Ultra wide band (UWB) radar signal super-resolution processing method and device, medium and vehicle control unit

Abstract

The invention relates to the technical field of intelligent cabins, in particular to a UWB radar signal super-resolution processing method and device, a medium and a vehicle control unit. The processing method comprises the following steps: acquiring UWB radar data; performing preprocessing on the UWB radar data; and after the preprocessed UWB radar data is processed by a deep learning network, super-resolution data are output. During deep learning network training, millimeter wave radar tensor data is used as a training target. According to the invention, based on the UWB radar, signal characteristics of the millimeter wave radar close to higher resolution can be obtained. Moreover, through mutual learning of the generative models, the generator can select a lightweight network, so that the training efficiency is improved, and the calculation complexity is reduced. The method is used for life body detection, is also suitable for application scenes related to distance measurement, speed measurement and angle measurement in an intelligent cabin, and is wide in application range.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent cockpits, and in particular to a UWB radar signal super-resolution processing method, device, medium and vehicle controller. Background Art

[0002] For cockpit radar, the mainstream solutions on the market are millimeter-wave radar and UWB radar. Relying on its frequency band advantage, millimeter-wave radar can achieve high sensitivity to micro-movements of targets in the cockpit, such as human heartbeat detection, while UWB radar operates in the centimeter-wave frequency band, and its detection performance is weaker than that of millimeter-wave radar. In addition, for the automotive-grade UWB chips currently on the market for the automotive industry, the resolution of distance and angle measurement is at a disadvantage compared to millimeter-wave radar. However, millimeter-wave radar has high costs and compliance issues. For the three operating frequency bands of millimeter-wave radar, 24GHz has been abolished in China, 77 / 79GHz is a dedicated radar for outside the car, and the 60GHz band has not yet been approved. For UWB radar, its cost is lower. If it is integrated into the UWB digital key system in the future, the cost of the cockpit radar can be greatly reduced. Therefore, the industry prefers to rely on lower-cost UWB radars and rely on technological innovation to explore more landing cockpit applications. Summary of the invention

[0003] The present invention discloses a UWB radar signal super-resolution processing method, device, medium and vehicle controller, which can enable the UWB radar to reach or approach the resolution of millimeter wave radar.

[0004] In order to achieve the above purpose, on the one hand, a UWB radar signal super-resolution processing method is provided, and the specific method is as follows:

[0005] Get UWB radar data;

[0006] Preprocess UWB radar data;

[0007] The pre-processed UWB radar data is processed by the deep learning network and output as super-resolution data;

[0008] When the deep learning network is trained, millimeter wave radar tensor data is used as a training target.

[0009] Furthermore, the UWB radar data is three-dimensional data, including:

[0010] The length of the distance dimension, the length of the time dimension and the number of UWB radar transmit and receive pairs.

[0011] Furthermore, the super-resolution data is three-dimensional data, including:

[0012] The length of the target distance dimension after super resolution, the length of the target velocity dimension after super resolution, and the length of the target angle dimension after super resolution.

[0013] Furthermore, the UWB radar data is preprocessed, and the specific method is as follows:

[0014] Perform linear interpolation processing on the distance dimension length data of the UWB radar data so that the length of the distance dimension is equal to the length of the target distance dimension data after super resolution;

[0015] Perform fast Fourier transform on the time dimension length data of the UWB radar data and convert it into target velocity dimension data, and the length is equal to the target velocity dimension length after super resolution;

[0016] The UWB radar transceiver pair number dimension data is transformed into target angle dimension length data by fast Fourier transform, and the length is equal to the target angle dimension length after super resolution.

[0017] Furthermore, the deep learning grid is an adversarial learning deep generative model, which uses a discriminator to identify the authenticity of the deep learning network output, and the discriminator uses millimeter wave tensor data as a discrimination target.

[0018] Furthermore, during training, the adversarial learning deep generative model uses a number of generative models with the same network structure, and each generative model has a random initialization parameter;

[0019] Train several generative models with the same training samples;

[0020] Feedback the learning information of the generative model that has made faster learning progress to the remaining generative models, or deep learning networks.

[0021] Furthermore, the loss function design of the deep learning network during training includes:

[0022] Pixel consistency between the output of the deep learning network model and the millimeter-wave radar tensor data;

[0023] The accuracy of the discriminator in distinguishing authenticity;

[0024] How fast the generative model learns;

[0025] After weighting, we get the loss function.

[0026] Furthermore, the method is used to distinguish life information acquired by a UWB radar in a car's smart cockpit, or distance, speed, and angle measurement information acquired by a UWB radar in a car's smart cockpit.

[0027] In order to achieve the above object, on the other hand, a UWB radar signal super-resolution processing device is provided, comprising:

[0028] Information acquisition module, to obtain UWB radar data;

[0029] Information processing module, pre-processing UWB radar data;

[0030] Information resolution module: After preprocessing UWB radar data, it is processed by deep learning network and outputs super-resolution data;

[0031] When the deep learning network is trained, millimeter wave radar tensor data is used as a training target.

[0032] In order to achieve the above object, on the other hand, a storage medium is provided, which stores a plurality of instructions, and a processor loads the plurality of instructions to execute the above UWB radar signal super-resolution processing method.

[0033] To achieve the above objectives, on the other hand, a vehicle controller is provided, which is installed on a car, and a number of UWB radars for life detection are installed in the car cabin, including the above-mentioned UWB radar signal super-resolution processing device, and / or the above-mentioned storage medium.

[0034] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0035] 1. The present invention aims to improve the accuracy of UWB radar signals. It uses a deep learning network to fit the functional relationship between UWB radar signals and millimeter wave signals, and then converts the UWB radar signals into millimeter wave radar signals with higher resolution based on the functional relationship fitted by the deep learning network.

[0036] 2. The present invention selects the adversarial learning deep generative model among many deep learning network models. Compared with other deep learning models, the adversarial learning deep generative model can realize mutual learning, and the generator selects a lightweight network, thereby improving training efficiency and reducing computational complexity.

[0037] 3. Based on the properties of the UWB radar measurement data, the processed UWB radar measurement data is not only used for life detection, but also suitable for application scenarios related to ranging, speed measurement, and angle measurement in smart cockpits, and has a wide range of uses.

[0038] It should be noted that the terms "first", "second" and similar terms used in this article are only for describing the various components in the technical solution, and do not constitute a limitation of the technical solution, nor can they be understood as an indication or suggestion of the importance of the corresponding elements; elements with "first", "second" and similar terms indicate that the corresponding technical solution contains at least one of the element. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solution of the present invention and facilitate further understanding of the technical effects, technical features and purposes of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings. The accompanying drawings constitute an essential part of the specification and are used together with the embodiments of the present invention to illustrate the technical solution of the present invention, but do not constitute a limitation to the present invention.

[0040] The same reference numerals in the accompanying drawings represent the same components, specifically:

[0041] Figure 1 Schematic diagram of the super-resolution inference process of the deep learning model.

[0042] Figure 2 Schematic diagram of the super-resolution training process of deep learning models.

[0043] Figure 3 It is a schematic diagram of the overall process of the embodiment. DETAILED DESCRIPTION

[0044] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described below are only for explaining the technical solutions of the present invention, rather than for limiting the present invention. In addition, the parts described in the embodiments or the accompanying drawings are only examples of the relevant parts of the present invention, rather than the whole of the present invention.

[0045] Embodiment 1:

[0046] A UWB radar signal super-resolution processing method is used to distinguish the life information obtained by the UWB radar in the car cabin, such as Figure 3 As shown, the specific steps are as follows:

[0047] S1. Obtain UWB radar data.

[0048] In this embodiment, the UWB radar data is three-dimensional data, including: distance dimension length, time dimension length and the number of UWB radar transceiver pairs.

[0049] Cube RAW Represents the original signal of UWB radar, whose tensor size is I u ×J u ×K u , I u Represents the length of the distance dimension, J u represents the length of the time dimension, K u Represents the number of radar transceiver pairs. Since the goal of this scheme is to map UWB radar signals into millimeter wave radar signals, the size of the target tensor is I m ×J m ×K m , I m Represents the target distance dimension length after super resolution, Jm represents the length of the target velocity dimension after super-resolution, K m Represents the length of the target angle dimension after super-resolution.

[0050] S2. Preprocess the UWB radar data.

[0051] Since the super-resolution model G0 requires the input and output tensors to be of the same size, we need to first RAW Do signal processing, the steps of signal processing are as follows:

[0052] S21. Performing linear interpolation processing on the distance dimension length data of the UWB radar data so that the length of the distance dimension is equal to the distance dimension length data of the target after super resolution.

[0053] That is, for Cube RAW Along I u Dimensions are linearly interpolated so that the length of the distance dimension is equal to I m .

[0054] S22, performing fast Fourier transform on the time dimension length data of the UWB radar data to convert it into target velocity dimension data, and the length of the data is equal to the target velocity dimension length after super resolution.

[0055] That is, for Cube RAW Along J u Perform FFT on the dimension and convert it into the velocity dimension so that its length is equal to J m .

[0056] S23, performing fast Fourier transform on the UWB radar receiving and transmitting pair quantity dimension data, and converting it into target angle dimension length data, and the length is equal to the target angle dimension length after super resolution.

[0057] That is, for Cube RAW Along K u Perform FFT on the dimension and convert it into an angle dimension so that its length is equal to K m .

[0058] The output of the super-resolution model G0 is Cube SR0 , this tensor is the result of super-resolution of the UWB radar raw data, which contains distance, speed, and angle information for subsequent target detection.

[0059] S3: The pre-processed UWB radar data is processed by the deep learning network and output as super-resolution data.

[0060] The core of this embodiment is to design an end-to-end super-resolution model. The input of the model is the UWB radar data after signal processing, and the output of the model is the UWB radar data after super-resolution processing. Figure 1 shown.

[0061] In step S3, when the deep learning network is trained, the millimeter wave radar tensor data is used as a training target.

[0062] The super-resolution data is three-dimensional data, including: the target distance dimension length after super-resolution, the target speed dimension length after super-resolution, and the target angle dimension length after super-resolution. The deep learning grid is an adversarial learning deep generation model, and the discriminator is used to identify the authenticity of the deep learning network output, and the discriminator uses millimeter wave tensor data as the discrimination target.

[0063] In this embodiment, the core of the super-resolution model G0 is its training process, such as Figure 2 As shown. The super-resolution model of this embodiment adopts a generative model based on a convolutional neural network, so a discriminator D is also required to participate in the training. The task of the generative model G0 is to map the low-resolution UWB radar tensor pixel by pixel to the high-resolution UWB radar tensor, while the task of the discriminator D is to identify the authenticity of the input. This embodiment uses the tensor data of the millimeter-wave radar as the learning target, denoted by Cube GT , whose size is I m ×J m ×K m , the goal of G0 is to make the generated Cube SR0 Get as close to the Cube as possible GT , so the probability of being identified as true after passing through the discriminator D is higher.

[0064] When training the adversarial learning deep generative model, several generative models with the same network structure are used, and the initialization parameters of each generative model are random; several generative models are trained separately with the same training samples; and the learning information of the generative model with faster learning progress is fed back to the remaining generative models, or the deep learning network.

[0065] However, in order to achieve better super-resolution results, it is usually necessary to design the generative model G0 to be very complex, which will increase the difficulty of training on the one hand and the amount of model calculation on the other hand. To solve this problem, this embodiment also proposes a mutual learning auxiliary training mechanism, which is specifically: the generative model G0 selects a lightweight convolutional neural network model, and designs N generative models G1, G2, ... G with the same network structure as G0. N ,Since the initialization parameters of each generative model are random, the learning progress of these N+1 models is not completely consistent. Based on this objective fact, during the learning process of G0, if there is a model G0 with a faster learning progress than G0 i , then G iThe learning information of G0 can be fed back to G0 to help it approach the target more efficiently. Conversely, the learning information of G0 can also be fed back to other models.

[0066] The loss function of the generative model G0 designed in this embodiment is designed as follows:

[0067] min G max D L(D,G)=αL a +βL b +γL c

[0068] in

[0069]

[0070] L b =-(Prob GT *logProb0+(1-Prob GT )*log(1-Prob0))

[0071]

[0072] α, β, γ represent L a ,L b ,L c The weights of the three loss functions. L a The results used to guide the model generation are consistent with the millimeter-wave radar signal at the pixel level; L b The results used to guide the model generation make it impossible for the discriminator D to distinguish the authenticity; L c The learning information used to guide G0 to learn the fastest learning progress among other generative models.

[0073] When the training process is completed, just deploy the trained super-resolution model G0 in Figure 1 The reasoning process shown is sufficient.

[0074] Embodiment 2:

[0075] A UWB radar signal super-resolution processing device, comprising:

[0076] Information acquisition module, to obtain UWB radar data;

[0077] Information processing module, pre-processing UWB radar data;

[0078] Information resolution module: After preprocessing UWB radar data, it is processed by deep learning network and outputs super-resolution data;

[0079] When the deep learning network is trained, millimeter wave radar tensor data is used as a training target.

[0080] It should be noted that the above embodiments are only for more clearly illustrating the technical solutions of the present invention. Those skilled in the art will understand that the implementation methods of the present invention are not limited to the above contents, and obvious changes, replacements or substitutions based on the above contents do not exceed the scope of the technical solutions of the present invention; other implementation methods will also fall within the scope of the present invention without departing from the concept of the present invention.

Claims

1. A UWB radar signal super-resolution processing method, characterized in that: The specific method is as follows: Get UWB radar data; Preprocess UWB radar data; The pre-processed UWB radar data is processed by the deep learning network and output as super-resolution data; When the deep learning network is trained, millimeter wave radar tensor data is used as a training target.

2. The UWB radar signal super-resolution processing method according to claim 1, characterized in that: The UWB radar data is three-dimensional data, including: The length of the distance dimension, the length of the time dimension and the number of UWB radar transmit and receive pairs.

3. The UWB radar signal super-resolution processing method according to claim 2, characterized in that: The super-resolution data is three-dimensional data, including: The length of the target distance dimension after super resolution, the length of the target velocity dimension after super resolution, and the length of the target angle dimension after super resolution.

4. The UWB radar signal super-resolution processing method according to claim 3, characterized in that: Preprocess the UWB radar data. The specific method is as follows: Perform linear interpolation processing on the distance dimension length data of the UWB radar data so that the length of the distance dimension is equal to the length of the target distance dimension data after super resolution; Perform fast Fourier transform on the time dimension length data of the UWB radar data and convert it into target velocity dimension data, and the length is equal to the target velocity dimension length after super resolution; The UWB radar transceiver pair number dimension data is transformed into target angle dimension length data by fast Fourier transform, and the length is equal to the target angle dimension length after super resolution.

5. The UWB radar signal super-resolution processing method according to claim 1, characterized in that: The deep learning grid is an adversarial learning deep generative model, which uses a discriminator to identify the authenticity of the deep learning network output, and the discriminator uses millimeter wave tensor data as a discrimination target.

6. The UWB radar signal super-resolution processing method according to claim 5, characterized in that: When training the adversarial learning deep generative model, several generative models with the same network structure are used, and the initialization parameters of each generative model are random; Train several generative models with the same training samples; Feedback the learning information of the generative model that has made faster learning progress to the remaining generative models, or deep learning networks.

7. The UWB radar signal super-resolution processing method according to claim 6, characterized in that: The loss function design of the deep learning network during training includes: Pixel consistency between the output of the deep learning network model and the millimeter-wave radar tensor data; The accuracy of the discriminator in distinguishing authenticity; How fast the generative model learns; After weighting, we get the loss function.

8. The UWB radar signal super-resolution processing method according to claim 1, characterized in that: The method is used to distinguish life information acquired by a UWB radar in a smart car cabin, or distance, speed and angle measurement information obtained by a UWB radar in a smart car cabin.

9. A UWB radar signal super-resolution processing device, characterized in that: include: Information acquisition module, to obtain UWB radar data; Information processing module, pre-processing UWB radar data; Information resolution module: After preprocessing UWB radar data, it is processed by deep learning network and outputs super-resolution data; When the deep learning network is trained, millimeter wave radar tensor data is used as a training target.

10. A storage medium, characterized in that: A plurality of instructions are stored, and a processor loads the plurality of instructions to execute the UWB radar signal super-resolution processing method according to any one of claims 1 to 8.

11. A vehicle controller is installed on a car, and a plurality of UWB radars for life detection are installed in the car cabin, characterized in that: It comprises the UWB radar signal super-resolution processing device as described in claim 9, and / or the storage medium as described in claim 10.

Images