offline CIR data is greater than a first threshold value or less than a second threshold value; the first threshold value is the sum of the average power and the fluctuation value, the second threshold value is the difference between the average power and the fluctuation value, and the fluctuation value is the difference between the maximum power and the minimum power.

[0060] In GRU, batch size (batch_size) is a key hyperparameter, which refers to the number of samples used in each iteration during model training. A suitable batch size can improve training efficiency, accelerate convergence, and affect the generalization ability of the model.

[0061] In this embodiment, when there is no interference from any object at the UWB radar antenna, a number of sample CIR data are taken (for example, 1000 sample CIR data are taken, and each sample CIR data is separated by 100ms), the power of the sample CIR data is calculated, and the average power avg_pow is further calculated, and the maximum power avg_max and the minimum power avg_min are selected therefrom. Next, the offline CIR data of the UWB radar with N kicks is collected offline, and the number of times the power value in the offline CIR data is greater than the first threshold value avg_pow+(avg_max-avg_min) or less than the second threshold value avg_pow-(avg_max-avg_min) is calculated. The total number of times num accumulated is finally divided by the number of kicks N, which is the batch_size=num / N of the GRU input. Among them, avg_max-avg_min is the fluctuation value.

[0062] In other embodiments, the hyperparameters in the GRU also include the time step (time_step) and feature dimension (feature_size). Among them, when time_step = n, it is considered that each value is related to its previous n values, and can be set according to actual needs; the feature dimension is determined according to the sampling rate and the height of the car trunk. If the sampling rate of the UWB radar is 2GHZ, it means that the sample is taken once every 0.5ns, and the 0.5ns light flies 15cm. Because it is a back and forth process (first transmitting and then receiving 2 times), it represents the reflection information of the object at the position of 7.5cm. 5 CIR data represent 5 samples, corresponding to the reflection information of five positions 7.5cm-37.5cm from the UWB radar antenna. If a farther distance is to be detected, the dimension of the CIR amplitude can be increased accordingly. In this embodiment, the CIR data dimension feature_size is temporarily set to 5 according to the height of the car trunk.

[0063] Set the GRU input data dimension to [batch_size, time_step, feature_size], and construct the time series of CIR data according to this dimension. Similarly, the output data is also constructed into the above time series.

[0064] Furthermore, after constructing the time series, the method also includes normalizing and standardizing the time series data, which is a method of feature scaling and a key step in data preprocessing. After the original data is normalized, each indicator is at the same order of magnitude, which is suitable for comprehensive comparative evaluation.

[0065] Since the amplitude of the UWB radar CIR data is non-negative, all input data are scaled to the (0, 1) interval, and the implementation method is as follows: a = (a-a_min) / (a_max-a_min). Among them, a represents the CIR amplitude to be normalized in the sample data, a_min is the minimum value of the CIR amplitude data in the sample data, and a_max is the maximum value of the CIR amplitude data in the sample data.

[0066] The labeled training CIR data is input as training data into the prediction model based on the GRU neural network. The training of the prediction model includes: inputting the training CIR data into the prediction model, using the forward propagation algorithm to obtain the output value corresponding to the CIR data; using the loss function to obtain the difference between the output value and the true value; using the directional propagation algorithm to update the parameters of the prediction model based on the loss function.

[0067] like Figure 3 As shown in the figure, it is a schematic diagram of the structure of the constructed GRU neural network. In the figure, the reset gate r t , which can help capture short-term dependencies in the sequence, is calculated as follows:

[0068] r t =σ(W r x t +U r h t-1 +b r )

[0069] Among them, x t represents the input at the current moment, i.e., the training CIR data; h t-1 is the hidden state of the hidden layer at the previous moment. The hidden state acts as the memory of the neural network and contains the data information that the node has seen before; W r and U r x t and h t-1 The weight of r is the bias parameter. t and h t-1 After linear addition, the sigmoid activation function is used to calculate the reset gate output. The output value of the sigmoid activation function is between 0 and 1, which is used to select how much information to keep.

[0070] Update gate z t , which can help preserve long-term dependencies in the sequence. The calculation formula is as follows:

[0071] z t =σ(W z x t +U z h t-1 +b z )

[0072] Calculation process and r t Similar. t Indicates the input at the current moment; h t-1 is the hidden state of the hidden layer at the previous moment; W z and U z x t and h t-1 The weight of z is the bias parameter. t and h t-1 After linear addition, the sigmoid activation function is used to calculate the update gate output.

[0073] is the updated value, and the calculation formula is as follows:

[0074]

[0075] By reset gate r t 、The output of the last moment is h t-1 、 Input x at the current time tCalculated together, the tanh nonlinear activation function is used here to ensure that the output value remains between -1 and 1.

[0076] h t is the hidden state at the current moment, and the calculation formula is as follows:

[0077]

[0078] By the updated Update gate z t 、The output of the last moment is h t-1 Calculated together.

[0079] The above is the forward calculation formula of the GRU neural network.

[0080] The GRU model can set the number of neurons in the hidden layer hidden_size and the number of layers in the hidden layer num_layers. In theory, a larger number of hidden layer units can improve the expressiveness of the model and better fit the data, but it will also lead to a larger number of parameters and a larger memory footprint, so parameter adjustment is required.

[0081] Since the input is a time series [x t_0 , x t_1 , x t_2 , x t_3 ], the output is the predicted value The data is not in one dimension, so a fully connected neural network is connected after GRU to form a prediction model based on the GRU neural network, mapping it into one-dimensional data

[0082] The above describes GRU's understanding of sequence information h. t The iterative process, output features at each moment and loss function are similar to those of the recurrent neural network RNN: the softmax activation function is used, and its output result is used as the model's prediction result at moment t; the deviation information between the prediction result at moment t and the actual output is represented by mean square error or cross entropy, and the accumulation of deviations at all moments constitutes a complete loss function.

[0083] The training of the GRU network includes three steps: forward propagation, loss function calculation, and back propagation. In the forward propagation stage, the input sequence passes through the GRU unit to calculate the hidden state, and then the prediction result is obtained through the output layer. In the loss calculation stage, the difference between the predicted result and the true result is used to calculate the loss. In the back propagation stage, the gradient descent algorithm updates the weights and biases.

[0084] The training samples include both the model input and the corresponding model target value. The forward propagation of the neural network will get the model's predicted value. Taking the first moment as an example, x tis the time series sample constructed in Section 2, h t-1 Initialize the array of corresponding dimensions to all 0s, and calculate the final output h one level at a time according to the forward calculation formula. t , and then through a fully connected neural network, mapped into one-dimensional data Use a cost function to evaluate the predictions of the neural network The difference between the target value [res] and the corresponding loss function loss. In this embodiment, the mean square error (MSE) is used as the loss function to evaluate the effectiveness of the model. The calculation formula is:

[0085]

[0086] Then, the gradient descent algorithm is used for optimization. The basic form of gradient descent is:

[0087] res n+1 =res n -αloss(res)'

[0088] Among them, α is the learning rate, and loss(res)' is the derivative of the loss function. The next step is to minimize the loss function, which requires the derivative of loss. When we approach the extreme point, the value of the derivative will be infinitely close to 0. As the absolute value of the derivative becomes smaller and smaller, the subsequent update amplitude will also become smaller and smaller, and eventually it will stay at the extreme point.

[0089] The error term is reduced again and again by gradient descent until it converges to obtain the training parameters Weight and bias, thereby training the overall prediction model.

[0090] In this embodiment, the principle of GRU to suppress gradient disappearance is the same as that of LSTM. First, as the depth of back propagation increases, the relevant gradient path will still grow exponentially, but the scale is significantly smaller than that of LSTM; and its gradient calculation process still has update gates and reset gates themselves participating in gradient calculations, thereby adjusting the configuration of each gradient component. Compared with the back propagation path of LSTM, GRU greatly reduces the time and space complexity. Secondly, GRU reduces the number of updates to model parameters compared to LSTM, reducing the risk of overfitting.

[0091] In one embodiment, data acquisition and processing is a critical first step in the construction and training of a neural network, which directly affects the final learning effect and performance of the model. Data processing refers to cleaning and processing data, which can eliminate data errors and noise, improve data accuracy and reliability, and make it suitable for analysis and modeling.

[0092] Therefore, this embodiment includes the step of performing data preprocessing on the training CIR data before inputting into the prediction model, including smoothing processing and Kalman filtering processing.

[0093] The smoothing process is to remove the p maximum values ​​and q minimum values ​​of the pow values ​​cached n times each time, and then average or accumulate the remaining pow values ​​as the pow value at the current moment. Where n can be the amount of sampled data within a one-second period, and p and q are both positive integers less than n. The values ​​of n, p and q can be set according to actual needs.

[0094] Kalman filter processing uses two processes, "prediction" and "update", to optimally estimate the state of the system. The processed pow value is the weighted average of the pow prediction value and the pow observation value, where the prediction value depends on historical data and is obtained through Kalman's state prediction equation, and the observation value is the measured value at the current moment. After the prediction and update process of Kalman filter, the pow value calculated through the CIR data collected by UWB will be closer to the true value.

[0095] The present invention relies on the powerful learning ability of the GRU neural network. When the prediction model learns enough samples, it can summarize the rules and then accurately predict the newly input data. At the same time, in terms of avoiding false triggering of the UWB kick radar, what the GRU learns depends entirely on the input and output data fed to it. The CIR data when the foot is kicked and the CIR data when there is no kick and false triggering are marked with different labels to train the prediction model to accurately determine the kicking action and filter out the interfering action to avoid false triggering, thereby improving the reliability of the car trunk opening control.

[0096] In terms of functional expansion, when the judgment conditions of the UWB kick radar change, you only need to collect the corresponding sample data and change the output label during training (for example, only detect the kicking action, and the foot scan does not meet the judgment conditions). The built network model can still be reused.

[0097] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0098] Based on the same inventive concept, the embodiment of the present application also provides a GRU-based UWB kick radar detection device for implementing the GRU-based UWB kick radar detection method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more GRU-based UWB kick radar detection device embodiments provided below can refer to the limitations of the GRU-based UWB kick radar detection method above, and will not be repeated here.

[0099] In one embodiment, Figure 4 As shown, a GRU-based UWB kick radar detection device is provided, comprising:

[0100] The data acquisition module 402 is used to obtain the feedback signal sequence received by the UWB radar in different scenarios and the training CIR data corresponding to the feedback signal, and mark the training CIR data with corresponding labels; the UWB radar is installed at the bottom of the trunk of the car, and the scenarios include non-interference scenarios and interference scenarios;

[0101] A model training module 404 is used to input the labeled training CIR data as training data into a prediction model based on a GRU neural network to train the prediction model;

[0102] The prediction output module 406 is used to predict the target CIR data collected in real time by using the trained prediction model, and determine whether to control the car trunk to open in response to the target CIR data according to the prediction result.

[0103] In one embodiment, the data acquisition module 402 is also used to obtain a short pulse signal sequence emitted by the transmitting antenna of the UWB radar and a corresponding feedback signal sequence received by the receiving antenna of the UWB radar; based on the correlation between the short pulse signal sequence and the corresponding feedback signal sequence, determine the training CIR data corresponding to the feedback signal sequence.

[0104] In one embodiment, the data acquisition module 402 is also used to divide the labeled training CIR data into several batches so that the labeled training CIR data can be input into the prediction model for training in batches; wherein, the step of determining the batch size includes: obtaining sample CIR data in an interference-free scenario, and obtaining average power, maximum power and minimum power based on the sample CIR data; collecting offline CIR data of the UWB radar that has been kicked several times, counting the amount of data in the offline CIR data that meets preset conditions, and determining the batch size based on the ratio of the data amount to the number of kicks; wherein, the preset condition is that the power value of the offline CIR data is greater than a first threshold value or less than a second threshold value; the first threshold value is the sum of the average power and the fluctuation value, the second threshold value is the difference between the average power and the fluctuation value, and the fluctuation value is the difference between the maximum power and the minimum power.

[0105] In one embodiment, the interference scenario includes: a number of behavior scenarios that trigger the opening action of the car trunk.

[0106] In one embodiment, the model training module 404 is also used to input the training CIR data into the prediction model, and use the forward propagation algorithm to obtain the output value corresponding to the CIR data; use the loss function to obtain the difference between the output value and the true value; and use the directional propagation algorithm to update the parameters of the prediction model based on the loss function.

[0107] In one embodiment, the feature dimension of the training CIR data is determined according to the sampling rate and the height of the car trunk.

[0108] Each module in the above-mentioned GRU-based UWB kick radar detection device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0109] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in all the above method embodiments when executing the computer program.

[0110] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in all the above method embodiments are implemented.

[0111] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in all the above method embodiments when executed by a processor.

[0112] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0113] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited thereto. The processors involved in each embodiment provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, data processing logic devices based on quantum computing, etc., but are not limited thereto.

[0114] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A GRU-based UWB kick radar detection method, characterized in that: The method comprises: Acquire feedback signal sequences received by the UWB radar in different scenarios and training CIR data corresponding to the feedback signals, and mark the training CIR data with corresponding labels; the UWB radar is installed at the bottom of the trunk of a car, and the scenarios include non-interference scenarios and interference scenarios; Input the labeled training CIR data as training data into a prediction model based on a GRU neural network to train the prediction model; The trained prediction model is used to predict the target CIR data collected in real time, and it is determined according to the prediction result whether to control the car trunk to complete the opening action in response to the target CIR data.

2. The method according to claim 1, characterized in that The step of acquiring feedback signal sequences received by the UWB radar in different scenarios and training CIR data corresponding to the feedback signals includes: Acquire a short pulse signal sequence emitted by a transmitting antenna of the UWB radar and a corresponding feedback signal sequence received by a receiving antenna of the UWB radar; Based on the correlation between the short pulse signal sequence and the corresponding feedback signal sequence, the training CIR data corresponding to the feedback signal sequence is determined.

3. The method according to claim 1, characterized in that After acquiring the training CIR data with labels, the method further includes: Divide the labeled training CIR data into several batches, so that the labeled training CIR data are input into the prediction model for training in batches; The steps for determining the batch size include: Obtain sample CIR data in a non-interference scenario, and obtain average power, maximum power, and minimum power based on the sample CIR data; Collecting offline CIR data of the UWB radar for several kicks, counting the amount of data that meets preset conditions in the offline CIR data, and determining a batch size according to the ratio of the amount of data to the number of kicks; Among them, the preset condition is that the power value of the offline CIR data is greater than the first threshold value or less than the second threshold value; the first threshold value is the sum of the average power and the fluctuation value, the second threshold value is the difference between the average power and the fluctuation value, and the fluctuation value is the difference between the maximum power and the minimum power.

4. The method according to claim 1, characterized in that: The interference scenario includes: one or at least two behavior scenarios that trigger the car trunk to complete the opening action.

5. The method according to claim 1, characterized in that The step of inputting the labeled training CIR data as training data into a prediction model based on a GRU neural network, and training the prediction model comprises: Input the training CIR data into the prediction model, and use a forward propagation algorithm to obtain an output value corresponding to the CIR data; The difference between the output value and the true value is obtained by using the loss function; The parameters of the prediction model are updated based on the loss function using a directional propagation algorithm.

6. The method according to claim 1, characterized in that The feature dimension of the training CIR data is determined according to the sampling rate and the height of the car trunk.

7. A GRU-based UWB kick radar detection device, characterized in that: The device comprises: A data acquisition module, used for acquiring feedback signal sequences received by the UWB radar in different scenarios and training CIR data corresponding to the feedback signals, and marking corresponding labels on the training CIR data; the UWB radar is installed at the bottom of the trunk of the car, and the scenarios include non-interference scenarios and interference scenarios; A model training module, used for inputting the labeled training CIR data as training data into a prediction model based on a GRU neural network to train the prediction model; The prediction output module is used to predict the target CIR data collected in real time by using the trained prediction model, and determine whether to control the car trunk to complete the opening action in response to the target CIR data according to the prediction result.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images