Key identification method based on millimeter wave radar

Abstract

The invention relates to a key identification method based on a millimeter wave radar. The method comprises the following steps: S1, obtaining distance angle spectrograms corresponding to different time frames; s2, obtaining an effective RA matrix corresponding to the distance angle spectrogram; s3, sorting the effective RA matrixes of the time frames, carrying out peak value detection based on the sum of the energy intensities of the effective RA matrixes, taking the effective RA matrix corresponding to each peak value as a seed matrix of single-bond data, and constructing a matrix sequence by using a single seed matrix and certain effective RA matrixes before and after the seed matrix; s4, respectively extracting spatial-temporal features based on each matrix sequence, and obtaining a key classification result based on the extracted spatial-temporal features and a first classifier; s5, extracting a continuous number part as a key sequence corresponding to one word according to a continuous key classification result; s6, generating a plurality of candidate words based on each key sequence; and S7, based on the candidate words of each key sequence, obtaining a text inference result based on a large language model. Compared with the prior art, the method has the advantages of high accuracy and the like.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave radar signal processing and intelligent human-computer interaction technology, and in particular to a key recognition method based on millimeter-wave radar. Background Technology

[0002] With the widespread adoption of smart terminals and IoT devices, users' demand for contactless input and privacy-secure interaction is constantly growing. Currently, common input recognition methods include touchscreen detection, camera-based visual recognition, and voice input. However, these methods still have shortcomings in terms of privacy, environmental adaptability, and energy consumption. For example, voice input is easily affected by environmental noise, and visual recognition is limited by factors such as lighting and occlusion, both of which can potentially lead to user privacy leaks.

[0003] In recent years, millimeter-wave radar has been widely studied in fields such as gesture recognition, posture perception, and vital sign detection due to its advantages such as high resolution, strong anti-interference ability, and non-contact measurement. The introduction of MIMO (Multiple-Input Multiple-Output) radar technology has enabled radar systems to acquire higher resolution target information in the spatial dimension. However, existing research on human behavior recognition based on millimeter-wave radar mainly focuses on large-amplitude movements (such as waving and walking), and the recognition of minute vibrations or fine operations (such as keystrokes) remains challenging. The main problem is that the vibration amplitude generated by keystrokes is extremely small and the duration is short, and the signal differences between different key positions are limited, making it difficult for traditional signal processing methods to effectively distinguish them.

[0004] Furthermore, existing methods largely rely on manually designed algorithms based on energy or frequency domain features, such as short-time Fourier transform, energy envelope analysis, or peak detection. These algorithms lack robustness when dealing with micro-vibration signals in complex noisy environments and struggle to automatically learn spatial-temporal correlation features. While some research has attempted to apply convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to radar signal recognition with the development of deep learning, existing models often neglect the two-dimensional spatial characteristics of radar RA (Range-Angle) data and the inter-frame dependencies of keystroke signals, resulting in room for improvement in recognition accuracy and generalization ability.

[0005] Therefore, there is an urgent need for a technical solution that can effectively extract micro-vibration features from millimeter-wave MIMO radar signals and achieve high-precision recognition of keystrokes in multiple categories, in order to overcome the shortcomings of existing methods in terms of resolution, robustness and model generalization. Summary of the Invention

[0006] The purpose of this invention is to provide a key recognition method based on millimeter-wave radar to overcome the shortcomings of the prior art.

[0007] The objective of this invention can be achieved through the following technical solutions: A key recognition method based on millimeter-wave radar includes: Step S1: Acquire millimeter-wave radar signals and obtain range-angle spectra corresponding to different time frames through three-dimensional fast Fourier transform; Step S2: For each distance angle spectrum, obtain the corresponding effective RA matrix based on signal energy filtering; Step S3: Sort the effective RA matrices of each time frame, perform peak detection based on the sum of the energy intensities of each effective RA matrix, take the effective RA matrix corresponding to each peak as a seed matrix of single-key data, and construct a matrix sequence by combining the single seed matrix and a certain number of effective RA matrices before and after the seed matrix. Step S4: Extract spatiotemporal features based on each matrix sequence, and obtain key classification results based on the extracted spatiotemporal features using the first classifier, wherein the key classification results include numbers and function keys; Step S5: Extract the consecutive numerical parts from the continuous keystroke classification results as the keystroke sequence corresponding to a word; Step S6: Generate multiple candidate words based on each key press sequence; Step S7: Based on the candidate words of each key sequence, obtain the text inference results based on the large language model.

[0008] Step S1 includes: Step S1-1: Acquire millimeter-wave radar signals and perform three-dimensional fast Fourier transform on the millimeter-wave radar signals to obtain a range-velocity-angle data block containing target range, velocity, and angle information; Step S1-2: Sort all distance, velocity, and angle data blocks by time; Step S1-3: For the sorted distance, velocity, and angle data blocks, select the portion with a velocity dimension of 0 to generate a distance-angle spectrum.

[0009] The process of extracting the effective RA matrix from a single distance-angle spectrum in step S2 includes: Step S2-1: In the distance angle spectrum, within the first range, calculate the cumulative signal energy value of each row, where the horizontal axis is the angle dimension and the vertical axis is the distance dimension, and the value of each coordinate represents the energy intensity. Step S2-2: Select the row with the largest energy accumulation value as the center row, and take the center row, the two rows below the center row, and the three rows above the center row as the second range; Step S2-3: Within the second range, calculate the cumulative signal energy value for each column; Step S2-4: Select the column with the largest accumulated signal energy value as the center column; Step S2-5: Using the center column as the center of the sliding window, slide the sliding window to both sides respectively, and calculate the column average of the energy accumulation value in the sliding window until the column average is lower than the first proportion of the center column. The column average of the energy accumulation value is the ratio of the energy accumulation value in the sliding window to the number of columns. Step S2-6: Use the union of all sliding windows as the seed matrix, spread it evenly to the left and right sides until the number of columns reaches the second target number, and use the resulting matrix as the effective RA matrix.

[0010] The first range is the 50 rows closest to the horizontal axis.

[0011] The second target quantity is 128.

[0012] Step S6 includes: Step S6-1: For each key in the key sequence, generate multiple extended keys based on the distance between the key and other keys, and obtain multiple rearranged sequences based on all extended keys; Step S6-2: Eliminate some rearranged sequences based on the maximum distance constraint; Step S6-3: For the remaining rearranged sequence, obtain multiple candidate words based on the word frequency of each letter.

[0013] Step S6-1 includes: Step S6-1-1: Select the first key in the key sequence; Step S6-1-2: Treat all keys whose distance from the currently selected key is less than the first pre-configured threshold as extended keys corresponding to the currently selected key, and use the distance from the currently selected key as the distance value of the extended key, wherein the distance between two horizontally or vertically adjacent keys is 1, and the distance between two diagonally adjacent keys is 1.5. Step S6-1-3: Determine whether the traversal of all keys in the key sequence has been completed. If yes, proceed to step S6-1-4; otherwise, select the next key in the key sequence and return to step S6-1-2. Step S6-1-4: Take each key and its extended keys in the key sequence as the candidate set for that position in the key sequence; Step S6-1-5: Traverse the candidate set for each position of the key sequence, exhaustively enumerate all combinations to obtain multiple rearranged sequences, and take the sum of the distance values ​​of all extended keys in the rearranged sequence as the total distance of the rearranged sequence.

[0014] Specifically, step S6-2 involves removing all rearranged sequences whose total distance is greater than the second pre-configured distance threshold.

[0015] A key recognition device based on millimeter-wave radar includes a memory, a processor, and a program stored in the memory. When the processor executes the program, it implements the method described above.

[0016] A storage medium having a program stored thereon, which, when executed, implements the method described above.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Non-contact input and high privacy: Input recognition is achieved through millimeter-wave radar, eliminating the need for touch or camera capture. This avoids visual leakage and touch wear issues, significantly improving user privacy and device durability.

[0018] 2. High-sensitivity micro-motion recognition: Combining RA domain transformation and deep feature extraction model, it can capture micro-vibration signals caused by keystrokes at the millimeter level, and achieve stable detection of extremely subtle movements.

[0019] 3. Joint learning of spatial and temporal features: By integrating CNN and Transformer structures, the model can simultaneously extract local spatial features and model temporal dependencies, resulting in higher recognition accuracy and generalization performance compared to traditional methods.

[0020] 4. A certain error correction mechanism is formed by adjacent extended buttons, and the running speed can be guaranteed by the total distance constraint. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the technical principle of the present invention; Figure 2 This is a network architecture diagram of the present invention; Figure 3 This is a diagram of the search tree structure of the present invention; Figure 4 The confusion matrix for single-bond inference results; Figure 5 This is a diagram illustrating the distance between the buttons; Figure 6 This is a schematic diagram of the main steps of the method of the present invention; Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0023] A key recognition method based on millimeter-wave radar achieves a complete mapping from millimeter-wave signals to readable text through radar signal acquisition, extraction of effective spatial and temporal signals, noise suppression, key event segmentation, single-key recognition, candidate word generation, and semantic reconstruction of the entire sentence. The overall process includes, in sequence, acquiring raw RA spectrum data and constructing a matrix sequence, extracting effective signals from the target device, removing environmental noise, single-key recognition of 32 frames of key events, generating and sorting candidate words, and constructing semantically coherent sentences based on a large-scale language model, ensuring the accuracy and robustness of key recognition and text reconstruction.

[0024] like Figure 1 and Figure 6 As shown, the method includes: Step S1: Acquire millimeter-wave radar signals and obtain range-angle spectra corresponding to different time frames using three-dimensional fast Fourier transform, including: Step S1-1: Acquire millimeter-wave radar signals and perform three-dimensional fast Fourier transform on the millimeter-wave radar signals to obtain a range-velocity-angle data block containing target range, velocity, and angle information; This module employs a multiple-transmit multiple-receive (MIMO) millimeter-wave radar, transmitting a 77GHz high-frequency signal with a bandwidth of 4GHz and a sampling frequency of 10MSps. Each frame contains 32 chirs. The radar array uses a 4×4 antenna arrangement to cover the target area. The radar system uses 12 transmit and 16 receive antennas. MIMO technology theoretically allows for 192 virtual channels, but the actual antenna spacing design causes some channel overlap, ultimately resulting in 86 effective virtual channels. Combined with waveform configuration, the system achieves a range resolution of 4.3cm and an angle resolution of 1.4°. The raw data acquired by the radar is stored in binary file format, recorded by four chips. The data on each chip is arranged in I / Q dual-channel configuration according to the transmit and receive antennas. After processing the raw data using a three-dimensional fast Fourier transform (3D-FFT), a range-velocity-angle data block containing target range, velocity, and angle information, i.e., the RDA (Range-Doppler-Angle) data block, is obtained.

[0025] Step S1-2: Sort all distance, velocity, and angle data blocks by time; Steps S1-3: For the sorted distance-velocity-angle data blocks, select the portion with a velocity dimension of 0 to generate a distance-angle spectrum. Each frame of data contains multiple chirp signals, which are processed to generate RDA data blocks, arranged in chronological order. To extract micro-vibration features on stationary objects, a zero-Doppler slice is selected from the RDA cube to generate a distance-angle (RA) spectrum, used to characterize the device's micro-vibration.

[0026] Step S2: For each distance angle spectrum, obtain the corresponding effective RA matrix based on signal energy filtering; During radar signal acquisition, the detection range covered by each frame of the RA spectrum is larger than the actual reflection area of ​​the target device. Therefore, it is necessary to extract effective signals in both spatial and temporal dimensions. Based on this, the process of extracting the effective RA matrix from a single range angle spectrum in step S2 includes: Step S2-1: In the distance angle spectrum, within the first range, calculate the cumulative signal energy value of each row. In the distance angle spectrum, the horizontal axis is the angle dimension and the vertical axis is the distance dimension. The value of each coordinate represents the energy intensity. In this embodiment, the first range is the 50 rows closest to the horizontal axis. Step S2-2: Select the row with the largest energy accumulation value as the center row, and take the center row, the two rows below the center row, and the three rows above the center row as the second range; Step S2-3: Within the second range, calculate the cumulative signal energy value for each column; Step S2-4: Select the column with the largest accumulated signal energy value as the center column; Step S2-5: Using the center column as the center of the sliding window, slide the sliding window to both sides respectively, and calculate the column average of the energy accumulation value in the sliding window until the column average is lower than the first proportion of the center column. The column average of the energy accumulation value is the ratio of the energy accumulation value in the sliding window to the number of columns. In this embodiment, the first proportion is selected as 0.2 to 0.4, preferably 0.3. Step S2-6: Use the union of all sliding windows as the seed matrix, and spread it evenly to the left and right sides until the number of columns reaches the second target number. The resulting matrix is ​​the effective RA matrix. In this embodiment, the second target number is 128, so the effective RA matrix is ​​a 6×128 matrix.

[0027] The above method of constructing an effective RA matrix helps to improve subsequent single-bond identification in terms of resolution, robustness, and model generalization. 1. Positive effect on resolution: Improves spatial focusing ability and enhances the accuracy of micro-vibration detection. Key Mechanism: Precise spatial focusing on the target device region is achieved through the cumulative calculation of rows and energy, and the positioning of the center row / column. While the distance-angle spectrum itself has high resolution, the full spectrum contains a large number of irrelevant regions. By selecting the local regions with the strongest energy, background interference is effectively removed, allowing the model to focus on the micro-vibration signals caused by millimeter-level keystrokes.

[0028] This spatial focusing improves the spatial resolution of the signal, enabling subsequent feature extraction to capture subtle displacements and angular changes in the keys more precisely. For example, in single-key inference, the model can distinguish subtle differences between adjacent keys (such as the number keys "1" and "2"), thereby improving classification accuracy.

[0029] 2. Positive effects on robustness: Suppresses noise and environmental changes, enhances signal stability. Key mechanisms: Dynamic sliding window and energy threshold filtering are employed. By calculating the column mean within the sliding window and comparing it with the proportion of the central column, the effective area is adaptively determined, avoiding the sensitivity of fixed thresholds to environmental changes. Simultaneously, the diffusion process of the seed matrix ensures data width consistency, reducing signal fluctuations caused by minor device position shifts or changes in illumination.

[0030] This method improves the system's robustness in complex environments. For example, in scenarios with static reflection noise or multiple people interfering, it can stably extract target vibration features and prevent false detections. Combined with a sliding window background mean denoising algorithm, it further suppresses environmental noise, ensures the reliability of key event segmentation, and avoids the susceptibility to interference inherent in visual or touch-based methods.

[0031] 3. Positive impact on model generalization ability: Provides standardized input, promoting consistency in feature learning. Key mechanism: Outputting a standardized and efficient RA matrix provides input data in a unified format for subsequent deep learning models (such as the CNN-Transformer architecture). Through energy accumulation and sliding window diffusion, the most representative signal regions in space are extracted. These regions contain common features of keystroke behavior (such as vibration patterns and energy distribution), rather than noise specific to a particular device or user.

[0032] Standardized matrices make it easier for models to learn general patterns across different scenarios, enhancing their generalization ability. For example, during training, models learn spatiotemporal features based on such matrices, adapting to different keyboard layouts or user habits; during inference, preprocessing maintains feature consistency even with slight changes in the input signal. As the front end for spatial feature extraction, it lays the foundation for the temporal modeling of the Transformer, thereby improving overall generalization performance. The process of extracting the effective RA matrix described above, through energy filtering, dynamic sliding windowing, and output standardization, plays a crucial role in resolution, robustness, and generalization ability. It not only improves the accuracy of single-key recognition but also provides reliable input for subsequent word generation and semantic inference, ultimately achieving efficient mapping from millimeter-wave signals to text. Through these designs, this application overcomes the limitations of existing technologies in micro-vibration recognition and environmental adaptability, demonstrating the advanced nature and practicality of non-contact input.

[0033] Step S3: Sort the effective RA matrices of each time frame, perform peak detection based on the sum of the energy intensities of each effective RA matrix, take the effective RA matrix corresponding to each peak as a seed matrix of single-key data, and construct a matrix sequence by combining the single seed matrix and a certain number of effective RA matrices before and after the seed matrix. In this embodiment, the first 7 valid RA matrices of the seed matrix, the seed matrix itself, and the last 16 valid RA matrices of the seed matrix are used to construct a 32-frame continuous 6×128 matrix sequence.

[0034] Step S4: Extract spatiotemporal features based on each matrix sequence, and obtain key classification results based on the extracted spatiotemporal features using the first classifier. The key classification results include numbers and function keys. In this embodiment, to extract the target vibration signal and suppress environmental noise, a sliding window background mean removal algorithm is used on the 32-frame matrix sequence. The first 10 frames of RA spectrum are selected as the initial window. The mean within the window is calculated and subtracted to obtain the initial background. Subsequently, the window slides with the frame sequence, and the window mean is recalculated and the new frame background is subtracted for each frame until all 32 frames of data are covered. Finally, the mean of the last window is used as the global background, and all 32 frames of the matrix sequence are uniformly subtracted to obtain the denoised RA signal matrix, which preserves the target vibration characteristics while suppressing static object reflection noise.

[0035] In the key event processing stage, to achieve single-key and whole-sentence input inference, the radar input is first segmented into independent key events. The Automatic Multi-Scale Peak Detection (AMPD) algorithm is used to identify the temporal energy peaks in the RA spectrum sequence, with each peak corresponding to a key operation. During text inference, only letter keys are retained, while function keys (such as space, delete, and enter) are removed. These function keys are identified by the time interval between key presses, thus constructing a continuous and clean key sequence for subsequent word and sentence reconstruction module input.

[0036] The single-bond inference module first processes 32 consecutive frames of 6×128-dimensional RA spectrum data. For example... Figure 2As shown, spatial features are extracted from each frame of data through a Convolutional Neural Network (CNN) module. The CNN module employs a hierarchical design, where the basic unit, SingleConvBlock, consists of a 2D convolutional layer (Conv2d), a batch normalization layer (BN), and a ReLU activation function. Two SingleConvBlocks and one max-pooling layer (Maxpool2d) constitute a DoubleConvBlock module. The entire CNN module is composed of three cascaded DoubleConvBlocks and one SingleConvBlock, used to process single-frame RA spectrum data. A similar function to Keras' TimeDistributed is implemented in the PyTorch framework to ensure that each of the 32 frames undergoes the same feature extraction process. After the CNN module outputs a 32×6×8 tensor, it is input into a three-layer stacked Transformer encoder for temporal dependency modeling. The Transformer adds positional information to the sequence through sine and cosine position encoding and uses a multi-head attention mechanism to extract temporal features. The encoding dimension is aligned with the number of CNN output channels, enabling direct processing of feature maps. Finally, the temporal features are mapped to 10 prediction results through a fully connected layer. The network is trained using the cross-entropy loss function to achieve accurate classification and recognition of single keys.

[0037] Step S5: Extract the consecutive numerical parts from the continuous keystroke classification results as the keystroke sequence corresponding to a word; Step S6: Generate multiple candidate words based on each key press sequence, including: Step S6-1: For each key in the key sequence, generate multiple extended keys based on the distance between the key and other keys. Based on all extended keys, obtain multiple rearranged sequences, such as... Figure 3 and Figure 4 As shown, it specifically includes: Step S6-1-1: Select the first key in the key sequence; Step S6-1-2: Treat all keys whose distance to the currently selected key is less than a first pre-configured threshold as extended keys corresponding to the currently selected key, and use the distance to the currently selected key as the distance value of the extended key, where, for example... Figure 5 As shown, the distance between two horizontally or vertically adjacent buttons is 1, and the distance between two diagonally adjacent buttons is 1.5. Step S6-1-3: Determine whether the traversal of all keys in the key sequence has been completed. If yes, proceed to step S6-1-4; otherwise, select the next key in the key sequence and return to step S6-1-2. Step S6-1-4: Take each key and its extended keys in the key sequence as the candidate set for that position in the key sequence; Step S6-1-5: Traverse the candidate set for each position of the key sequence, exhaustively enumerate all combinations to obtain multiple rearranged sequences, and take the sum of the distance values ​​of all extended keys in the rearranged sequence as the total distance of the rearranged sequence.

[0038] Step S6-2: Eliminate some rearranged sequences based on the maximum distance constraint, specifically: eliminate all rearranged sequences whose total distance is greater than the second pre-configured distance threshold.

[0039] Step S6-3: For the remaining rearranged sequence, obtain multiple candidate words based on the word frequency of each letter.

[0040] In this embodiment, candidate words are sorted using both edit distance and word frequency. First, they are sorted by edit distance, and then, if the edit distances are the same, they are sorted by word frequency to prioritize frequently used words. Finally, the top K candidate words are selected from the sorted results as the output, providing an input sequence for text reconstruction and achieving efficient and accurate sentence inference.

[0041] Step S7: Based on the candidate words of each key sequence, obtain the text inference results based on the large language model.

[0042] In the text inference process, Top-K candidate words are input into a Large Language Model (LLM) to construct semantically coherent complete sentences. The model employs a self-generated suggestion strategy, iteratively optimizing suggestions based on previously generated results and user context to guide the model in selecting the optimal word arrangement, ensuring semantic and grammatical consistency in the output sentence. This method achieves accurate mapping from keystroke sequences to text by constraining word selection and arrangement order, and improves the robustness and accuracy of whole-sentence inference.

[0043] The following experiments verify the proposed solution. This application achieves high-resolution spatial and temporal positioning by accurately capturing key vibrations of target devices using multiple-transmit multiple-receive (MIMO) millimeter-wave radar. Compared to traditional optical or touch recording methods, this invention can acquire complete user operation data in a non-contact state, effectively avoiding visual obstruction and interference problems. By extracting a spatial 6×128 matrix and performing time series analysis on 32 frames of RA spectrum for each key press, single-key operations can be accurately identified (as shown in Table 1), with an accuracy exceeding 77%. Furthermore, the reconstruction of entire sentences from letter key press sequences exhibits high semantic fidelity. The ranking of Top-K candidate words is optimized by combining edit distance and word frequency, effectively improving the accuracy of text inference.

[0044] Table 1 Table 2 Meanwhile, the sliding window background mean denoising algorithm of this invention can dynamically suppress environmental noise and static reflection signals, ensuring the integrity and continuity of the target vibration features. In the text mapping process from single keystrokes to complete sentences, a Trie dictionary tree and edit distance pruning strategy are adopted to achieve fault tolerance and efficient search, resulting in fast candidate word generation and minimal redundant computation. Combined with the semantic reconstruction strategy of a large-scale language model, it can automatically correct minor keystroke errors while maintaining semantic coherence, achieving high-accuracy reconstruction of complete sentences. Table 2 shows the inference results for texts of different lengths.

[0045] This invention has a wide range of applications, including but not limited to key input recognition, non-contact input behavior monitoring, text input assistance systems, keyboard operation behavior analysis, and security monitoring for touch devices such as smartphones and tablets. This technology can acquire operation information without the user's awareness, and features high precision, non-contact operation, real-time performance, and adaptability to complex environments, making it highly feasible and worthy of widespread adoption.

[0046] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a 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, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for key recognition based on millimeter wave radar, characterized by, The method comprises the following steps: Step S1: acquiring millimeter wave radar signals, and obtaining distance-angle spectrum diagrams corresponding to different time frames through three-dimensional fast Fourier transform; Step S2: obtaining corresponding effective RA matrices of each distance-angle spectrum diagram based on signal energy screening; Step S3: sorting the effective RA matrices of each time frame, performing peak detection based on the sum of the energy intensities of the effective RA matrices, taking the effective RA matrices corresponding to each peak as a seed matrix of a single key data, and constructing a matrix sequence with the single seed matrix and a certain number of effective RA matrices before and after the seed matrix; Step S4: extracting space-time features based on each matrix sequence, and obtaining key classification results based on the first classifier based on the extracted space-time features, wherein the key classification results include numeric keys and function keys; Step S5: extracting a continuous numeric part as a key sequence corresponding to a single word from continuous key classification results; Step S6: generating multiple candidate words based on each key sequence; Step S7: obtaining a text inference result based on a large language model based on the candidate words of each key sequence. 2.The key recognition method based on millimeter wave radar according to claim 1, characterized in that, The step S1 comprises the following steps: Step S1-1: acquiring millimeter wave radar signals, and performing three-dimensional fast Fourier transform on the millimeter wave radar signals to obtain distance-velocity-angle data blocks containing target distance, velocity and angle information; Step S1-2: sorting all distance-velocity-angle data blocks according to time; Step S1-3: generating a distance-angle spectrum diagram by selecting the part with a velocity of 0 from the sorted distance-velocity-angle data blocks. 3.The key recognition method based on millimeter wave radar according to claim 1, characterized in that, The process of extracting the effective RA matrix from a single distance-angle spectrum diagram in step S2 comprises the following steps: Step S2-1: in the distance-angle spectrum diagram, calculate the signal energy accumulation value of each row in the first range, wherein the horizontal axis of the distance-angle spectrum diagram is the angle dimension, the vertical axis is the distance dimension, and the value of each coordinate represents the energy intensity; Step S2-2: select the row with the maximum energy accumulation value as the center row, and select the center row and two rows below the center row and three rows above the center row as the second range; Step S2-3: calculate the signal energy accumulation value of each column in the second range; Step S2-4: select the column with the maximum signal energy accumulation value as the center column; Step S2-5: take the center column as the center of the sliding window, slide the sliding window to both sides respectively, calculate the column mean of the energy accumulation value in the sliding window, and stop until the column mean is lower than the first proportion of the center column, wherein the column mean of the energy accumulation value is the ratio of the energy accumulation value in the sliding window to the number of columns; Step S2-6: take the union of all sliding windows as a seed matrix, average diffuse to both sides respectively until the number of columns reaches a second target number, and take the obtained matrix as an effective RA matrix. 4.The key recognition method based on millimeter wave radar according to claim 3, characterized in that, The first range is the nearest 50 rows close to the horizontal axis. 5.The key recognition method based on millimeter wave radar according to claim 3, characterized in that, The second target number is 128. 6.The key recognition method based on millimeter wave radar according to claim 1, wherein, The step S6 comprises the following steps: Step S6-1: for each key in the key sequence, generate multiple extended keys according to the distance between the key and other keys, and obtain multiple reordered columns based on all the extended keys; Step S6-2: eliminate part of the reordered columns based on the maximum distance constraint; Step S6-3: For the remaining reordering sequences, a plurality of candidate words are obtained based on the word frequency of each letter. 7.The key recognition method based on millimeter wave radar according to claim 6, wherein, The step S6-1 includes: Step S6-1-1: Selecting the first key in the key sequence; Step S6-1-2: Selecting all keys with a distance less than a first preconfigured threshold to the currently selected key as extended keys corresponding to the currently selected key, and taking the distance to the currently selected key as the distance value of the extended key, wherein the distance between two horizontally or vertically adjacent keys is 1, and the distance between two diagonally adjacent keys is 1.5; Step S6-1-3: Determining whether the traversal of all keys in the key sequence is completed, if yes, executing step S6-1-4, otherwise, selecting the next key in the key sequence and returning to step S6-1-2; Step S6-1-4: Taking each key in the key sequence and its extended keys as the candidate set at the position in the key sequence; Step S6-1-5: Traversing the candidate set at each position in the key sequence, and obtaining a plurality of reordering sequences by enumerating all combination cases, and taking the sum of the distance values of all extended keys in the reordering sequence as the total distance of the reordering sequence. 8.The key recognition method based on millimeter wave radar according to claim 7, wherein, The step S6-2 is specifically: removing all reordering sequences with a total distance greater than a second preconfigured distance threshold. 9.A millimeter wave radar-based key recognition device, comprising a memory, a processor, and a program stored in the memory, wherein, The processor implements the method of any one of claims 1-8 when executing the program.

10. A storage medium having stored thereon a program, characterized by The program is executed to implement the method of any one of claims 1-8.

Images