Personnel positioning method and device used in indoor environment and computer readable storage medium

Abstract

The invention relates to the technical field of smart home, and discloses a personnel positioning method used in an indoor environment, and the method comprises the steps: obtaining an original echo signal collected by a low-resolution millimeter wave radar in the indoor environment; performing front-end signal enhancement processing on the original echo signal to obtain the distance between the target object and the radar, the radial speed and the existence probability of the target object; based on the distance between the target object and the radar, the radial speed and the existence probability of the target object, performing high-resolution angle estimation on the target object meeting a preset existence condition so as to calculate an azimuth angle and a pitch angle of the target object; and fusing the azimuth angle, the pitch angle, and the distance and the radial speed corresponding to the target object meeting the preset existence condition, and generating a continuous trajectory of the target object in the three-dimensional space through trajectory tracking processing. Therefore, accurate, continuous and stable three-dimensional positioning can be carried out on the personnel. The invention further discloses a personnel positioning device used in the indoor environment and a computer readable storage medium.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart home, for example, to a personnel positioning method and device in an indoor environment and a computer readable storage medium. BACKGROUND

[0002] At present, low-resolution millimeter wave radar shows application potential in the field of indoor personnel positioning due to its low hardware cost and good power consumption control. However, its angular resolution is limited by the physical antenna array size, and at the same time, its distance resolution is also limited to about 1.2 meters under the indoor compliance requirement. In a complex indoor scene, dense multi-target, significant multipath effect and frequent occlusion problem together cause distortion, attenuation and even loss of radar echo signals, so that the positioning system based on the traditional signal processing method faces the technical difficulties of target confusion, track interruption and sharp decline in measurement accuracy.

[0003] In order to improve the performance of low-resolution radar, the related technology discloses a deep learning-based solution, including: processing sparse radar data through a deep recurrent network to form time-correlated trajectories, and using a classifier to realize user identification and intrusion detection. The core of this solution is to use a deep model to learn and extract complex nonlinear features from low-quality raw data to make up for the inherent shortcomings of hardware resolution.

[0004] In the process of implementing the embodiments of the present disclosure, it is found that at least the following problems exist in the related art: Although the related technology realizes processing capability beyond the physical resolution at the data level, the deep model is highly dependent on large-scale and high-quality training data, and its performance is extremely sensitive to data preprocessing quality, and its generalization ability is limited in actual dynamic environment. Moreover, this method cannot break through the inherent information missing bottleneck of low-resolution radar in the front-end signal acquisition stage, resulting in poor positioning stability in real complex scenes with serious multipath interference and dense target intersection.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present application, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] In order to have a basic understanding of some aspects of the disclosed embodiments, the following is a simple summary. The summary is not a general review, nor is it intended to determine the key / important components or delineate the scope of protection of these embodiments, but as a prelude to the detailed description below.

[0007] This disclosure provides a method, apparatus, and computer-readable storage medium for personnel positioning in indoor environments, offering a technical solution that overcomes the hardware limitations of low-resolution millimeter-wave radar, suppresses multipath and obstruction interference, and enables accurate, continuous, and stable positioning of multiple targets in complex indoor environments.

[0008] In some embodiments, the method for personnel positioning in an indoor environment includes: acquiring raw echo signals collected by a low-resolution millimeter-wave radar in an indoor environment; performing front-end signal enhancement processing on the raw echo signals to obtain the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence; based on the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence, performing high-resolution angle estimation on the target object that meets preset existence conditions to calculate its azimuth and elevation angles; fusing the azimuth angle, elevation angle, and the distance and radial velocity corresponding to the target object that meets the preset existence conditions, and generating a continuous trajectory of the target object in three-dimensional space through trajectory tracking processing.

[0009] In some embodiments, the personnel positioning device for indoor environments includes: an acquisition module configured to acquire raw echo signals collected by a low-resolution millimeter-wave radar in an indoor environment; a processing module configured to perform front-end signal enhancement processing on the raw echo signals to obtain the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence; a calculation module configured to perform high-resolution angle estimation on target objects that meet preset existence conditions based on the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence, to calculate its azimuth and elevation angles; and a fusion module configured to fuse the azimuth angle, elevation angle, and the distance and radial velocity corresponding to the target object that meet the preset existence conditions, and generate a continuous trajectory of the target object in three-dimensional space through trajectory tracking processing.

[0010] In some embodiments, the personnel positioning device for indoor environments includes a processor and a memory storing program instructions, characterized in that the processor is configured to execute the aforementioned personnel positioning method for indoor environments when running the program instructions.

[0011] In some embodiments, the computer-readable storage medium stores program instructions that, when executed, cause the computer to perform the aforementioned method for locating people in an indoor environment.

[0012] The personnel positioning method, apparatus, and computer-readable storage medium for indoor environments provided in this disclosure can achieve the following technical effects: This solution estimates the target's distance, velocity, and probability of existence through front-end signal enhancement processing, effectively suppressing signal distortion and attenuation caused by multipath and obstruction. It then uses this prior information to compress the search space and perform high-resolution angle estimation, overcoming the hardware limitations of low-resolution radar. Finally, by fusing multi-dimensional observation information for trajectory tracking, it solves the problems of target confusion, trajectory interruption, and poor positioning stability caused by missing front-end information, achieving accurate, continuous, and stable three-dimensional positioning of personnel in complex indoor environments.

[0013] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0014] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a schematic diagram of a personnel positioning method for an indoor environment provided by an embodiment of this disclosure; Figure 2 This is a schematic diagram of a method for outputting distance, radial velocity, and existence probability provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram of a method for calculating azimuth and elevation angles provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of a method for generating continuous trajectories provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of a personnel positioning device for indoor environments provided in an embodiment of this disclosure; Figure 6 This is a schematic diagram of another personnel positioning device for indoor environments provided in this embodiment of the disclosure. Detailed Implementation

[0015] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0016] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0017] Unless otherwise stated, the term "multiple" means two or more.

[0018] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0019] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0020] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0021] Combination Figure 1 As shown, this disclosure provides a method for locating people in an indoor environment, including: S11, the server acquires the raw echo signal collected by the low-resolution millimeter-wave radar in an indoor environment.

[0022] S12, the server performs front-end signal enhancement processing on the original echo signal to obtain the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence.

[0023] S13, the server performs high-resolution angle estimation on target objects that meet preset existence conditions based on the distance between the target object and the radar, the radial velocity, and the probability of the target object's existence, in order to calculate its azimuth and elevation angles.

[0024] S14, the server integrates the azimuth angle, pitch angle, and distance and radial velocity of the target object that meet the preset existence conditions, and generates a continuous trajectory of the target object in three-dimensional space through trajectory tracking processing.

[0025] In this solution, the server can acquire raw echo signals from low-resolution millimeter-wave radar collected in an indoor environment. The server is deployed within smart home appliances. Alternatively, it can be a cloud server. The raw echo signals include two types of raw signals: range-Doppler profile data and angle channel data output from the radar front-end. Here, the range-Doppler profile data characterizes the energy distribution of the target in the range and velocity dimensions, while the angle channel data contains phase difference information between multiple antenna channels for angle estimation. This solution provides a high-quality data foundation for subsequent signal processing, effectively supporting accurate personnel positioning in complex indoor environments.

[0026] Furthermore, the server can perform front-end signal enhancement processing on the raw echo signal. First, the raw signal, containing range-Doppler profiles and angle channels, is converted into a three-dimensional data matrix fusing range, velocity, and angle information. Then, this three-dimensional data matrix is ​​input into a trained deep learning model. The deep learning model extracts the target's motion features in the spatial and temporal dimensions, ultimately outputting the precise distance, radial velocity, and presence probability between the target and the radar. This effectively improves the robustness of extracting target motion parameters from low-quality echoes, providing reliable input for subsequent positioning.

[0027] Furthermore, the server first filters out high-confidence targets from the target objects based on their existence probability. Then, it uses the range and velocity information corresponding to these high-confidence targets to compress the angle search space. Finally, within the compressed search space, a two-dimensional MUSIC algorithm incorporating temporal consistency constraints is used to calculate the target's azimuth and elevation angles. In this way, multipath interference and spurious spectral peaks are effectively suppressed through prior information guidance and motion constraints, significantly improving the angle resolution accuracy and estimation stability of low-resolution radar in complex scenarios.

[0028] Furthermore, the server integrates azimuth, elevation, and distance and radial velocity corresponding to the target object that meet preset existence conditions, and generates a continuous trajectory of the target object in three-dimensional space through trajectory tracking processing. This process first acquires the dynamic state vector of the established target trajectory and predicts the state at the next moment based on this. Then, it fuses the three-dimensional position observations obtained from azimuth, elevation, and distance conversions with the radial velocity observations, and performs data association matching between the fused observations and the predicted state. According to the matching result, when the observation successfully matches the predicted state of the target trajectory, the dynamic state vector of the trajectory is updated using the fused observations, and the updated state is used as the current point of the trajectory. When the observation does not successfully match any trajectory, it is used as the starting point of a new trajectory, and the corresponding dynamic state vector is initialized. In this way, multi-source information fusion and dynamic state estimation effectively suppress observation noise and abnormal data jumps, achieving accurate construction of continuous and stable trajectories for multiple targets in a low-resolution radar environment.

[0029] The personnel positioning method, apparatus, and computer-readable storage medium for indoor environments provided in this disclosure estimate the target's distance, velocity, and probability of existence through front-end signal enhancement processing, effectively suppressing signal distortion and attenuation caused by multipath and obstruction. Furthermore, this prior information is used to compress the search space and perform high-resolution angle estimation, overcoming the hardware limitations of low-resolution radar. Finally, by fusing multi-dimensional observation information for trajectory tracking, the problems of target confusion, trajectory interruption, and poor positioning stability caused by missing front-end information are solved, achieving accurate, continuous, and stable three-dimensional positioning of personnel in complex indoor environments.

[0030] Combination Figure 2 As shown, optionally, in S12, the server performs front-end signal enhancement processing on the original echo signal to obtain the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence, including: S21, the server converts the raw echo signal into a three-dimensional data matrix, which includes distance information, velocity information and angle information.

[0031] S22, the server inputs the three-dimensional data matrix into the trained deep learning model to extract the spatial features and temporal motion features of the target object.

[0032] S23, the server outputs the distance between the target object and the radar, radial velocity, and probability of presence based on the spatial and temporal motion characteristics of the target object.

[0033] In this scheme, when the server performs front-end signal enhancement processing on the raw echo signal, it first fuses and reconstructs the range-Doppler profile data and angle channel data acquired by the radar to form a three-dimensional data matrix containing range, velocity, and angle information. This data reconstruction process unifies the signal features that were originally scattered across multiple dimensions. Its core fusion method is to directly integrate the three physical dimensions—range unit, velocity unit, and antenna channel—into a unified data structure by stacking the channel dimensions. At the same time, it combines timestamp synchronization to ensure the temporal alignment of multi-dimensional data and normalizes the values ​​of each dimension to eliminate dimensional differences. Finally, it constructs a three-dimensional data matrix with a unified mathematical representation, establishing a complete data foundation for subsequent deep feature extraction.

[0034] The three-dimensional data matrix is ​​then input into a specially designed CNN-LSTM (Convolutional Long Short-Term Memory) hybrid network model for processing. This network structure fully utilizes the advantages of convolutional neural networks in spatial feature extraction. Through convolution operations on multi-channel radar data, it effectively captures the distribution characteristics of the target in the range-velocity-angle three-dimensional space. At the same time, it utilizes the temporal modeling capability of the long short-term memory network to learn the motion patterns of the target between consecutive frames, thereby suppressing instantaneous noise and abnormal jumps.

[0035] Optionally, the structural parameters of the CNN-LSTM hybrid network are as follows: the convolutional layers use 5×3×3 kernels, corresponding to the normal size of human subjects in the range-Doppler-angle profile of low-resolution millimeter-wave radar. The initial CNN layer is responsible for fusing information from the three angular channels, producing a compressed output matrix of shape 64×16×3. Subsequently, three ConvLSTM layers are used to extract spatial information and learn temporal correlations, with a kernel size of 5×3, a stride of 2, and feature dimensions of 16, 32, and 64 respectively. The last fully connected layer performs classification, producing multi-head outputs corresponding to the estimates of range, radial velocity, and presence probability, respectively. This approach can effectively capture subtle features of the target in the range-Doppler-angle space, improve the accuracy and robustness of feature extraction, thereby improving the accuracy of range, velocity, and presence probability estimation under low-resolution radar and reducing false alarms and missed detections caused by multipath and occlusion.

[0036] In one example, the training of this hybrid network model employs a carefully designed curriculum learning strategy, progressively increasing model complexity through three stages. The first stage uses single-user data without multipath interference to train the network, allowing it to initially grasp the basic motion characteristics of human targets. The second stage introduces data with mild multipath interference, enabling the network to learn to maintain the stability of feature extraction in environments with some interference. The third stage uses multi-user, dense scene data to train the network's ability to distinguish multiple neighboring targets in complex environments. This progressive training strategy ensures the model's robustness in real-world, complex scenarios.

[0037] Furthermore, at the network output level, based on the learned spatial and temporal motion features, the distance, radial velocity, and presence probability of the target are simultaneously output through a multi-task learning framework. Distance and velocity are accurately estimated through a regression task, while the presence probability is assessed using a classification task to evaluate the target's credibility. The entire process also employs a composite loss function for optimization, balancing the weight relationships between different output tasks.

[0038] This approach, by fusing spatiotemporal features with a progressive training strategy through deep neural networks, overcomes the performance limitations of traditional signal processing methods under low-resolution radar, significantly improving the accuracy and robustness of target motion parameter extraction in complex indoor environments, and providing a reliable parameter basis for subsequent angle estimation and trajectory generation.

[0039] Optionally, when performing front-end signal enhancement processing on the original echo signal, the construction of the three-dimensional data matrix also includes a two-dimensional fusion of amplitude and phase information: after temporally aligning the amplitude features of the range-Doppler profile with the phase features of the angle channel, they are mapped to a unified numerical range, forming a multi-dimensional data representation including range, velocity, angle, amplitude, and phase. This dual-information fusion strategy breaks through the limitations of traditional methods that rely solely on amplitude information. By using phase information to reflect the relative positional relationship between the target and the radar, it collaboratively constructs a more complete target feature model with amplitude information. This approach solves the feature ambiguity problem caused by the hardware limitations of low-resolution radar, significantly improves the distinction between clutter and real targets, reduces the false alarm rate and missed detection rate in multipath and obstructed environments, and provides a more reliable data foundation for subsequent accurate estimation of range, velocity, and presence probability.

[0040] In one optimized approach, an attention-guided multi-scale feature fusion method is introduced based on the construction of the 3D data matrix. Residual connections and channel attention modules are added to the CNN-LSTM hybrid network, enabling the network to adaptively focus on effective target features at different distance scales. This approach, by introducing an attention mechanism and multi-scale feature fusion, allows the network to adaptively focus on effective target features at different distance scales, significantly improving the ability to distinguish between real targets and multipath artifacts in complex scenes, and enhancing the accuracy of feature extraction and the robustness of the model.

[0041] In one optimized approach, a dynamic difficulty assessment mechanism is introduced at each training stage to automatically adjust the complexity distribution of the training data based on the model's real-time performance. A constraint term based on radar physics characteristics is incorporated into the loss function design, integrating the physical correlation between distance and velocity as prior knowledge into the model optimization process. This approach, through the synergistic effect of the dynamic difficulty assessment mechanism and the physically constrained loss function, enables the model to adaptively learn its progress and follow the physical laws of radar signals during training, effectively improving the accuracy of motion parameter estimation and the model's generalization ability in complex environments.

[0042] Optionally, the trained deep learning model is a CNN-LSTM hybrid network, which is trained using a curriculum learning strategy.

[0043] In this scheme, the trained deep learning model adopts a CNN-LSTM hybrid network structure, and its training process employs a curriculum learning strategy. This hybrid network uses a convolutional neural network module to extract multi-level spatial features from the input 3D data matrix, utilizing convolutional kernels to scan features in distance, velocity, and angle dimensions to capture the distribution characteristics of the target in 3D space. Simultaneously, a long short-term memory network module performs temporal modeling of the convolutional features across multiple frames, learning the dynamic patterns of target motion. The curriculum learning strategy is implemented in three orderly stages: the first stage uses single-user scene data without multipath interference to enable the network to initially establish basic feature extraction capabilities for human targets; the second stage introduces training data containing mild multipath interference to improve the network's feature preservation ability under signal distortion environments; the third stage uses multi-user dense scene data to train the network's feature discrimination ability under conditions of target cross-interference. This progressive training method ensures a smooth transition of the network from simple features to complex scenes, effectively avoiding gradient anomalies and overfitting during model training. The entire training process employs a composite loss function to jointly optimize distance regression, velocity regression, and existence probability classification tasks, and coordinates the learning progress of different tasks through a weight balancing mechanism. This design, through the organic combination of spatiotemporal feature extraction and a progressive training strategy, significantly improves the accuracy and environmental adaptability of low-resolution radar in estimating target motion parameters in complex indoor environments.

[0044] In one optimized approach, an attention-guided cross-modal fusion mechanism can be introduced into the CNN-LSTM hybrid network. By embedding a bidirectional attention module between convolutional and recurrent layers, the network can dynamically adjust the contribution weights of spatial and temporal features. Simultaneously, the learning strategy is improved by introducing adversarial sample generation technology in each training stage. This enhances model robustness by generating challenging samples that match the statistical characteristics of real-world scenarios. Furthermore, an information entropy-based regularization term can be added to the composite loss function to constrain the model's prediction uncertainty in complex scenarios. Through this approach, the synergistic optimization of the attention mechanism and adversarial training enables the network to adaptively focus on key features and enhance its generalization ability to unknown scenarios, while significantly reducing the false alarm probability in complex environments.

[0045] Optionally, course learning strategies include: In the first stage, a CNN-LSTM hybrid network was trained using single-user data free from multipath interference; and, In the second stage, a CNN-LSTM hybrid network was trained using data with mild multipath interference; and, In the third stage, the CNN-LSTM hybrid network was trained using dense scene data with multiple users and small user spacing.

[0046] In this scheme, the learning strategy is implemented in three stages, progressively increasing the complexity of the environment and the target. The first stage uses single-user data without multipath interference. The training environment is set in an open indoor space, with only a single target moving at a constant speed in a straight line. Any wall reflections or furniture scattering interference are avoided in the radar signal propagation path to enable the network to establish an accurate mapping relationship between the basic motion pattern of the human target and the radar echo characteristics. The second stage introduces training data with mild multipath interference. In a standard room environment with basic furniture, the multipath signal strength is controlled within 20%-40% of the original signal energy. The target moves in a variable-speed curve within a limited area to train the network's ability to maintain core target characteristics and suppress interference under signal distortion conditions. The third stage uses multi-user dense scene data. In a complex indoor environment, at least three targets are placed at intervals of less than 1.5 meters, moving intermittently, while preserving complete wall reflection and furniture scattering paths. This stage focuses on strengthening the network's ability to separate features of nearby targets and its continuous tracking performance under strong interference conditions. It should be noted that the advancement of each training stage is predicated on the network reaching a stable training state within the data environment defined in the previous stage, thereby systematically building its generalization ability. In this way, through this step-by-step, interconnected data environment design, the deep learning model can learn smoothly and effectively, ultimately significantly improving its resolution accuracy and robustness against low-resolution radar signals in real-world, complex indoor scenarios.

[0047] In practical implementation, the training data construction is specifically manifested as follows: Single-user data without multipath interference is collected from an open corridor scene exceeding 10 meters in length, with only a single tester performing regular back-and-forth movements along a line-of-sight path. Data with mild multipath interference is collected in a standard office environment, with 2-3 workstations and filing cabinets in the room. Testers move along a pre-defined curved path between workstations, and the radar received signal includes a primary reflection component from the furniture surface. Multi-user dense scene data is collected in a conference room environment, with 3-5 testers randomly walking at intervals of 0.8-1.2 meters. The signal contains multiple reflection paths from the conference table, walls, and human bodies, and the target spacing continuously changes dynamically. This learning strategy, through progressive environmental complexity design, allows the network to gradually adapt from ideal environments to real-world complex scenarios, significantly improving the robustness and environmental adaptability of multi-target motion parameter estimation under low-resolution radar conditions.

[0048] In this scheme, a real-time difficulty estimator can be constructed to perform multi-dimensional and complex quantification of training samples. The difficulty distribution of the training dataset is dynamically adjusted based on the real-time performance during model training. Simultaneously, a meta-learning algorithm is employed to predict the optimal difficulty scheduling strategy based on the model's gradient state, forming a closed-loop curriculum adjustment system based on the model's learning state. This scheme achieves precise matching between training courses and model capabilities, ensuring the model remains in its optimal learning range while maintaining stable training. Ultimately, this significantly improves the model's convergence speed and generalization performance in unknown scenarios.

[0049] Combination Figure 3 As shown, optionally, in S13, the server performs high-resolution angle estimation on target objects that meet preset existence conditions based on the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence, to calculate its azimuth and elevation angles, including: S31, the server filters out high-confidence targets from the target objects based on the probability of their existence.

[0050] S32, the server uses the distance and velocity information corresponding to high-confidence targets to compress the angle search space.

[0051] S33, within the compressed angle search space, the server uses a two-dimensional MUSIC algorithm based on temporal consistency constraints to calculate the azimuth and elevation angles of high-confidence targets.

[0052] In this scheme, when the server performs high-resolution angle estimation based on the distance between the target and the radar, radial velocity, and presence probability, it first executes a high-confidence target screening process. Specifically, the server can distinguish the confidence level of targets by setting a presence probability threshold. This threshold is determined based on the balance between the statistical characteristics of radar target detection and the system's accuracy requirements. As an example, targets with a presence probability greater than 0.8 can be classified as high-confidence targets. This threshold setting effectively retains true targets while filtering out most false alarm targets.

[0053] Furthermore, after target selection, the server intelligently compresses the angle search space using the distance and velocity information of high-confidence targets. Specifically, the server retains only the angle channel data within the distance and velocity units corresponding to these high-confidence targets, excluding data from other regions from the calculation scope. This space compression strategy based on prior information effectively reduces the angle search range, significantly lowers the computational complexity of subsequent algorithms, and ensures that no useful information is lost.

[0054] Optionally, before executing the two-dimensional MUSIC algorithm within the compressed angle search space, the target quantity is automatically identified using the maximum eigenvalue difference method: the eigenvalue sequence of the covariance matrix of the angle channel data is calculated, the difference between adjacent eigenvalues ​​is solved, and the critical point with the largest difference is used as the target quantity segmentation threshold to determine the true number of targets in the current scene, thereby dividing the signal subspace and noise subspace. This automatic identification mechanism avoids subspace misjudgment caused by manually setting the target quantity and adapts to the dynamic switching between multi-target dense and single-target scenes for low-resolution radar.

[0055] Furthermore, within the compressed angle search space, the server employs a two-dimensional MUSIC algorithm based on temporal consistency constraints for high-precision angle estimation. This algorithm introduces a kinematic constraint mechanism on top of the traditional MUSIC algorithm. First, it obtains candidate angle pairs by calculating the spatial spectrum function, and then analyzes the angular velocity changes of these candidate angles across consecutive frames. The system compares the calculated instantaneous angular velocity with an angular velocity threshold derived from human motion principles; candidate angles whose angular velocities exceed the limits of human physical motion are identified as false spectral peaks and eliminated. Simultaneously, the algorithm enhances the stability of angle estimation by calculating the covariance matrix of multi-frame angle data, ultimately outputting accurate azimuth and pitch angle estimates.

[0056] This approach, through the synergistic effect of prior information guidance and physical motion constraints, significantly improves the angle resolution accuracy and anti-interference capability of low-resolution radar in complex indoor environments while ensuring computational efficiency.

[0057] In one optimized scheme, a probabilistic correlation model of target distance-velocity-angle is established to deeply integrate the multidimensional features output by the front-end signal enhancement module with the angle search process. At the same time, the static temporal constraint is improved into an adaptive threshold adjustment mechanism based on individual motion features, and the angular velocity threshold is dynamically updated by learning the target motion pattern online. In addition, a multipath suppression filter is introduced into the two-dimensional MUSIC algorithm. By prior modeling the multipath features of the environment, the false spectral peaks generated by multipath interference are directly suppressed in the spectral estimation stage.

[0058] This scheme, through multimodal deep coupling and dynamic adaptive constraint mechanism, significantly improves the accuracy and stability of angle estimation under strong multipath interference environment, while enhancing the adaptability to targets with different motion modes.

[0059] Combination Figure 4 As shown, optionally, trajectory tracking processing includes: S41, the server obtains the dynamic state vector corresponding to the established target trajectory.

[0060] S42, the server predicts the target object's state at the next moment based on the dynamic state vector.

[0061] S43, the server will fuse the three-dimensional position observations obtained from the conversion of azimuth, pitch, and distance with the radial velocity observations, and perform data association and matching with the predicted state.

[0062] S44, the server generates a continuous trajectory of the target object in three-dimensional space based on the association matching results.

[0063] In this scheme, when the server performs trajectory tracking, it first obtains the dynamic state vector corresponding to the established target trajectory. This dynamic state vector is a system state description maintained by an improved extended Kalman filter. Based on this dynamic state vector, the server predicts the target's state at the next moment. This prediction process employs a dynamically adaptive noise matrix, whose process noise and observation noise matrices are dynamically adjusted according to the variance of the distance and angle observations measured by the radar in real time, thus adapting to the large fluctuations in the signal-to-noise ratio of low-resolution radar. Furthermore, when the server fuses azimuth, elevation, and range information to obtain 3D position observations, it first converts the polar coordinate data to Cartesian coordinates and then performs exponentially weighted sliding window filtering. For example, setting the window length to 5-8 frames, weights that decay over time are assigned to the 3D coordinate data of consecutive frames, and coordinate jumps caused by random noise are smoothed through weighted averaging to output stable 3D position observations. The weight coefficient for recent data is 0.6-0.8, and the weight coefficient for long-term data is 0.2-0.4. This approach can effectively suppress the jumps in positioning data caused by the sparsity and high noise ratio of low-resolution radar data, improve the stability of three-dimensional position observations, provide a smoother input for dynamic state estimation of subsequent trajectory tracking, and reduce trajectory interruptions and abnormal deviations.

[0064] Furthermore, the server fuses the converted 3D position observations with the radial velocity observations and performs data association matching with the predicted state. As an example, the Hungarian algorithm can be used for data association matching. This involves constructing a cost matrix based on Mahalanobis distance and incorporating the confidence weights of the observation points, calculating the matching cost between each observation point and each predicted state, and setting a maximum cost threshold to exclude abnormal matches. Further, after a successful match, when updating the trajectory state using the fused observations, this scheme performs multiple rounds of iterative linearization on the EKF observation model to reduce nonlinear errors. After the state update, physical constraints based on human kinematics are applied to the velocity component, limiting the velocity within a reasonable range through truncation operations to suppress non-physical trajectory jumps.

[0065] Understandably, data association matching will produce three results: when an observation point successfully matches a trajectory prediction state, it is used to update the corresponding trajectory; when an observation point does not match any existing trajectory, it is initialized as a new trajectory; when an existing trajectory does not match any observation point, a loss count is started, and if there is no match for 3 consecutive periods, the trajectory is determined to terminate.

[0066] This scheme, through the synergy of improved extended Kalman filtering and data association, achieves continuous and stable tracking of multiple target trajectories under low-resolution radar conditions, effectively overcoming the problems of trajectory confusion and interruption in complex scenarios such as target intersection and brief obstruction.

[0067] In one specific embodiment, the server tracks the trajectories of three intersecting individuals in a conference room environment: First, it acquires the established dynamic state vector of the trajectory. Here, the dynamic state vector of the trajectory includes position, velocity, and covariance; based on this vector, an improved extended Kalman filter is used to predict the state at the next moment. The process noise matrix is ​​dynamically adjusted according to the real-time observation variance; then, the three-dimensional coordinates (2.1, 4.8, 1.5) obtained by converting the azimuth angle of 12.5 degrees, the elevation angle of 3.2 degrees, and the distance of 5.2 meters are fused with the radial velocity of 1.2 m / s to form an observation vector, and the Mahalanobis distance between it and each predicted state is calculated using the Hungarian algorithm to complete the association matching; after a successful match, the state vector is updated using multi-round iterative linearization, and the output velocity is constrained by human motion to be limited to the range of 0-3 m / s; when a new observation point is not matched, a new trajectory is initialized, and when an existing trajectory is not matched for 3 consecutive frames, the trajectory is terminated. With this scheme, by improving filtering and multi-source information fusion, the continuity of multi-target trajectories and positioning stability are effectively improved under low-resolution radar, and the trajectory confusion problem caused by target intersection occlusion is solved.

[0068] Optionally, S44, the server generates a continuous trajectory of the target object in three-dimensional space based on the association matching results, including: When the fused observations successfully match the predicted state of the target trajectory, the server updates the dynamic state vector of the target trajectory using the fused observations and uses the updated state as the current point of the continuous trajectory.

[0069] When the fused observation does not match the predicted state of any target trajectory, the server uses it as the starting point of the new target trajectory and initializes the corresponding dynamic state vector.

[0070] In this scheme, when the server generates the continuous trajectory of the target object in 3D space based on the association matching results, it adopts differentiated trajectory processing strategies for different matching situations. When the fused observation value successfully matches the predicted state of the target trajectory, the server initiates a state update process, using an improved extended Kalman filter algorithm to process the fused observation value. This process involves gradually correcting the linearization error of the observation model through a multi-round iterative linearization method, while dynamically adjusting the weight coefficients of the noise matrix based on the real-time radar observation variance. Finally, the optimized state estimation result is used as the continuous trajectory point of the trajectory at the current moment, and the position, velocity, and covariance information in the dynamic state vector are updated synchronously.

[0071] When the fused observations fail to match the predicted state of any target trajectory, the server identifies it as a newly emerging target and executes a new trajectory initialization process. This process first establishes a new dynamic state vector, initializing its position component to the current 3D coordinate observations, its velocity component to the projection estimate of the radial velocity observations in Cartesian coordinates, and setting the covariance matrix to a preset large initial value to reflect the uncertainty of the initial state. Simultaneously, a unique identifier is assigned to this new trajectory, thus forming the starting point for trajectory tracking.

[0072] It should be noted that during continuous trajectory tracking, the server also establishes a trajectory lifecycle management mechanism. For existing trajectories that fail to match observations, the system will activate a loss counting mechanism. When no valid match is obtained for three consecutive tracking cycles, it is determined that the target has left the monitoring area or is continuously occluded, and the corresponding trajectory tracking will be terminated and related system resources will be released.

[0073] This approach combines the trajectory generation mechanism with an improved state estimation method and complete lifecycle management, enabling continuous and stable construction of multi-target trajectories in a low-resolution radar environment. This effectively solves the problems of trajectory breakage and identity confusion caused by target intersection, obstruction, and entry / exit from the monitoring area.

[0074] In one optimized scheme, after generating the continuous trajectory of the target object in three-dimensional space, the server initiates a reverse optimization process. This involves evaluating trajectory smoothness by calculating the rate of change of curvature and the variance of velocity fluctuation. If the rate of change of curvature exceeds a preset threshold or the variance of velocity fluctuation is greater than 0.2 m² / s², the weights α, β, γ, and δ of the composite loss function of the distance-Doppler estimator are dynamically adjusted, while the temporal consistency constraint threshold ω of the angle-of-arrival estimator is corrected. max Repeatedly iterate and optimize until the trajectory smoothness meets the preset standard. The preset threshold can be set to 0.3-0.5 rad / m based on the limits of human motion.

[0075] As an example, the parameter adjustment logic is precisely adapted to human motion characteristics and radar signal processing rules based on the trajectory smoothness anomaly type: when the rate of curvature change exceeds 0.3-0.5 rad / m, the trajectory smoothing constraint loss weight β in the composite loss function is increased from 0.25 to 0.35-0.4, and the distance regression loss weight α is decreased from 0.65 to 0.55-0.6. Simultaneously, the temporal consistency constraint threshold ω of the angle of arrival estimator module is increased. max Reduce the weight by 15%-20% to suppress abrupt angle changes; when the velocity fluctuation variance is greater than 0.2 m² / s², increase the physical constraint loss weight δ from 0.1 to 0.15-0.2, maintain the existence probability loss weight γ at 0.1 or finely adjust it to 0.08-0.12, and simultaneously reduce α by 5%-10%, ω max Adjust by 10%-15% to adapt to reasonable speed changes. After all weight adjustments, α+β+γ+δ=1 must be satisfied. After each round of adjustment, signal processing and trajectory generation are re-executed, iterating until the trajectory smoothness meets the standard.

[0076] Combination Figure 5 As shown, this embodiment of the present disclosure provides a personnel positioning device 200 for indoor environments, including an acquisition module 51, a processing module 52, a calculation module 53, and a fusion module 54. The acquisition module 51 is configured to acquire the raw echo signal collected by a low-resolution millimeter-wave radar in an indoor environment; the processing module 52 is configured to perform front-end signal enhancement processing on the raw echo signal to obtain the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence; the calculation module 53 is configured to perform high-resolution angle estimation on the target object that meets preset existence conditions based on the distance between the target object and the radar, the radial velocity, and the probability of the target object's presence, to calculate its azimuth and elevation angles; the fusion module 54 is configured to fuse the azimuth angle, elevation angle, and the distance and radial velocity corresponding to the target object that meets the preset existence conditions, and generate a continuous trajectory of the target object in three-dimensional space through trajectory tracking processing.

[0077] The personnel positioning device 200 for indoor environments provided in this disclosure estimates the distance, velocity, and probability of existence of the target through front-end signal enhancement processing, effectively suppressing signal distortion and attenuation caused by multipath and obstruction; then, it uses this prior information to compress the search space and perform high-resolution angle estimation, breaking through the hardware limitations of low-resolution radar; finally, it uses the fusion of multi-dimensional observation information for trajectory tracking, solving the problems of target confusion, trajectory interruption, and poor positioning stability caused by missing front-end information, and realizing accurate, continuous, and stable three-dimensional positioning of personnel in complex indoor environments.

[0078] Combination Figure 6As shown, this disclosure provides a personnel positioning device 300 for indoor environments, including a processor 301 and a memory 302. Optionally, the device 300 may further include a communication interface 303 and a bus 304. The processor 301, communication interface 303, and memory 302 can communicate with each other via the bus 304. The communication interface 303 can be used for information transmission. The processor 301 can call logical instructions in the memory 302 to execute the personnel positioning method for indoor environments described in the above embodiments.

[0079] Furthermore, the logic instructions in the aforementioned memory 302 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0080] The memory 302, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor 301 executes functional applications and data processing by running the program instructions / modules stored in the memory 302, thereby implementing the personnel positioning method for indoor environments described in the above embodiments.

[0081] The memory 302 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 302 may include high-speed random access memory and may also include non-volatile memory.

[0082] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to perform the above-described method for locating people in an indoor environment.

[0083] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more 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 method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, such as a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., and other media capable of storing program code.

[0084] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0086] The methods and products disclosed in the embodiments herein (including but not limited to devices and equipment) can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0087] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for locating a person in an indoor environment, characterized by, The method comprises the following steps: acquiring original echo signals collected by a low-resolution millimeter wave radar in an indoor environment; performing front-end signal enhancement processing on the original echo signals to obtain the distance between a target object and the radar, the radial velocity of the target object, and the existence probability of the target object; performing high-resolution angle estimation on the target object satisfying a preset existence condition based on the distance between the target object and the radar, the radial velocity of the target object, and the existence probability of the target object to calculate the azimuth angle and the elevation angle of the target object; fusing the azimuth angle, the elevation angle, and the distance and the radial velocity corresponding to the target object satisfying the preset existence condition, and generating a continuous trajectory of the target object in a three-dimensional space through trajectory tracking processing.

2. The method of claim 1, wherein, The front-end signal enhancement processing on the original echo signals to obtain the distance between a target object and the radar, the radial velocity of the target object, and the existence probability of the target object comprises the following steps: converting the original echo signals into a three-dimensional data matrix, wherein the three-dimensional data comprises distance information, velocity information, and angle information; inputting the three-dimensional data matrix into a trained deep learning model to extract the spatial features and the time-dimension motion features of the target object; outputting the distance between the target object and the radar, the radial velocity of the target object, and the existence probability of the target object based on the spatial features and the time-dimension motion features of the target object.

3. The method of claim 2, wherein, The trained deep learning model is a CNN-LSTM hybrid network, and the CNN-LSTM hybrid network is trained through a curriculum learning strategy.

4. The method of claim 3, wherein, The curriculum learning strategy comprises the following steps: training the CNN-LSTM hybrid network using single-user data without multipath interference in a first stage; and training the CNN-LSTM hybrid network using data with light multipath interference in a second stage; and training the CNN-LSTM hybrid network using multi-user and dense scene data with small user spacing in a third stage.

5. The method of claim 1, wherein, The high-resolution angle estimation on the target object satisfying the preset existence condition based on the distance between the target object and the radar, the radial velocity of the target object, and the existence probability of the target object comprises the following steps: selecting a high-confidence target from the target object based on the existence probability of the target object; compressing an angle search space by using the distance and velocity information corresponding to the high-confidence target; calculating the azimuth angle and the elevation angle of the high-confidence target by using a two-dimensional MUSIC algorithm based on time sequence consistency constraint in the compressed angle search space.

6. The method of claim 1, wherein, The trajectory tracking processing comprises the following steps: acquiring a dynamic state vector corresponding to an established target trajectory; predicting a predicted state of the target object at a next time based on the dynamic state vector; fusing a three-dimensional position observation value converted based on the azimuth angle, the elevation angle, and the distance with the radial velocity observation value, and performing data association matching with the predicted state; generating a continuous trajectory of the target object in a three-dimensional space based on the association matching result.

7. The method of claim 6, wherein, The generating of the continuous trajectory of the target object in the three-dimensional space based on the association matching result comprises the following steps: When the fused observation value matches the predicted state of the target trajectory successfully, the dynamic state vector of the target trajectory is updated using the fused observation value, and the updated state is taken as the current point of the continuous trajectory; When the fused observation value does not match the predicted state of any target trajectory successfully, it is taken as the starting point of a new target trajectory, and the corresponding dynamic state vector is initialized.

8. A personnel positioning device for use in an indoor environment, characterized by The method comprises the following steps: An acquisition module is configured to acquire original echo signals collected by a low-resolution millimeter wave radar in an indoor environment; A processing module is configured to perform front-end signal enhancement processing on the original echo signals to obtain the distance between a target object and the radar, the radial velocity of the target object, and the existence probability of the target object; A calculation module is configured to perform high-resolution angle estimation on the target object satisfying the preset existence condition based on the distance between the target object and the radar, the radial velocity of the target object, and the existence probability of the target object, to calculate the azimuth angle and the elevation angle of the target object; A fusion module is configured to fuse the azimuth angle, the elevation angle, and the distance and the radial velocity corresponding to the target object satisfying the preset existence condition, and generate a continuous trajectory of the target object in a three-dimensional space through trajectory tracking processing.

9. A device for locating a person in an indoor environment, comprising a processor and a memory having stored program instructions, characterized in that, The processor is configured to execute the program instructions to perform the personnel positioning method in an indoor environment according to any one of claims 1 to 7 when the program instructions are executed.

10. A computer readable storage medium storing program instructions, characterized in that, The program instructions, when executed, cause the computer to perform the personnel positioning method in an indoor environment according to any one of claims 1 to 7.

Images