Unsupervised learning method, electronic device and storage medium for millimeter wave radar perception tasks

Abstract

The present invention provides an unsupervised learning method, electronic device and storage medium applied to millimeter wave radar perception tasks, which can be applied to the field of millimeter wave radar technology. The method includes: using a target antenna array combination to obtain a radar echo signal, and performing signal processing on the radar echo signal to obtain a positive sample pair; using an encoder and a momentum encoder to extract features of the positive sample pair to obtain a positive sample feature pair, and projecting the positive sample feature pair to obtain a positive sample projection pair; using a predefined contrast loss function to process the positive sample projection pair and the negative sample projection pair stored in the queue memory library to obtain a contrast loss value; based on the contrast loss value, the encoder is parameterized, based on the encoder after parameter update, the momentum encoder is parameterized, and the positive sample projection pair is stored in the queue memory library; the above operations are iterated unsupervisedly until the preset training conditions are met to obtain a pre-trained encoder.

Description

Technical Field

[0001] The present invention relates to the field of millimeter wave radar technology, and more specifically to an unsupervised learning method, electronic device and storage medium applied to millimeter wave radar perception tasks. Background Art

[0002] In recent years, millimeter-wave radar has made significant progress in wireless sensing. With the development of deep learning, learning-based radar sensing has achieved encouraging results and realized innovative applications in various scenarios. Most of these innovative applications are achieved through supervised learning, which requires a large amount of labeled RF data. However, the uninterpretability of wireless data makes labeling very difficult, and this characteristic of wireless data severely limits the practical application of learning-based wireless sensing. Although unsupervised learning schemes do not rely on labeled data, they require different radar signals to be generated in order to be suitable for unsupervised positive sample pairs, which leads to large limitations in the application of unsupervised learning in millimeter-wave radar sensing, which in turn leads to the lack of versatility of millimeter-wave radar sensing schemes based on unsupervised learning and their inability to adapt to various types of wireless sensing scenarios. Summary of the invention

[0003] In view of the above problems, the present invention provides an unsupervised learning method, electronic device and storage medium applied to millimeter wave radar perception tasks.

[0004] According to a first aspect of the present invention, there is provided an unsupervised learning method applied to a millimeter wave radar perception task, comprising:

[0005] The determined target antenna array combination is used to obtain a radar echo signal corresponding to the millimeter-wave radar perception task, and the radar echo signal is processed to obtain a positive sample pair;

[0006] Using an encoder and a momentum encoder corresponding to the millimeter-wave radar perception task to extract features from the positive sample pair to obtain a positive sample feature pair, and projecting the positive sample feature pair to obtain a positive sample projection pair;

[0007] The positive sample projection pairs and the negative sample projection pairs stored in the queue memory library are processed using a predefined contrast loss function to obtain a contrast loss value;

[0008] The encoder is updated based on the contrast loss value, and the momentum encoder is updated based on the momentum method based on the updated encoder, and the positive sample projection pair is stored in the queue memory library;

[0009] Radar echo signal acquisition and signal processing operations, feature extraction and projection operations, loss calculation operations, parameter update operations, and positive sample projection pair storage operations are performed iteratively in an unsupervised manner until preset training conditions are met to obtain a pre-trained encoder.

[0010] According to an embodiment of the present invention, the above-mentioned method of using the determined target antenna array combination to obtain a radar echo signal corresponding to the millimeter wave radar perception task, and performing signal processing on the radar echo signal to obtain a positive sample pair includes:

[0011] Determine constraint parameters according to the millimeter-wave radar perception task, and select the antenna array synthesis scheme according to the constraint parameters to obtain the target antenna array combination;

[0012] The target antenna array combination is used to transmit radar signals to the target object involved in the millimeter-wave radar perception task, and the millimeter-wave radar echo signal of the target object is received;

[0013] Signal processing is performed on the radar echo signal to obtain a positive sample pair including a first component and a second component.

[0014] According to an embodiment of the present invention, the constraint parameters are determined according to the millimeter wave radar perception task, and the antenna array synthesis scheme is selected according to the constraint parameters to obtain the target antenna array combination, including:

[0015] The virtual antenna array synthesis strategy of millimeter-wave radar is analyzed based on the antenna array diagram and half-power beamwidth, and the side lobes whose amplitude difference with the main lobe of the millimeter-wave radar is within a preset decibel value range are defined as generalized grating lobes.

[0016] The minimum half-power beamwidth and the position of the generalized grating lobe closest to the main lobe are taken as constraint parameters of the millimeter-wave radar perception task. The constraint parameters are used to select the virtual antenna array synthesis strategy of the millimeter-wave radar to obtain the target antenna array combination.

[0017] According to an embodiment of the present invention, the encoder and momentum encoder corresponding to the millimeter wave radar perception task are used to extract features from the positive sample pairs to obtain positive sample feature pairs, and the positive sample feature pairs are projected to obtain positive sample projection pairs, including:

[0018] The neural network structure corresponding to the millimeter-wave radar perception task is used as an encoder, and the parameters of the encoder are transformed to obtain a momentum encoder;

[0019] Using an encoder to extract features from a first component of a positive sample pair to obtain first component features, and using a momentum encoder to extract features from a second component of a positive sample pair to obtain second component features;

[0020] Combine the first component feature with the second component feature to obtain a positive sample feature pair;

[0021] The first component feature and the second component feature in the positive sample feature pair are projected through different projection heads respectively to obtain a first component projection and a second component projection, and the first component projection and the second component projection are combined to obtain a positive sample projection pair.

[0022] According to an embodiment of the present invention, the above-mentioned use of the predefined contrast loss function to process the positive sample projection pairs and all negative sample projection pairs stored in the queue memory library to obtain the contrast loss value includes:

[0023] Initialize the parameters of the predefined contrast loss function to obtain a contrast loss function based on dot product measurement of similarity;

[0024] When the queue memory is not full, fill the unfilled part of the queue memory by random initialization, and use all negative sample projection pairs in the queue memory as the negative sample projection pair set of the current training round;

[0025] The negative sample projection pair set and the positive sample projection pair of the current training round are input into the contrast loss function based on the dot product measurement similarity for calculation to obtain the contrast loss value.

[0026] According to an embodiment of the present invention, the above-mentioned updating the parameters of the encoder based on the contrast loss value, updating the parameters of the momentum encoder based on the momentum method based on the encoder after the parameter update, and storing the positive sample projection pair in the queue memory library includes:

[0027] Based on the contrast loss value, the encoder parameters are updated to obtain an encoder with updated parameters;

[0028] The parameters of the encoder after parameter update are updated using the momentum method to obtain a momentum encoder after parameter update;

[0029] The positive sample projection pairs of the current training round are stored in the queue memory library, and the head element in the queue memory library is removed.

[0030] According to an embodiment of the present invention, the above-mentioned unsupervised learning method applied to the millimeter wave radar perception task also includes:

[0031] The millimeter-wave radar annotated data associated with the millimeter-wave radar perception task is obtained, and the millimeter-wave radar annotated data is used to adjust the parameters of the pre-trained encoder to obtain the final trained encoder.

[0032] According to an embodiment of the present invention, the above-mentioned millimeter wave radar perception task includes at least one of the following: a non-contact electrocardiogram real-time monitoring task, a human body posture estimation task and a human body contour segmentation task.

[0033] A second aspect of the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the above method.

[0034] The third aspect of the present invention further provides a computer-readable storage medium having a computer program or instructions stored thereon, which implements the steps of the above method when the computer program or instructions are executed by a processor.

[0035] The unsupervised learning method for millimeter-wave radar perception tasks provided by the present invention realizes unsupervised comparative learning of encoders related to millimeter-wave radar perception tasks by constructing positive sample projection pairs and utilizing negative sample projection pairs stored in a queue memory library, and can efficiently train related encoders; in addition, since the richness and effectiveness of the sample acquisition process are guaranteed, the method provided by the present invention can train a general encoder. At the same time, since the method provided by the present invention does not rely on wireless data tags, it can still provide efficient and accurate encoder training when there is a lack of tags or fewer tags in the downstream tasks of millimeter-wave radar perception. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0037] Figure 1 is a flowchart of an unsupervised learning method applied to a millimeter wave radar perception task according to an embodiment of the present invention;

[0038] Figure 2 is a schematic diagram of an unsupervised learning method according to an embodiment of the present invention;

[0039] Figure 3 is a detailed process diagram of an unsupervised learning method according to an embodiment of the present invention;

[0040] Figure 4 is a schematic diagram of a non-contact ECG monitoring task according to an embodiment of the present invention;

[0041] Figure 5 is a schematic diagram of a 3D human body posture estimation task according to an embodiment of the present invention;

[0042] Figure 6 is a schematic diagram of a human body contour segmentation task according to an embodiment of the present invention;

[0043] Figure 7 is a schematic diagram of the structure of an unsupervised learning device applied to a millimeter wave radar perception task according to an embodiment of the present invention;

[0044] Figure 8 is a block diagram of an electronic device suitable for implementing an unsupervised learning method applied to millimeter wave radar perception tasks according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0046] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0047] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0048] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0049] In recent years, wireless sensing technology solutions based on millimeter-wave radar have achieved many innovative applications in human posture estimation, personnel recognition, and non-contact cardiac signal monitoring. However, the above innovative applications require a large amount of annotated RF data. Unlike RGB image data in visual tasks, the uninterpretability of wireless data makes annotation extremely difficult. This feature severely limits the practical application of learning-based wireless sensing. Unsupervised learning provides a potential solution to this major challenge. As an unsupervised learning method, contrastive learning has achieved great success in the computer vision community. This method rejects different images (negative sample pairs) while attracting different views of the same image (positive sample pairs), thereby extracting more generalized effective information from unlabeled data. The key to unsupervised learning methods is to construct different representations for the same sample to extract consistent information, which is particularly widely used in the field of computer vision. For example, simple data enhancement (such as cropping, color jittering, Gaussian blur, etc.) can be applied to a single image to generate valuable positive samples.

[0050] However, these methods of constructing different views are not applicable to RF signals. Related research in this field has shown that directly applying data augmentation techniques designed for visual images to RF data leads to contrastive learning that tends to take shortcuts and thus fails to generate useful feature embeddings. To address this challenge, in the prior art, there are methods that use various signal processing techniques to create positive and negative samples, demonstrating the effectiveness of contrastive learning in the field of RF perception. However, it requires the use of different signal processing to generate effective signal representations to generate positive sample pairs, which is not applicable to all applications. Therefore, there is an urgent need for a general method that is applicable to various perception applications and hardware to construct an effective signal representation to solve the dilemma of scarce annotated data for millimeter-wave radar perception.

[0051] In order to solve the technical problems in the prior art, the present invention provides an unsupervised learning method with scene universality and applied to millimeter-wave radar perception. The method uses unlabeled wireless data for pre-training, thereby improving the fine-tuning effect of downstream tasks, especially when there is less labeled data.

[0052] In the technical solution of the present invention, the user-related radar signals and data involved are all signals and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0053] In the scenario of using radar signals and data related to the user for automated decision-making, the methods, devices, and systems provided by the embodiments of the present invention provide users with corresponding operation portals for the user to choose to agree or reject the automated decision results; if the user chooses to reject, the expert decision-making process is entered. The expression "automated decision-making" here refers to the activity of automatically analyzing and evaluating the user's posture, ECG information, contour segmentation, etc. through a computer program and making decisions. The expression "expert decision-making" here refers to the activity of making decisions by people who specialize in a certain field, have specialized experience, knowledge and skills, and have reached a certain level of professionalism.

[0054] Figure 1 4 is a flowchart of an unsupervised learning method applied to millimeter wave radar perception tasks according to an embodiment of the present invention.

[0055] like Figure 1 As shown, the above-mentioned unsupervised learning method applied to the millimeter wave radar perception task includes operations S110 to S150.

[0056] In operation S110, a radar echo signal corresponding to the millimeter wave radar perception task is acquired using the determined target antenna array combination, and signal processing is performed on the radar echo signal to obtain a positive sample pair.

[0057] In an embodiment of the present invention, before acquiring the radar echo signal, the consent or authorization of the target user may be obtained, for example, when performing human posture estimation or human contour segmentation. For example, before operation S110, a request to acquire the radar echo signal may be issued to the user. When the user agrees or authorizes that the radar echo signal may be acquired, the operation S110 is performed.

[0058] In an embodiment of the present invention, a corresponding operation entry may be provided for the user to choose to agree or reject the automated decision result. That is, before the radar echo signal of the user is processed, an instruction of the user to agree or reject the processing / decision input through the corresponding operation entry may be obtained. If the user agrees to perform the processing, the user radar echo signal is processed, that is, operations S110 to S120 are executed. If the user rejects the processing, the expert decision process is entered.

[0059] Commercial millimeter-wave radars have a multi-transmit and multi-receive configuration, but many perception tasks do not require the angular resolution that all antennas can achieve. After analyzing commercial millimeter-wave radars, some antenna array combinations can be selected to achieve perception tasks.

[0060] In operation S120, feature extraction is performed on the positive sample pairs using an encoder corresponding to the millimeter wave radar perception task and a momentum encoder to obtain positive sample feature pairs, and the positive sample feature pairs are projected to obtain positive sample projection pairs.

[0061] The encoder may optionally be a neural network for millimeter wave sensing tasks, and technicians in this field may select a suitable neural network based on the actual sensing task.

[0062] The momentum encoder has the same structure as the encoder, but the parameters are different. The momentum encoder is obtained by processing the parameters of the encoder.

[0063] In operation S130, the positive sample projection pairs and the negative sample projection pairs stored in the queue memory library are processed using a predefined contrast loss function to obtain a contrast loss value.

[0064] All projection pairs in the queue memory library are regarded as negative sample projection pairs. When calculating the contrast loss value, all projection pairs in the queue memory library are used as the negative sample projection pair set of the current training round to calculate the contrast loss value.

[0065] In operation S140, parameters of the encoder are updated based on the contrast loss value, parameters of the momentum encoder are updated based on the momentum method based on the encoder after parameter update, and the positive sample projection pair is stored in the queue memory library.

[0066] Before the positive sample projection pair is stored in the queue memory, the element at the head of the queue memory is removed.

[0067] In operation S150, radar echo signal acquisition and signal processing operations, feature extraction and projection operations, loss calculation operations, parameter update operations, and positive sample projection pair storage operations are performed in an unsupervised and iterative manner until preset training conditions are met to obtain a pre-trained encoder.

[0068] The unsupervised learning method for millimeter-wave radar perception tasks provided by the present invention realizes unsupervised comparative learning of encoders related to millimeter-wave radar perception tasks by constructing positive sample projection pairs and utilizing negative sample projection pairs stored in a queue memory library, and can efficiently train related encoders; in addition, since the richness and effectiveness of the sample acquisition process are guaranteed, the method provided by the present invention can train a general encoder. At the same time, since the method provided by the present invention does not rely on wireless data tags, it can still provide efficient and accurate encoder training when there is a lack of tags or fewer tags in the downstream tasks of millimeter-wave radar perception.

[0069] The following is a specific embodiment and combined with the attached Figure 2The unsupervised learning method for millimeter wave radar perception tasks provided by the present invention is further described in detail.

[0070] Figure 2 is a schematic diagram of an unsupervised learning method according to an embodiment of the present invention.

[0071] like Figure 2 As shown, first, the process of synthesizing the virtual antenna array of the millimeter-wave radar is analyzed to determine the constraint parameters for selecting the antenna array combination; the constraint parameters are used to generate different signal representations; the encoder is trained using the contrastive learning pre-training framework; finally, the pre-trained encoder is fine-tuned using labeled data to obtain the final trained encoder.

[0072] According to an embodiment of the present invention, the above-mentioned method of using the determined target antenna array combination to obtain a radar echo signal corresponding to a millimeter-wave radar perception task, and performing signal processing on the radar echo signal to obtain a positive sample pair includes: determining constraint parameters according to the millimeter-wave radar perception task, and selecting an antenna array synthesis scheme according to the constraint parameters to obtain a target antenna array combination; using the target antenna array combination to transmit a radar signal to a target object involved in the millimeter-wave radar perception task, and receiving the millimeter-wave radar echo signal of the target object; performing signal processing on the radar echo signal to obtain a positive sample pair including a first component and a second component.

[0073] A target antenna array combination is determined according to the constraint parameters, and a radar echo signal is obtained by using the target antenna array combination. The radar echo signal is processed to obtain a first radar echo component and a second radar echo component, and then a positive sample pair is constructed.

[0074] According to an embodiment of the present invention, the above-mentioned determining constraint parameters according to the millimeter-wave radar perception task, and selecting the antenna array synthesis scheme according to the constraint parameters to obtain the target antenna array combination includes: analyzing the virtual antenna array synthesis strategy of the millimeter-wave radar based on the antenna array diagram and the half-power beamwidth, and defining the side lobe whose amplitude difference with the main lobe of the millimeter-wave radar is within a preset decibel value range as a generalized grating lobe; using the minimum half-power beamwidth and the position of the generalized grating lobe closest to the main lobe as the constraint parameters of the millimeter-wave radar perception task, and using the constraint parameters to select the virtual antenna array synthesis strategy of the millimeter-wave radar to obtain the target antenna array combination.

[0075] The following is a specific embodiment and combined with the attached Figure 3 The selection of the above antenna array combination provided by the present invention is further explained in detail.

[0076] Figure 3 Detailed process diagram of the unsupervised learning method according to an embodiment of the present invention.

[0077] Commercial millimeter-wave radars have multiple transmit and multiple receive antenna configurations. The main function of multiple transmit and multiple receive antennas is to improve the angular resolution of the radar by increasing the aperture size of the virtual antenna array. When performing perception tasks, all the transmit and receive antennas are usually used directly. However, many tasks do not require the angular resolution that can be achieved by all antennas. Relevant technicians have found that if a portion of suitable antennas are selected and then signal processing is performed, the corresponding tasks can be completed well. Based on the above key findings, the present invention selects a portion of the transmit and receive antennas of the millimeter-wave radar, generates different signal features after signal processing, and constructs positive sample pairs for comparative learning. In addition, relevant technicians have noticed that not all antenna combinations can generate representations that are useful for downstream tasks. This is mainly because the use of some antennas will lead to the generation of grating lobes and a decrease in resolution. On the one hand, the use of some antennas may reduce the aperture of the virtual antenna array, resulting in a decrease in resolution. On the other hand, the use of some antennas may cause the distance between antennas in the virtual antenna array to be greater than 1 / 2 of the antenna wavelength, which will lead to the generation of grating lobes. In response to these two problems, the present invention proposes two constraint parameters and . represents the minimum half-power beamwidth, Represents the position of the generalized grating lobe closest to the main lobe. The present invention limits the selection of antenna combinations through these two parameters, thereby ensuring that useful signal representation can be generated.

[0078] First, the millimeter wave radar antenna synthesis is analyzed: Figure 3 As shown in the figure, by moving the receiving antenna of the MIMO antenna array according to the spacing between the transmitting antennas, the corresponding virtual antenna array can be synthesized. If all available antennas are not used directly, but a combination of some transmitting and receiving antennas is used, a variety of antenna arrays will be obtained. Using some antennas will cause the synthesized virtual array to no longer be uniform. At this time, the angular resolution formula of the radar is will no longer apply. In order to facilitate the analysis of the angular resolution and grating lobes of non-uniform antenna arrays, the antenna array diagram and half-power beamwidth are introduced here, and the generalized grating lobes are defined. For N non-directional equal-amplitude array elements, they are The vector sum of the radiation field at a point in the directional far field is shown in formula (1):

[0079] (1),

[0080] in, It represents the field strength of each antenna at that point in the far field. represents the phase difference between the nth antenna and the 0th antenna. By normalizing the above equation, we can get the radar antenna pattern. The beam width can be used as a measure to evaluate the angular resolution of the antenna array. In general, the angular resolution of the array is defined by the half-power beam width (HPBW), which is used in the following section. In order to determine , move down 3 dB from the peak, and measure the angular distance. When the antenna distribution is uneven, such as the minimum spacing is λ / 2 and the maximum spacing is 1.5λ, there is no grating lobe in the antenna radiation pattern. However, the amplitude of the side lobe is very high, even exceeding the half-power level. Although no grating lobe is generated in this case, its impact on perception is significant. Therefore, the present invention introduces the concept of "generalized grating lobe", which is defined as a side lobe with an amplitude difference of 3 dB from the main lobe. The impact of antenna array synthesis on radar mainly involves the reduction of angular resolution and the appearance of grating lobes. These two effects will cause some arrays to produce poor signal representation, thereby losing information related to downstream tasks, which will be detrimental to contrastive learning. Therefore, taking these two effects into account, the present invention designs a strategy to guide antenna array synthesis.

[0081] Secondly, determine the antenna selection constraint parameters according to the task: Figure 3 As shown, when designing positive samples, it is crucial to maintain the integrity of task-related information to ensure that sample construction does not lose information required for downstream tasks. It can be seen that antenna combinations cannot be arbitrarily selected to form positive examples. If the positive samples are not constructed properly, the neural network may learn information that is irrelevant to downstream tasks. Therefore, a direct and effective selection strategy is urgently needed. As mentioned above, the synthesis of antenna arrays mainly leads to two effects: reduced resolution and the appearance of grating lobes. Both of these effects are closely related to the characteristics of the task to be completed. Therefore, the present invention introduces two constraints for the synthesis of antenna arrays: minimum half-power beamwidth and the position of the generalized grating lobe closest to the main lobe , to guide the synthesis of antenna arrays. corresponds to the minimum angular resolution required for the task, and corresponds to the minimum angular position where a generalized grating lobe can appear for a given task. Antenna arrays that do not satisfy these constraints will not be used. This ensures that the obtained positive samples satisfy the constraints and retain information relevant to the downstream task. On the other hand, different antenna combinations will produce different observations of the same sample, making it possible to learn valid information relevant to the downstream task.

[0082] According to an embodiment of the present invention, the above-mentioned use of an encoder corresponding to a millimeter-wave radar perception task and a momentum encoder to extract features from a positive sample pair to obtain a positive sample feature pair, and projecting the positive sample feature pair to obtain a positive sample projection pair includes: using a neural network structure corresponding to the millimeter-wave radar perception task as an encoder, and transforming the parameters of the encoder to obtain a momentum encoder; using the encoder to extract features from a first component in the positive sample pair to obtain first component features, and using the momentum encoder to extract features from a second component in the positive sample pair to obtain second component features; combining the first component features and the second component features to obtain a positive sample feature pair; projecting the first component features and the second component features in the positive sample feature pair through different projection heads respectively to obtain first component projections and second component projections, and combining the first component projections and the second component projections to obtain a positive sample projection pair.

[0083] According to an embodiment of the present invention, the above-mentioned use of a predefined contrast loss function to process the positive sample projection pairs and all negative sample projection pairs stored in the queue memory library to obtain the contrast loss value includes: initializing the parameters of the predefined contrast loss function to obtain a contrast loss function based on dot product measurement of similarity; when the queue memory library is not full, filling the unfull part of the queue memory library by random initialization, and using all negative sample projection pairs in the queue memory library as a set of negative sample projection pairs for the current training round; inputting the set of negative sample projection pairs and the positive sample projection pairs of the current training round into the contrast loss function based on dot product measurement of similarity for calculation to obtain the contrast loss value.

[0084] According to an embodiment of the present invention, the above-mentioned updating of parameters of the encoder based on the contrast loss value, updating of parameters of the momentum encoder based on the momentum method based on the encoder after the parameter update, and storing the positive sample projection pairs in the queue memory library include: updating the parameters of the encoder based on the contrast loss value to obtain the encoder after the parameter update; updating the parameters of the encoder after the parameter update using the momentum method to obtain the momentum encoder after the parameter update; storing the positive sample projection pairs of the current training round into the queue memory library, and removing the head element in the queue memory library.

[0085] The following is a specific embodiment and attached Figure 3 The process of obtaining the positive sample projection pairs provided by the present invention is further described in detail.

[0086] Pre-training using constrained parameters and contrastive learning framework: This paper uses a momentum encoder and a queue memory library to construct a contrastive learning framework. By comparing positive and negative samples, effective feature embedding is extracted from the data, thereby improving the performance of downstream tasks. The pre-training process can be divided into three parts, such as Figure 3As shown in the figure: sample pair construction, feature extraction and loss calculation. (1) Sample pair construction. First, two constraint parameters are determined according to the task characteristics, and then the antenna combination that meets the conditions is determined. For each sample, the present invention randomly selects two antenna combinations to synthesize an antenna array that meets the constraints, and then uses the data of the selected antennas to perform signal processing corresponding to the task. After signal processing, a positive sample pair is obtained from the same sample. and , while the negative sample pairs are naturally formed by different samples. Since the signals of all antennas are collected during the data collection process, it is only necessary to collect data once and reuse these data to obtain different representations. (2) Feature extraction. After obtaining and After that, the present invention uses encoders respectively and Extract features and get and In order to seamlessly integrate with the contrastive learning pre-training framework, the present invention adopts the network structure commonly used in the corresponding task as the encoder. After encoding, Also projected through a projection head and . Momentum Encoder And its corresponding projection head No gradient backpropagation occurs. Their parameters Use the momentum method to update, as shown in formula (2):

[0087] (2),

[0088] in For encoder Parameters, is the momentum coefficient. (3) Loss calculation. Form a positive pair, the queue is a negative sample. The similarity between When the similarity between is low, the contrast loss will decrease. Using the dot product to measure similarity, the InfoNCE contrast loss is expressed as shown in formula (3):

[0089] (3),

[0090] in, represents the temperature parameter. After calculating the loss, Enter the queue, the first feature that enters the queue will be removed from the queue.

[0091] According to an embodiment of the present invention, the above-mentioned unsupervised learning method applied to millimeter-wave radar perception tasks also includes: obtaining millimeter-wave radar annotation data associated with the millimeter-wave radar perception task, and using the millimeter-wave radar annotation data to adjust the parameters of the pre-trained encoder to obtain the final trained encoder.

[0092] Fine-tune downstream tasks using labeled data: After pre-training, a pre-trained encoder is obtained. The encoder has a valid initial distribution and is fine-tuned using labeled data so that the pre-trained encoder can complete downstream tasks.

[0093] According to an embodiment of the present invention, the above-mentioned millimeter wave radar perception task includes at least one of the following: a non-contact electrocardiogram real-time monitoring task, a human body posture estimation task and a human body contour segmentation task.

[0094] The present invention evaluates the effectiveness and versatility of the method proposed in the present invention through three millimeter-wave radar perception tasks: non-contact ECG monitoring, human posture estimation, and human contour segmentation. The results show that the universal unsupervised learning method of radio frequency signals based on millimeter-wave radar antenna array synthesis proposed in the present invention exhibits excellent performance in multiple different application scenarios, different models, and different scales of annotated data, and can better solve the problem of small amount of annotated data in millimeter-wave radar perception.

[0095] The following is a specific embodiment and combined with the attached Figure 4 The experimental process and effects of the non-contact ECG monitoring task are explained.

[0096] Figure 4 is a schematic diagram of a non-contact ECG monitoring task according to an embodiment of the present invention.

[0097] In verifying the effect of the present invention on the non-contact ECG monitoring task: first, it is necessary to analyze the two constraint parameters of ECG monitoring. Due to the point-by-point regression of the electrocardiogram, this task focuses more on learning the distribution of time signals. Therefore, the requirements for the angular resolution of ECG monitoring are relatively low. However, too low an angular resolution will cause the monitoring process to be interfered with by other parts of the body outside the chest cavity, which is undesirable. Therefore, it is necessary to extract the signal reflection from the chest with the minimum angular resolution. When using millimeter-wave radar to acquire human ECG signals, the radar is usually about 50 cm away from the body, and the human chest is about 30 cm high and 30 cm wide. Assuming that the radar is facing directly at the center of the chest, the analysis process can be modeled as Figure 4 The base of the triangle is 30 cm and the height is 50 cm, forming an isosceles triangle with a vertex angle of Therefore, the minimum angular resolution should not be less than ,Right now Since the position of the human body relative to the radar does not change significantly, the presence of grating lobes is allowed to a certain extent in this task. In this case, the steering angle range required by the radar is relatively small, basically covering the entire chest area, about Therefore, if the grating lobe appears In addition, the ECG monitoring task will not be affected. Also take .

[0098] exist , Under the condition of , unlabeled radar data is used for pre-training. After pre-training, the model is fine-tuned using labeled data. The experimental results are shown in Figure 4 shown. Figure 4 The middle is the regression result of a sample. It can be seen that the time when each peak of the present invention occurs is more consistent with the true label. The right side is the statistical result of the ECG event timing error. It can be observed that the performance of the present invention is better than the result without pre-training. Obviously, the present invention can improve the lower limit of timing accuracy, which proves the effectiveness of the present invention.

[0099] The following is a specific embodiment and combined with the attached Figure 5 The experimental process and results of the 3D human pose estimation task are explained.

[0100] Figure 5 is a schematic diagram of a 3D human posture estimation task according to an embodiment of the present invention.

[0101] After verifying the effect of the present invention on the task of 3D human posture estimation: First, it is necessary to analyze the two constraint parameters of 3D posture estimation. For the task of 3D human posture estimation, a higher angular resolution is usually required to distinguish the positions between joints. The radar should be able to distinguish the two closest fixed joints. The two closest joints in the human body are likely to be the nose and the neck, which are about 20 centimeters apart. When estimating the posture, the distance between the radar and the person is usually within 5 meters. Figure 5 As shown, the required angular resolution can be calculated .therefore, Can be When performing pose estimation, the subject is usually moving or walking. This requires the radar to have no grating lobes over a large detection angle. Therefore, for the task of 3D pose estimation, Should be , which shows that Generalized grating lobes are not allowed in the region.

[0102] Figure 5The table shows the effects of the comparison method and the present invention without pre-training. R3D-L, R3D-B, and R3D-T correspond to ResNet3D networks of three different sizes, large, medium, and small, respectively. For 100%, 50%, and 10% labeled data, the average error of R3D-B is reduced by 12.37mm, 27.8mm, and 53.6mm, respectively, compared with training from scratch. Obviously, as the labeled training data decreases, the effectiveness of the present invention becomes more obvious. For three models of different sizes: R3D-T, R3D-B, and R3D-L, the performance of R3D-B is the best among the three models. This is due to the overfitting of R3D-L to the existing training data. The present invention improves the performance of all three models, proving its robustness to different model sizes. The overall experimental results show that the present invention can make full use of unlabeled RF signals and improve the performance of downstream tasks.

[0103] The following is a specific embodiment and combined with the attached Figure 6 The experimental process and results of the human body contour segmentation task are explained.

[0104] Figure 6 is a schematic diagram of a human body contour segmentation task according to an embodiment of the present invention.

[0105] This example verifies the effect of the present invention on the task of human body contour segmentation. The actual scenario of the human body contour segmentation task is the same as the actual scenario of the 3D human body posture estimation task, such as Figure 6 Therefore, the human silhouette is generated and It can take the same value as the 3D pose estimation, that is, and .

[0106] Without pre-training, the performance of the comparison method and the present invention in the human contour segmentation task is as follows: Figure 6 As shown. Compared with the non-pre-training and comparison methods, the present invention achieves better results on different proportions of labeled data. For 100%, 50% and 10% labeled data, the IoU of R3D-L increased by 0.0233, 0.027 and 0.1389 respectively compared with no pre-training. For different models, the performance is improved as the model size increases. Even with only 10% training data, the IoU of all models pre-trained using the present invention remains above 0.6. These results demonstrate the effectiveness of the present invention under different model sizes and labeled data ratios, especially in the case of limited labeled training data.

[0107] The above-mentioned Figures 2~6The real-life images involved in the process have been authorized by the parties concerned and processed with their permission. The entire process strictly complies with the relevant provisions of laws and regulations, and corresponding confidentiality measures have been taken. In addition, the above-mentioned real-life images are only used to illustrate the advantages and effectiveness of the method provided by the present invention and will not be used in commercial scenarios or other profit-making scenarios.

[0108] The above experiments show that the present invention performs very well in extracting effective features from unsupervised millimeter-wave radar signal pre-training. For three different millimeter-wave radar perception tasks (non-contact ECG monitoring, human posture estimation, and human contour segmentation), the present invention can improve the perception performance. The two proposed parameters can also effectively constrain the selection of antennas, ensuring the richness of samples while ensuring the validity of samples. In particular, when there are fewer labels for downstream tasks, the present invention can significantly improve the performance. Through experiments on various tasks, the present invention verifies the feasibility of using antenna array synthesis to achieve unsupervised learning of millimeter-wave radar signals. In addition, a general unsupervised learning method for millimeter-wave radar signals based on antenna array synthesis proposed in the present invention can be applied to a wide range of millimeter-wave radars and various perception tasks, effectively improving the perception performance of millimeter-wave radars based on learning under unsupervised conditions, and to a certain extent solving the problem of scarce millimeter-wave radar annotated data.

[0109] Based on the above unsupervised learning method applied to millimeter wave radar perception tasks, the present invention also provides an unsupervised learning device applied to millimeter wave radar perception tasks. Figure 7 The device is described in detail.

[0110] Figure 7 is a structural diagram of an unsupervised learning device applied to millimeter wave radar perception tasks according to an embodiment of the present invention.

[0111] like Figure 7 As shown, the unsupervised learning device 700 applied to the millimeter wave radar perception task includes a signal acquisition and processing module 710, a feature extraction and projection module 720, a loss value calculation module 730, a parameter updating module 740 and an iteration module 750.

[0112] The signal acquisition and processing module 710 is used to acquire the radar echo signal corresponding to the millimeter wave radar perception task using the determined target antenna array combination, and perform signal processing on the radar echo signal to obtain a positive sample pair. In one embodiment, the signal acquisition and processing module 710 can be used to perform the operation S110 described above, which will not be repeated here.

[0113] The feature extraction and projection module 720 is used to extract features from the positive sample pairs using the encoder and momentum encoder corresponding to the millimeter wave radar perception task to obtain positive sample feature pairs, and to project the positive sample feature pairs to obtain positive sample projection pairs. In one embodiment, the loss value calculation module 720 can be used to perform the operation S120 described above, which will not be repeated here.

[0114] The loss value calculation module 730 is used to process the positive sample projection pairs and the negative sample projection pairs stored in the queue memory library using a predefined contrast loss function to obtain a contrast loss value. In one embodiment, the loss value calculation module 730 can be used to perform the operation S130 described above, which will not be repeated here.

[0115] The parameter updating module 740 is used to update the parameters of the encoder based on the contrast loss value, update the parameters of the momentum encoder based on the momentum method based on the encoder after the parameter update, and store the positive sample projection pair in the queue memory library. In one embodiment, the parameter updating module 740 can be used to perform the operation S140 described above, which will not be repeated here.

[0116] The iteration module 750 is used to iterate the radar echo signal acquisition and signal processing operations, feature extraction and projection operations, loss calculation operations, parameter update operations, and positive sample projection pair storage operations in an unsupervised manner until the preset training conditions are met to obtain a pre-trained encoder. In one embodiment, the iteration module 750 can be used to perform the operation S150 described above, which will not be repeated here.

[0117] According to an embodiment of the present invention, any multiple modules of the signal acquisition and processing module 710, the feature extraction and projection module 720, the loss value calculation module 730, the parameter update module 740 and the iteration module 750 can be combined in one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present invention, at least one of the signal acquisition and processing module 710, the feature extraction and projection module 720, the loss value calculation module 730, the parameter update module 740 and the iteration module 750 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or can be implemented in any one of the three implementation methods of software, hardware and firmware or in any appropriate combination of any of them. Alternatively, at least one of the signal acquisition and processing module 710, the feature extraction and projection module 720, the loss value calculation module 730, the parameter update module 740 and the iteration module 750 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is executed.

[0118] Figure 8 is a block diagram of an electronic device suitable for implementing an unsupervised learning method applied to millimeter wave radar perception tasks according to an embodiment of the present invention.

[0119] like Figure 8 As shown, the electronic device 800 according to an embodiment of the present invention includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 802 or a program loaded from a storage part 808 to a random access memory (RAM) 803. The processor 801 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 801 may also include an onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0120] In RAM 803, various programs and data required for the operation of electronic device 800 are stored. Processor 801, ROM 802 and RAM 803 are connected to each other via bus 804. Processor 801 performs various operations of the method flow according to the embodiment of the present invention by executing the programs in ROM 802 and / or RAM 803. It should be noted that the program can also be stored in one or more memories other than ROM 802 and RAM 803. Processor 801 can also perform various operations of the method flow according to the embodiment of the present invention by executing the programs stored in the one or more memories.

[0121] According to an embodiment of the present invention, the electronic device 800 may further include an input / output (I / O) interface 805, which is also connected to the bus 804. The electronic device 800 may further include one or more of the following components connected to the input / output (I / O) interface 805: an input portion 806 including a keyboard, a mouse, etc.; an output portion 807 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage portion 808 including a hard disk, etc.; and a communication portion 809 including a network interface card such as a LAN card, a modem, etc. The communication portion 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output (I / O) interface 805 as needed. A removable medium 811, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 810 as needed, so that the computer program read therefrom is installed into the storage portion 808 as needed.

[0122] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiment; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present invention is implemented.

[0123] According to an embodiment of the present invention, the computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, an apparatus or a device. For example, according to an embodiment of the present invention, the computer-readable storage medium may include the ROM 802 and / or RAM 803 described above and / or one or more memories other than ROM 802 and RAM 803.

[0124] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0125] It will be appreciated by those skilled in the art that the features described in the various embodiments of the present invention may be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features described in the various embodiments of the present invention may be combined and / or combined in various ways. All of these combinations and / or combinations fall within the scope of the present invention.

[0126] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination advantageously. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.

Claims

1. An unsupervised learning method for millimeter wave radar perception tasks, characterized in that: The method comprises: Determine constraint parameters according to the millimeter wave radar perception task, and select an antenna array synthesis scheme according to the constraint parameters to obtain a target antenna array combination; Using the target antenna array combination to transmit a radar signal to a target object involved in the millimeter-wave radar sensing task, and receiving a millimeter-wave radar echo signal of the target object; Performing signal processing on the radar echo signal to obtain a positive sample pair including a first component and a second component; Using an encoder and a momentum encoder corresponding to the millimeter-wave radar perception task to extract features from the positive sample pair to obtain a positive sample feature pair, and projecting the positive sample feature pair to obtain a positive sample projection pair; Processing the positive sample projection pair and the negative sample projection pair stored in the queue memory library using a predefined contrast loss function to obtain a contrast loss value; Based on the contrast loss value, the encoder is updated with parameters, based on the encoder after parameter update, the momentum encoder is updated with parameters based on the momentum method, and the positive sample projection pair is stored in the queue memory bank; Radar echo signal acquisition and signal processing operations, feature extraction and projection operations, loss calculation operations, parameter update operations, and positive sample projection pair storage operations are performed iteratively in an unsupervised manner until preset training conditions are met to obtain a pre-trained encoder.

2. The method according to claim 1, characterized in that The constraint parameters are determined according to the millimeter wave radar perception task, and the antenna array synthesis scheme is selected according to the constraint parameters, and the target antenna array combination is obtained, including: The virtual antenna array synthesis strategy of the millimeter wave radar is analyzed based on the antenna array diagram and the half-power beam width, and the side lobe whose amplitude difference with the main lobe of the millimeter wave radar is within a preset decibel value range is defined as a generalized grating lobe; The minimum half-power beamwidth and the position of the generalized grating lobe closest to the main lobe are used as constraint parameters of the millimeter-wave radar perception task, and the virtual antenna array synthesis strategy of the millimeter-wave radar is selected using the constraint parameters to obtain a target antenna array combination.

3. The method according to claim 1, characterized in that: The encoder and momentum encoder corresponding to the millimeter wave radar perception task are used to extract features from the positive sample pair to obtain a positive sample feature pair, and the positive sample feature pair is projected to obtain a positive sample projection pair, including: Using a neural network structure corresponding to the millimeter-wave radar perception task as the encoder, and transforming parameters of the encoder to obtain the momentum encoder; Using the encoder to extract features from a first component of the positive sample pair to obtain first component features, and using the momentum encoder to extract features from a second component of the positive sample pair to obtain second component features; Combining the first component feature and the second component feature to obtain the positive sample feature pair; The first component feature and the second component feature in the positive sample feature pair are projected through different projection heads respectively to obtain a first component projection and a second component projection, and the first component projection and the second component projection are combined to obtain the positive sample projection pair.

4. The method according to claim 3, characterized in that The positive sample projection pair and all negative sample projection pairs stored in the queue memory library are processed using a predefined contrast loss function to obtain contrast loss values ​​including: Initializing parameters of the predefined contrast loss function to obtain a contrast loss function based on dot product similarity measurement; When the queue memory is not full, fill the unfilled part of the queue memory by random initialization, and use all negative sample projection pairs in the queue memory as the negative sample projection pair set of the current training round; The negative sample projection pair set and the positive sample projection pair of the current training round are input into the contrast loss function based on dot product measurement of similarity for calculation to obtain the contrast loss value.

5. The method according to claim 1, characterized in that: The step of updating the parameters of the encoder based on the contrast loss value, updating the parameters of the momentum encoder based on the momentum method based on the encoder after the parameter update, and storing the positive sample projection pair in the queue memory bank comprises: Performing parameter updating on the encoder based on the contrast loss value to obtain an encoder with updated parameters; Using the momentum method to update the parameters of the encoder after the parameter update, to obtain a momentum encoder after the parameter update; The positive sample projection pairs of the current training round are stored in the queue memory library, and the head element in the queue memory library is removed.

6. The method according to claim 1, characterized in that Also includes: Millimeter-wave radar annotation data associated with the millimeter-wave radar perception task is obtained, and the millimeter-wave radar annotation data is used to adjust parameters of the pre-trained encoder to obtain a final trained encoder.

7. The method according to claim 1, characterized in that The millimeter wave radar perception task includes at least one of the following: a non-contact electrocardiogram real-time monitoring task, a human body posture estimation task, and a human body contour segmentation task.

8. An electronic device comprising: one or more processors; a memory for storing one or more computer programs, It is characterized in that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: When the computer program or instruction is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images